WO2019047710A1 - 有限冲激响应滤波器系数矢量的可持续更新方法及装置 - Google Patents
有限冲激响应滤波器系数矢量的可持续更新方法及装置 Download PDFInfo
- Publication number
- WO2019047710A1 WO2019047710A1 PCT/CN2018/101491 CN2018101491W WO2019047710A1 WO 2019047710 A1 WO2019047710 A1 WO 2019047710A1 CN 2018101491 W CN2018101491 W CN 2018101491W WO 2019047710 A1 WO2019047710 A1 WO 2019047710A1
- Authority
- WO
- WIPO (PCT)
- Prior art keywords
- signal
- subband
- fir filter
- time
- far
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Ceased
Links
Images
Classifications
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L21/0216—Noise filtering characterised by the method used for estimating noise
- G10L21/0232—Processing in the frequency domain
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04M—TELEPHONIC COMMUNICATION
- H04M9/00—Arrangements for interconnection not involving centralised switching
- H04M9/08—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic
- H04M9/082—Two-way loud-speaking telephone systems with means for conditioning the signal, e.g. for suppressing echoes for one or both directions of traffic using echo cancellers
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; ELECTRIC HEARING AIDS; PUBLIC ADDRESS SYSTEMS
- H04R27/00—Public address systems
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; ELECTRIC HEARING AIDS; PUBLIC ADDRESS SYSTEMS
- H04R3/00—Circuits for transducers
- H04R3/02—Circuits for transducers for preventing acoustic reaction, i.e. acoustic oscillatory feedback
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L2021/02082—Noise filtering the noise being echo, reverberation of the speech
-
- G—PHYSICS
- G10—MUSICAL INSTRUMENTS; ACOUSTICS
- G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
- G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
- G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
- G10L21/0208—Noise filtering
- G10L21/0216—Noise filtering characterised by the method used for estimating noise
- G10L21/0224—Processing in the time domain
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; ELECTRIC HEARING AIDS; PUBLIC ADDRESS SYSTEMS
- H04R2227/00—Details of public address [PA] systems covered by H04R27/00 but not provided for in any of its subgroups
- H04R2227/009—Signal processing in [PA] systems to enhance the speech intelligibility
-
- H—ELECTRICITY
- H04—ELECTRIC COMMUNICATION TECHNIQUE
- H04R—LOUDSPEAKERS, MICROPHONES, GRAMOPHONE PICK-UPS OR LIKE ACOUSTIC ELECTROMECHANICAL TRANSDUCERS; ELECTRIC HEARING AIDS; PUBLIC ADDRESS SYSTEMS
- H04R2499/00—Aspects covered by H04R or H04S not otherwise provided for in their subgroups
- H04R2499/10—General applications
- H04R2499/11—Transducers incorporated or for use in hand-held devices, e.g. mobile phones, PDA's, camera's
Definitions
- the present disclosure relates to the field of signal processing technologies, and in particular, to a sustainable update method for a finite impulse response filter coefficient vector and a sustainable update device for a finite impulse response filter coefficient vector.
- the Acoustic Echo Canceller is a key component of a full-duplex voice communication system whose primary function is to remove the echo signal of the far-end signal coupled into the microphone (microphone) by the speaker (horn).
- the echo path is adaptively modeled using a finite impulse response (FIR) linear filter, and a valid estimate of the echo is synthesized, and then the estimate is subtracted from the received signal of the microphone. Thereby completing the task of echo cancellation.
- FIR finite impulse response
- adaptive variable-step learning techniques are very sensitive to parameter initialization; in addition, this type of technique is not suitable for dealing with "double talk" situations where there is near-end (non-stationary) speech in echo cancellation applications.
- Other schemes in the related art such as the generalized normalized gradient descent (GNGD) algorithm, do not handle the "double talk" situation with near-end (non-stationary) speech.
- GNGD generalized normalized gradient descent
- Some embodiments of the present disclosure provide a method and apparatus for continuously updating a finite impulse response filter coefficient vector, so as to solve the problem that the adaptive learning mode of the related FIR filter cannot guarantee the stability of the performance of the FIR filter, thereby affecting the signal. Handling reliability issues.
- some embodiments of the present disclosure provide a sustainable adaptive update method for a finite impulse response FIR filter coefficient vector, including:
- the FIR filter coefficient vector is updated according to the time varying regularization factor.
- the preset signal comprises: a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone, a noise reference signal and a system input signal in an adaptive noise cancellation system, and adaptive interference.
- a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone
- a noise reference signal and a system input signal in an adaptive noise cancellation system
- adaptive interference One of a combination of the interference reference signal and the system input signal in the cancellation system, the excitation input signal in the adaptive system identification, and the unknown system output signal to be identified.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone, where the acquiring is used to perform the FIR filtering when the preset signal is processed by using the FIR filter coefficient vector.
- the step of the iteratively updated time-varying regularization factor of the vector coefficient vector includes:
- the method for obtaining the power of the signal received by the microphone is:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the effective estimation value of the coupling factor is obtained by:
- the step of obtaining a biased estimation value of the coupling factor according to the cross-correlation method includes:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the step of acquiring a correction factor for the biased estimate compensation of the coupling factor comprises:
- the step of acquiring the candidate value of the square of the correlation coefficient between the error signal of the AEC output and the far-end reference signal includes:
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the candidate value according to the square of the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal is obtained, and the square of the effective amplitude of the correlation coefficient between the error signal of the AEC output and the far-end reference signal is obtained. Steps, including:
- the step of obtaining a valid estimated value of the coupling factor according to the biased estimation value of the coupling factor and the correction factor including:
- a valid estimate of the coupling factor a biased estimate of the coupling factor based on cross-correlation techniques;
- the effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the time-varying regularization factor for the iterative update of the FIR filter coefficient vector when the preset signal is processed by using the FIR filter coefficient vector is obtained according to the power of the microphone received signal and the effective estimated value of the coupling factor. Steps include:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0
- t is the digital signal sample time index number.
- the step of updating the FIR filter coefficient vector according to the time varying regularization factor includes:
- the step of updating the FIR filter coefficient vector according to the time varying regularization factor includes:
- a conjugate transposed matrix of X state (t) I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of a far-end reference voice signal, where the FIR filter coefficient vector is a sub-band FIR filter coefficient.
- the time-varying regularization factor is a sub-band time-varying regularization factor
- the obtaining is a time-varying regularity for performing iterative update of the FIR filter coefficient vector when the preset signal is processed by using the FIR filter coefficient vector
- the steps of the factor including:
- the subband domain FIR filter coefficient vector used for processing the preset signal by using the subband FIR filter coefficient vector Iteratively updated subband time-varying regularization factor.
- the subband power spectrum of the received signal of the microphone is obtained by:
- the obtaining of the effective estimated value of the sub-band coupling factor includes:
- the step of acquiring the biased estimation value of the sub-band coupling factor according to the cross-correlation method includes:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- the step of acquiring a correction factor for the biased estimate compensation of the subband domain coupling factor comprises:
- the step of acquiring the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far reference signal subband spectrum includes:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the AEC output error signal subband spectrum and the far reference signal subband are obtained according to the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far end reference signal subband spectrum.
- the steps of squaring the effective magnitude of the correlation coefficient between spectra including:
- n is the signal frame time index variable.
- the step of obtaining a valid estimated value of the sub-band coupling factor according to the biased estimation value of the sub-band coupling factor and the correction factor including:
- n is the signal frame time index variable.
- the obtaining, according to the sub-band power spectrum of the microphone received signal and the effective estimated value of the sub-band coupling factor, acquiring the sub-band when the sub-band FIR filter coefficient vector is used to process the preset signal includes:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- the step of updating the FIR filter coefficient vector according to the time varying regularization factor includes:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the n signal frame time, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the n
- the step of updating the FIR filter coefficient vector according to the time varying regularization factor includes:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- some embodiments of the present disclosure also provide a sustainable adaptive update apparatus for a finite impulse response FIR filter coefficient vector, including a memory, a processor, and a memory stored on the processor and a computer program running thereon; wherein the processor executes the computer program to implement the following steps:
- the FIR filter coefficient vector is updated according to the time varying regularization factor.
- the preset signal includes: a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone, a noise reference signal and a system input signal in an adaptive noise cancellation system, and adaptive interference cancellation.
- a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone
- a noise reference signal and a system input signal in an adaptive noise cancellation system
- adaptive interference cancellation One of a combination of the interference reference signal and the system input signal in the system, the excitation input signal in the adaptive system identification, and the unknown system output signal to be identified.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- a valid estimate of the coupling factor a biased estimate of the coupling factor based on cross-correlation techniques;
- the effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the processor may further implement the following steps when the computer program is executed by the processor:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0
- t is the digital signal sample time index number.
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- a conjugate transposed matrix of X state (t) I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of the far-end reference voice signal, where the FIR filter coefficient vector is a sub-band FIR filter.
- the time-varying regularization factor is a sub-band time-varying regularization factor, and the processor may further implement the following steps when the computer program is executed by the processor:
- the subband domain FIR filter coefficient vector used for processing the preset signal by using the subband FIR filter coefficient vector Iteratively updated subband time-varying regularization factor.
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- the processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps when the computer program is executed by the processor:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the processor may further implement the following steps when the computer program is executed by the processor:
- n is the signal frame time index variable.
- the processor may further implement the following steps when the computer program is executed by the processor:
- n is the signal frame time index variable.
- the processor may further implement the following steps when the computer program is executed by the processor:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- the processor may further implement the following steps when the computer program is executed by the processor:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the n signal frame time, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the n
- the processor may further implement the following steps when the computer program is executed by the processor:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- some embodiments of the present disclosure further provide a computer readable storage medium having stored thereon a computer program, wherein the processor implements the following steps when the computer program is executed by a processor:
- the FIR filter coefficient vector is updated according to the time varying regularization factor.
- some embodiments of the present disclosure further provide a sustainable adaptive updating apparatus for a finite impulse response FIR filter coefficient vector, including:
- An obtaining module configured to acquire a time-varying regularization factor used to perform the iterative update of the FIR filter coefficient vector when the preset signal is processed by using the FIR filter coefficient vector;
- an update module configured to update the FIR filter coefficient vector according to the time varying regularization factor.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone, where the acquiring module includes:
- a first acquiring unit configured to separately obtain a power of a microphone received signal and a valid estimated value of a coupling factor
- a second acquiring unit configured to acquire, according to the power of the microphone received signal and the effective estimated value of the coupling factor, when the FIR filter coefficient vector is used to perform the iterative update of the FIR filter coefficient vector when the preset signal is processed Change the regularization factor.
- the acquiring manner of the power of the microphone receiving signal in the first acquiring unit is:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the obtaining manner of the effective estimated value of the coupling factor in the first acquiring unit is:
- the manner of obtaining the biased estimation value of the coupling factor according to the cross-correlation method is:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the manner of obtaining the correction factor for the biased estimate compensation of the coupling factor is:
- the manner of obtaining the candidate value of the square of the correlation coefficient between the error signal of the AEC output and the far-end reference signal is:
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the candidate value according to the square of the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal is obtained, and the square of the effective amplitude of the correlation coefficient between the error signal of the AEC output and the far-end reference signal is obtained.
- the manner of obtaining the effective estimated value of the coupling factor according to the biased estimation value of the coupling factor and the correction factor is:
- a valid estimate of the coupling factor a biased estimate of the coupling factor based on cross-correlation techniques;
- the effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the second obtaining unit is configured to:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0
- t is the digital signal sample time index number.
- the update module is configured to:
- the update module is configured to:
- a conjugate transposed matrix of X state (t) I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of the far-end reference voice signal, where the FIR filter coefficient vector is a sub-band FIR filter.
- a coefficient vector, the time-varying regularization factor is a sub-band time-varying regularization factor, and the obtaining module includes:
- a third acquiring unit configured to respectively obtain a sub-band power spectrum of the microphone received signal and a valid estimated value of the sub-band coupling factor
- a fourth acquiring unit configured to obtain, according to a sub-band power spectrum of the microphone received signal and a valid estimated value of a sub-band coupling factor, to obtain the sub-band FIR filter coefficient vector for performing the sub-band processing
- the sub-band time-varying regularization factor is updated iteratively with the band FIR filter coefficient vector.
- the acquiring the subband power spectrum of the microphone receiving signal in the third acquiring unit is:
- the obtaining manner of the effective estimated value of the sub-band coupling factor in the third acquiring unit includes:
- the manner of obtaining the biased estimation value of the sub-band coupling factor according to the cross-correlation method is:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- the manner of obtaining the correction factor for the biased estimation of the sub-band coupling factor is:
- the manner of obtaining the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far reference signal subband spectrum is:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the AEC output error signal subband spectrum and the far reference signal subband are obtained according to the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far end reference signal subband spectrum.
- the way to square the effective magnitude of the correlation coefficient between spectra is:
- n is the signal frame time index variable.
- the manner of obtaining the effective estimated value of the sub-band coupling factor according to the biased estimation value of the sub-band coupling factor and the correction factor is:
- n is the signal frame time index variable.
- the fourth obtaining unit is configured to:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- the update module is configured to:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the n signal frame time, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the n
- the update module is configured to:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- the time-varying regularization factor for performing the iterative update of the FIR filter coefficient vector is obtained, and the FIR filter coefficient vector is continuously adaptively updated according to the time-varying regularization factor, thereby ensuring the performance of the FIR filter.
- the stability of the signal processing improves the reliability of the signal processing.
- Figure 1 shows a schematic diagram of the working principle of the AEC
- FIG. 2 is a flow chart showing a method of sustainable adaptive updating of FIR filter coefficient vectors of some embodiments of the present disclosure
- Figure 3 is a schematic diagram showing the structure of the time domain LAEC
- Figure 4 is a schematic view showing the structure of a sub-band AEC
- Figure 5 is a schematic view showing the LAEC structure of the sub-band k
- FIG. 6 is a block diagram showing the structure of a sustainable adaptive updating apparatus for a FIR filter coefficient vector of some embodiments of the present disclosure
- FIG. 7 is a block diagram showing a sustainable adaptive update device for FIR filter coefficient vectors of some embodiments of the present disclosure.
- the AEC works as shown in Figure 1. It usually consists of a linear echo canceller (LAEC), a "double talk” detector (DTD) and a nonlinear residual echo suppressor (RES).
- LAEC uses FIR linear filtering.
- the downlink signal ie, the reference signal, denoted by x(t)
- x(t) the reference signal
- RES nonlinear residual echo suppressor
- the coefficient of the FIR filter is usually only used in the case of "single talk” (ie no near-end speech signal (represented by s(t))
- the signal represented by e(t)
- NLMS echo cancellation
- ⁇ is a constant learning rate parameter
- ⁇ is a small normal amount of regularization factor parameter
- L is the number of filter coefficients.
- the AEC FIR filter coefficient adaptive update must be performed only in the absence of a near-end speech signal (ie, "single talk” mode); in the case of a near-end speech signal (ie "double talk”) Mode), the adaptive line of its coefficients must be stopped to avoid the divergence of the filter coefficients and cause damage to the near-end speech signal.
- the detection of "double talk” mode is usually done by DTD.
- DTD processing delay of DTD and its possible misjudgment will seriously affect the adaptive learning behavior of AEC filter, which will affect the performance of AEC.
- many adaptive variable step size learning techniques have been proposed to continuously iteratively update the coefficients of the FIR filter.
- GNGD generalized normalized gradient descent
- ⁇ (t) is the learning rate
- ⁇ (t) is the learning rate control parameter
- ⁇ 1 is the adaptive iterative step size parameter of ⁇ (t).
- the GNGD algorithm is not sensitive to the initialization of its related parameters, the learning parameter ( ⁇ (t)) of the learning rate ( ⁇ (t)) is adaptively adjusted based on the behavioral characteristics of the NLMS stochastic gradient, so it can only be relatively Slower adaptive adjustments (usually requiring tens or hundreds of samples). For this reason, the GNGD algorithm does not handle the "double talk" situation with near-end (non-stationary) speech.
- the adaptive learning method of the related FIR filter of the present disclosure cannot guarantee the stability of the performance of the FIR filter, affect the reliability of the signal processing, and provide a sustainable update of the coefficient vector of the finite impulse response filter.
- some embodiments of the present disclosure provide a sustainable adaptive update method for a finite impulse response FIR filter coefficient vector, including steps 21-22.
- Step 21 Acquire a time-varying regularization factor for performing iterative update of the FIR filter coefficient vector when the FIR signal coefficient vector is used to process the preset signal.
- the preset signal includes: a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone, a noise reference signal and a system input signal in an adaptive noise cancellation system, and adaptive interference.
- a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone
- a noise reference signal and a system input signal in an adaptive noise cancellation system
- adaptive interference One of a combination of the interference reference signal and the system input signal in the cancellation system, the excitation input signal in the adaptive system identification, and the unknown system output signal to be identified.
- Step 22 Update the FIR filter coefficient vector according to the time varying regularization factor.
- the time-varying regularization factor for performing the iterative update of the FIR filter coefficient vector is obtained, and the FIR filter coefficient vector is continuously adaptively updated according to the time-varying regularization factor, thereby ensuring the performance of the FIR filter.
- the stability of the signal processing improves the reliability of the signal processing.
- the following takes the processing of the far-end reference speech signal input in the AEC and the speech signal received by the near-end microphone as an example, starting from the time domain and the sub-band domain, respectively, and the present disclosure is explained as follows.
- the present disclosure mainly discusses its persistent adaptive learning problem with good robustness to "double talk" and echo path changes, and proposes a time domain LAEC self.
- Adaptive real-time online calculation method for time-varying regularization factors the FIR filter coefficient vector adaptive iterative update algorithm in LAEC composed of the time-varying regularization factor has good robustness to "double talk” and echo path changes.
- the normalized least mean square (NLMS) algorithm of time domain LAEC adaptive learning is taken as an example to elaborate the method of solving the time-varying regularization factor in real time, and then apply the time-varying regularization factor to LAEC persistence.
- NLMS normalized least mean square
- a time-varying regularization factor NLMS algorithm and a time-varying regularization factor affine projection (AP) algorithm are presented.
- Equation 6 the FIR filter coefficient vector (or FIR filter complex coefficient vector) at time t is expressed by Equation 6 as:
- Equation 7 the NLMS learning algorithm for updating the FIR filter coefficients
- e(t) is the error signal output by the AEC at the time t sample
- y(t) is the microphone received signal at the time t sample
- Conjugate transposed matrix For the far-end reference signal vector, the symbol * represents the complex conjugate operation
- H is the complex conjugate transpose operator
- 0 ⁇ ⁇ ⁇ 2 is the predetermined coefficient update step parameter
- ⁇ > 0 is called a "regularization"
- the small positive real constant of the factor "" avoids the denominator of the formula eight to be zero.
- Equation 9 Expressed in Equation 9 as:
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample.
- d(t) is the echo signal at time t
- s(t) is the near-end speech signal at time t.
- LAEC distortion-free error or system error of the filter (ie residual echo).
- Equation 8 replace ⁇ with ⁇ (t) to get the formula fourteen:
- Equation 14 the formula 2 is obtained by using the 2 norm of the complex vector:
- Equation 18 Equation 18
- the optimal regularization factor ⁇ opt (t) is expressed by the formula twenty-one:
- the formula twenty-four The power of the signal received by the microphone at time t, The power of the echo signal at time t. Since the power of the echo signal It is not directly available, so in order to facilitate the realization of the project, the present disclosure simplifies the formula twenty-four to:
- the control variable ⁇ opt (t), which updates the iterative step size as a filter coefficient vector, can be estimated using the formula twenty-fifth, although the filter coefficient is over-suppressed slightly in the presence of echo. Iterative, but at the higher level of the near-end speech signal, the desired effect of slowing the filter coefficient vector update iteration can still be achieved.
- Engineering calculations can be implemented online in real time through a first-order recursive model, considering The change in the level of the near-end speech signal must be tracked in time, and the present disclosure estimates the variables in real time online using the following "fast attack/slow decay" method. That is, there are formula twenty-six:
- the formula twenty-five also involves parameters The estimation of this parameter is actually an estimation problem of the echo coupling factor ⁇ (t).
- the coupling factor ⁇ (t) is defined by its formula and can be directly estimated by the following formula 27:
- Equation 28 uses the statistical irrelevance between the near-end signal s(t) and the reference signal x(t) to remove the influence of the near-end speech s(t), this estimate is a biased estimate and thus exists deviation. To do this, we need to make correction compensation to improve the estimation accuracy of the coupling factor ⁇ (t).
- Equation 29 shows that in the case of “single talk”, the ratio of ⁇ (t)
- the (normalized) correlation coefficient magnitude between e(t) and the far-end reference signal x(t) is squared.
- the formula 30 used to estimate the coupling factor is not only applicable to the “single talk” case, but also to the “double talk” case, as long as the correlation coefficient between e(t) and x(t) can be effectively estimated. Squared by magnitude. How to correctly and effectively estimate the square of the correlation coefficient can be obtained through the following analysis and discussion. Notice the fact that in the case of "single talk",
- Equation 28 the estimate of the ⁇ (t)
- Equation thirty-two can be expressed as the formula thirty-five:
- T s ⁇ L is a positive integer, which is used to estimate the number of samples used in Equation 34 to Formula 35.
- Formula 30, Equation thirty-three to Formula 35 constitute a kind of scene that can be used for both “double talk” and echo path changes.
- High-precision estimation algorithm obtained by the algorithm The estimated value is substituted into the formula twenty-five to get the formula thirty-six:
- step 21 the specific implementation of step 21 in the present disclosure is:
- the method for obtaining the power of the microphone receiving signal is:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the step of obtaining a biased estimation value of the coupling factor according to the cross-correlation method includes:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the step of acquiring a correction factor for the biased estimation value compensation of the coupling factor includes:
- the step of acquiring the candidate value of the square of the correlation coefficient between the error signal of the AEC output and the far-end reference signal includes:
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the step of obtaining the square of the effective amplitude of the correlation coefficient between the error signal of the AEC output and the far-end reference signal according to the candidate value of the square of the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal includes:
- the step of obtaining a valid estimated value of the coupling factor according to the biased estimation value of the coupling factor and the correction factor comprises:
- the time-varying regularization factor for the iterative update of the FIR filter coefficient vector when the FIR signal coefficient vector is used for the preset signal processing is acquired. Steps, including:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0.
- step 22 when the FIR filter coefficient vector is continuously adaptively updated using the NLMS algorithm, the specific implementation manner of step 22 is:
- the FIR filter coefficient vector before updating; ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample;
- T is the transpose operator; for Conjugate transposed matrix;
- ⁇ opt (t) is a time-varying regularization factor;
- e * (t) is a complex conjugate of e(t); e(t) is an error signal outputted by the AEC at the time t sample;
- y(t) is a microphone received signal at the time t sample;
- the first step initialization
- Step 2 Calculate related variables online
- Step 3 Calculate the regularization factor online
- Step 4 Update the iterative FIR filter coefficient vector
- Equation 37 the optimal time-varying regularization factor determined by Equation 37 is also applicable to the AP algorithm for time-domain LAEC adaptive learning.
- the time domain LAEC adaptive learning AP algorithm can be characterized by Equation 38 and Equation 39:
- ⁇ (t) is the regularization factor
- Receive a signal vector for a P-dimensional microphone For the state matrix of L ⁇ P dimension, I P ⁇ P is a P ⁇ P dimension unit matrix, and P is the order of the AP algorithm.
- the P-dimensional near-end speech signal vector is the P-dimensional near-end speech signal vector.
- Equation 18 ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇ ⁇
- step 22 Comparing the forty-seventh formula with the formula twenty-four, that is, the optimal time-varying regularization factor determined by the formula thirty-seven is also applicable to the AP algorithm; specifically, the FIR filter coefficient vector is continued using the AP algorithm.
- the specific implementation of step 22 is:
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t is the digital signal
- the first step initialization
- Step 2 Calculate related variables online
- Step 3 Calculate the regularization factor ⁇ opt (t) online
- Step 4 Update the iterative FIR filter coefficient vector
- the time domain affine projection (AP) algorithm has a certain degree of compromise in the convergence speed of the algorithm and its computational complexity, but in the case that the affine projection order P value is large (so that the AP algorithm converges faster) Its computational complexity is also difficult to implement on current commercial DSPs. This has prompted people to study and explore the frequency domain implementation technology of AEC, especially the current technology of sub-band domain.
- the adaptive learning algorithm in the subband domain, certain characteristics of the speech signal and the acoustic echo path transfer function (eg, the number of echo path transfer function coefficients and the weight update rate in each subband, the subband signal dynamic range) It is smaller than its time domain signal, etc.) It will be beneficial for the adaptive learning algorithm to quickly estimate and track the acoustic echo path transfer function.
- the following describes the adaptive learning problem of applying the time domain LAEC time-varying regularization factor adaptive learning algorithm discussed above to the sub-band LAEC, and proposes a time-varying regularization of sub-band LAEC adaptive learning. Factor NLMS algorithm and AP algorithm.
- the structure of the subband AEC is shown in Fig. 4. It consists of an analysis filter bank (AFB), a subband linear echo canceller (LAEC), a subband residual echo suppressor (RES), and a synthesis filter bank (SFB).
- AFB analysis filter bank
- LAEC subband linear echo canceller
- RES subband residual echo suppressor
- SFB synthesis filter bank
- the near-end time domain microphone signal y(t) is transformed into the sub-band signal X(k,n) and the Y(k,n) input to the sub-band LAEC for linear echo cancellation, and the output signal E(k,n) After sub-band RES processing, it is transformed into time domain signal by SFB.
- the LAEC structure of subband k is shown in Fig. 5.
- the far reference signal subband spectrum at the n signal frame time in the k subband is X(k, n), and the microphone receive signal subband spectrum is Y(k, n).
- the near-end speech signal sub-band spectrum is S(k,n)
- the echo signal sub-band spectrum is D(k,n)
- the number of coefficients of the FIR filter is L s , and the output thereof
- Table 1 shows that the subband LAEC is on a fixed k subband, which is actually equivalent to a time domain LAEC (along the signal frame n-axis). Since each subband is mutually uncorrelated, for a given subband k, we can directly apply the adaptive algorithm for time-varying regularization factors discussed in the previous section to solve the FIR adaptation in subband k. Filtering learning problems. According to the corresponding relationship in Table 1, the calculation formula for the time-varying regularization factor in the easy-to-know LAEC should be corresponding to the formula forty-nine in the sub-band k:
- N s is the number of signal frames used in calculating the average
- time-varying regularization factor NLMS learning algorithm for updating the FIR filter coefficients in sub-band k can be expressed by the formula fifty-four and the formula fifty-five:
- ⁇ opt (k,n) is a time-varying regularization factor determined by the formula forty-nine; Can be defined by the formula fifty-six:
- the preset signal includes a subband spectrum of a speech signal received by a near-end microphone input in the AEC and a sub-band spectrum of a far-end reference speech signal
- the FIR filter coefficient vector is a sub-band FIR.
- the filter coefficient vector when the time-varying regularization factor is a sub-band time-varying regularization factor, the specific implementation manner of step 21 in the present disclosure is: acquiring the sub-band power spectrum and the sub-band coupling factor of the microphone receiving signal respectively
- the effective estimated value is obtained according to the subband power spectrum of the microphone received signal and the effective estimated value of the subband coupling factor, and is used to perform the subband domain when the subband FIR filter coefficient vector is used for the preset signal processing.
- the FIR filter coefficient vector iteratively updates the sub-band time-varying regularization factor.
- the subband power spectrum of the microphone receiving signal is obtained by:
- the manner of obtaining the effective estimated value of the sub-band coupling factor includes:
- the step of obtaining a biased estimation value of the sub-band coupling factor according to the cross-correlation method includes:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- the step of acquiring a correction factor for the biased estimation of the subband domain coupling factor comprises: obtaining a correlation coefficient amplitude between the AEC output error signal subband spectrum and the far reference signal subband spectrum An alternative value of the square; obtaining an AEC output error signal subband spectrum and a far reference signal subband according to an alternative value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far reference signal subband spectrum
- the effective magnitude of the correlation coefficient between the spectra is squared and used as a correction factor for the biased estimate of the sub-band coupling factor.
- the step of acquiring the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far reference signal subband spectrum includes:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the AEC output error signal subband spectrum and the far reference signal subband spectrum are obtained according to the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far end reference signal subband spectrum.
- the steps of squaring the effective magnitude of the correlation coefficient including:
- n is the signal frame time index variable.
- the step of obtaining a valid estimated value of the sub-band coupling factor according to the biased estimation value of the sub-band coupling factor and the correction factor including:
- n is the signal frame time index variable.
- the sub-band power spectrum and the effective estimation value of the sub-band coupling factor according to the received signal of the microphone are used to perform the sub-band domain when the sub-band FIR filter coefficient vector is used to process the preset signal.
- the FIR filter coefficient vector iteratively updates the sub-band time-varying regularization factor steps, including:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is the preset hour constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real Constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- step 22 when the FIR filter coefficient vector is continuously adaptively updated using the NLMS algorithm, the specific implementation manner of step 22 is:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the n signal frame time, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the n
- the first step initialization
- Equation 57 the time-varying regularization factor AP algorithm for updating the FIR filter coefficients in sub-band k can be expressed as Equation 57 and Equation 58 respectively:
- Receive a signal vector for a P-dimensional microphone It is an L ⁇ P-dimensional state matrix, and I P ⁇ P is a P ⁇ P-dimensional unit matrix.
- step 22 when the AP algorithm is used to continuously update the FIR filter coefficient vector, the specific implementation of step 22 is as follows:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone received signal at the time of the n-n3 signal frame;
- the first step initialization
- the time domain and subband LAEC time-varying regularization factor adaptive learning algorithm proposed by the present disclosure is not sensitive to the initialization process
- the LAEC time-varying regularization factor adaptive learning algorithm proposed by the present disclosure reduces the complexity of M 2 /K, where M is a decimation factor, and K is The total number of subbands); considering that the subband spectrum of the real signal satisfies the conjugate symmetry characteristic, the learning algorithm of the subband LAEC only needs to run in the first (K/2+1) subbands, so that the algorithm complexity can be further Reduced (about half); this low computational complexity subband LAEC learning algorithm is easy to implement on commercial DSP chips; in addition, the subband algorithm proposed by the present disclosure is easier to use in view of the parallelism of the subband processing structure. Integrated circuit (ASIC) implementation.
- ASIC Integrated circuit
- some embodiments of the present disclosure further provide a sustainable adaptive updating apparatus for a finite impulse response FIR filter coefficient vector, including a memory 61, a processor 62, and a memory 61.
- a computer program running on the processor 62; and the memory 61 is coupled to the processor 62 via a bus interface 63; wherein the processor 62 executes the computer program to implement the following steps: acquiring FIR filtering The time-varying regularization factor used to perform the iterative update of the FIR filter coefficient vector when the preset signal is processed; and the FIR filter coefficient vector is updated according to the time-varying regularization factor.
- the preset signal includes: a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone, a noise reference signal and a system input signal in an adaptive noise cancellation system, and adaptive interference cancellation.
- a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone
- a noise reference signal and a system input signal in an adaptive noise cancellation system
- adaptive interference cancellation One of a combination of the interference reference signal and the system input signal in the system, the excitation input signal in the adaptive system identification, and the unknown system output signal to be identified.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone.
- the processor 62 may further implement the following steps: respectively acquiring the microphone receiving signal An effective estimated value of the power and the coupling factor; according to the power of the microphone received signal and the effective estimated value of the coupling factor, the acquisition is performed by using the FIR filter coefficient vector for the preset signal processing to perform the iterative update of the FIR filter coefficient vector Time-varying regularization factor.
- the processor 62 can also implement the following steps:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- a valid estimate of the coupling factor a biased estimate of the coupling factor based on cross-correlation techniques;
- the effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the processor 62 can also implement the following steps:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0
- t is the digital signal sample time index number.
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- a conjugate transposed matrix of X state (t) I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of a far-end reference voice signal, where the FIR filter coefficient vector is a sub-band FIR filter coefficient.
- the vector, the time-varying regularization factor is a sub-band time-varying regularization factor, and the processor 62 can also implement the following steps when the computer program is executed by the processor 62:
- the subband domain FIR filter coefficient vector used for processing the preset signal by using the subband FIR filter coefficient vector Iteratively updated subband time-varying regularization factor.
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- the processor 62 can also implement the following steps:
- the processor 62 can also implement the following steps:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the processor 62 can also implement the following steps:
- n is the signal frame time index variable.
- the processor 62 can also implement the following steps:
- n is the signal frame time index variable.
- the processor 62 can also implement the following steps:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- the processor 62 can also implement the following steps:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the time of the n signal frame, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the
- the processor 62 can also implement the following steps:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- some embodiments of the present disclosure further provide a persistent adaptive updating apparatus 70 for a finite impulse response FIR filter coefficient vector, including: an obtaining module 71, configured to acquire a coefficient pair using FIR filter coefficients. And a time-varying regularization factor for performing the iterative update of the FIR filter coefficient vector when the signal processing is preset; the updating module 72 is configured to update the FIR filter coefficient vector according to the time-varying regularization factor.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone
- the acquiring module 71 includes: a first acquiring unit, configured to respectively acquire power and coupling of the microphone received signal a valid estimation value of the factor; a second obtaining unit, configured to obtain, according to the power of the microphone received signal and the effective estimated value of the coupling factor, the FIR filter for performing the processing on the preset signal by using the FIR filter coefficient vector The time-varying regularization factor of the coefficient vector iteratively updated.
- the acquiring manner of the power of the microphone receiving signal in the first acquiring unit is:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the obtaining manner of the effective estimated value of the coupling factor in the first acquiring unit is:
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- x(t-t1) is the far-end reference signal at the time of the t-t1 signal sample
- x * (t-t1) is Complex conjugate of x(t-t1)
- e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time
- e(t) is an error signal outputted by the AEC at the time t sample
- y(t) is a microphone received signal at the time t sample
- for Conjugate transposed matrix Is the far-end reference signal vector
- x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample
- T is the transpose operator
- the method for obtaining the square of the effective amplitude of the correlation coefficient between the error signal of the AEC output and the far-end reference signal according to the candidate value of the square of the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal for:
- a valid estimate of the coupling factor a biased estimate of the coupling factor based on cross-correlation techniques;
- the effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the second obtaining unit is configured to:
- ⁇ ppt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0; ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0;
- t is the digital signal sample time index number.
- update module 72 is configured to:
- update module 72 is configured to:
- a conjugate transposed matrix of X state (t) I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of a far-end reference voice signal, where the FIR filter coefficient vector is a sub-band FIR filter coefficient.
- the time-varying regularization factor is a sub-band time-varying regularization factor
- the obtaining module 71 includes: a third acquiring unit, configured to respectively acquire a sub-band power spectrum of the microphone received signal and an effective sub-band coupling factor
- An estimation unit configured to perform, according to the sub-band power spectrum of the microphone received signal and the effective estimated value of the sub-band coupling factor, to obtain the sub-band FIR filter coefficient vector for performing preset signal processing
- the sub-band FIR filter coefficient vector iteratively updates the sub-band time-varying regularization factor.
- the acquiring manner of the subband power spectrum of the microphone receiving signal in the third acquiring unit is:
- the obtaining manner of the effective estimated value of the sub-band coupling factor in the third acquiring unit includes:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- the AEC output error signal subband spectrum and the far reference signal subband spectrum are obtained according to the candidate value of the square of the correlation coefficient between the AEC output error signal subband spectrum and the far end reference signal subband spectrum.
- the way in which the effective magnitude of the correlation coefficient is squared is:
- n is the signal frame time index variable.
- n is the signal frame time index variable.
- the fourth obtaining unit is configured to:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- update module 72 is configured to:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the n signal frame time, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the n
- update module 72 is configured to:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- the embodiment of the device is a device corresponding to the foregoing method embodiment. All the implementation manners in the foregoing method embodiments are applicable to the embodiment of the device, and the same technical effects can be achieved.
- Some embodiments of the present disclosure also provide a computer readable storage medium having stored thereon a computer program, the computer program being executed by a processor to implement the following steps: acquiring a preset signal processing using a FIR filter coefficient vector And performing a time-varying regularization factor of the FIR filter coefficient vector iterative update; updating the FIR filter coefficient vector according to the time-varying regularization factor.
- the preset signal includes: a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone, a noise reference signal and a system input signal in an adaptive noise cancellation system, and adaptive interference cancellation.
- a far-end reference voice signal input in an acoustic echo canceller AEC and a voice signal received by a near-end microphone
- a noise reference signal and a system input signal in an adaptive noise cancellation system
- adaptive interference cancellation One of a combination of the interference reference signal and the system input signal in the system, the excitation input signal in the adaptive system identification, and the unknown system output signal to be identified.
- the preset signal includes a far-end reference voice signal input in the AEC and a voice signal received by the near-end microphone.
- the processor may further implement the following steps: respectively acquiring the microphone receiving signal Effective estimates of power and coupling factors;
- the processor may further implement the following steps: according to the formula:
- the power of the signal received by the microphone is the microphone received signal; y(t) is the microphone received signal; ⁇ a and ⁇ d are preset recursive constants, and 0 ⁇ ⁇ a ⁇ ⁇ d ⁇ 1; t is the digital signal time index number.
- the processor may further implement the steps of: obtaining a biased estimation value of the coupling factor according to the cross-correlation method; and acquiring a correction factor for the biased estimation value compensation of the coupling factor; Obtaining a valid estimate of the coupling factor based on the biased estimate of the coupling factor and the correction factor.
- the processor may further implement the following steps: According to the formula: Obtaining a biased estimate of the coupling factor;
- T s L is the number of FIR filter coefficients; e(t-t1) is the error signal of the AEC output at the t-t1 signal sample time, e(t) is an error signal outputted by the AEC at the time t sample; y(t) is a microphone received signal at the time t sample; for Conjugate transposed matrix; Is the far-end reference signal vector, and x(t-t2) is the far-end reference signal at the time of the t-t2 signal sample; T is the transpose operator; Is the FIR filter coefficient vector, w t2 (t) is the t2+1th coefficient of the FIR filter at the time t sample, t
- the processor may further implement the following steps: acquiring an alternative value of the square of the correlation coefficient between the error signal outputted by the AEC and the remote reference signal; and the error signal output according to the AEC An alternative value of the square of the correlation coefficient between the remote reference signal and the square of the effective amplitude of the correlation coefficient between the error signal of the AEC output and the far-end reference signal, and this is used as a correction for the biased estimation of the coupling factor factor.
- the processor may further implement the following steps: according to the formula:
- the processor may further implement the following steps: according to the formula: Obtain a valid estimate of the coupling factor; a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques; The effective amplitude squared the correlation coefficient between the error signal outputted by the AEC and the far-end reference signal; t is the digital signal sample time index number.
- the processor may further implement the following steps: according to the formula:
- ⁇ opt (t) is a time-varying regularization factor
- L is the number of FIR filter coefficients
- the power of the signal received by the microphone a valid estimate of the coupling factor; a biased estimate of the coupling factor based on cross-correlation techniques;
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small real constants, respectively, and ⁇ >0, ⁇ 0 >0
- t is the digital signal sample time index number.
- processor may further implement the following steps when the computer program is executed by the processor:
- the processor may further implement the following steps: according to the formula:
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error vector;
- Receive a signal vector for a P-dimensional microphone, and y(t-t3) is the microphone receiving signal at the time of the t-t3 signal sample;
- t is the digital signal
- the preset signal includes a subband spectrum of a voice signal received by a near-end microphone input in the AEC and a subband spectrum of the far-end reference voice signal
- the FIR filter coefficient vector is a sub-band FIR filter coefficient.
- the vector, the time-varying regularization factor is a sub-band time-varying regularization factor, and the processor may further implement the following steps when the computer program is executed by the processor:
- the subband domain FIR filter coefficient vector used for processing the preset signal by using the subband FIR filter coefficient vector Iteratively updated subband time-varying regularization factor.
- processor may further implement the following steps when the computer program is executed by the processor:
- processor may further implement the following steps when the computer program is executed by the processor:
- processor may further implement the following steps when the computer program is executed by the processor:
- E(k, n-n1) is the AEC at the time of the n-n1 signal frame Output error signal subband spectrum
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time
- Y(k,n) is the microphone received signal subband spectrum at the n signal frame time; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame
- T is the transpose operator
- processor may further implement the following steps when the computer program is executed by the processor:
- processor may further implement the following steps when the computer program is executed by the processor:
- E(k, n-n1) is the AEC output at the n-n1 signal frame time Error signal subband spectrum;
- E(k,n) is the AEC output error signal subband spectrum at the n signal frame time;
- Y(k,n) is the microphone received signal subband spectrum; for Conjugate transposed matrix; a spectral vector for the far-end reference signal subband, and X(k, n-n2) is the far-end reference signal subband spectrum at the time of the n-n2 signal frame;
- T is the transpose
- processor may further implement the following steps when the computer program is executed by the processor:
- n is the signal frame time index variable.
- processor may further implement the following steps when the computer program is executed by the processor:
- n is the signal frame time index variable.
- processor may further implement the following steps when the computer program is executed by the processor:
- ⁇ opt (k, n) is a sub-band time-varying regularization factor
- Subband power spectrum for receiving signals for the microphone a valid estimate of the subband coupling factor, a biased estimate of the subband coupling factor
- ⁇ min is a preset small real constant, and ⁇ min >0
- ⁇ 0 and ⁇ are preset small Real constant, and ⁇ >0, ⁇ 0 >0
- n is the signal frame time index variable.
- processor may further implement the following steps when the computer program is executed by the processor:
- ⁇ is a predetermined coefficient update step parameter, and 0 ⁇ ⁇ ⁇ 2;
- X(k,n-n2) is the far-end reference signal subband spectrum at the n-n2 signal frame time;
- n2 0,1,...,L s -1,L s is the FIR filter coefficient in each subband Number
- T is the transpose operator;
- E * (k, n) is a complex conjugate of E(k, n);
- E(k, n) is an AEC output error signal subband spectrum at the time of the n signal frame, and
- Y(k,n) is the subband spectrum of the microphone received signal at the time of the n signal frame;
- W n2 (k,n) is the n2+1th coefficient of the FIR filter in subband k at the
- processor may further implement the following steps when the computer program is executed by the processor:
- the affine projection AP algorithm is used to continuously update the sub-band FIR filter coefficient vector
- ⁇ is the predetermined coefficient update step parameter, and 0 ⁇ 2;
- ⁇ opt (k,n) is the sub-band time-varying regularization factor;
- X state (k, n) is an L ⁇ P dimensional state matrix in subband k, and
- I P ⁇ P is a P ⁇ P dimension unit matrix; for Complex conjugate, and a P-dimensional error signal subband spectral vector;
- Receiving a signal subband spectral vector for a P-dimensional microphone, and Y(k, n-n3) is a microphone receiving signal at the time of the n-n3 signal frame;
- W n2 (k,n) is the n
- the computer readable storage medium referred to by this disclosure may be volatile or non-volatile, transient or non-transitory.
Landscapes
- Engineering & Computer Science (AREA)
- Signal Processing (AREA)
- Physics & Mathematics (AREA)
- Acoustics & Sound (AREA)
- Health & Medical Sciences (AREA)
- Audiology, Speech & Language Pathology (AREA)
- Computational Linguistics (AREA)
- Quality & Reliability (AREA)
- Human Computer Interaction (AREA)
- Multimedia (AREA)
- Otolaryngology (AREA)
- General Health & Medical Sciences (AREA)
- Circuit For Audible Band Transducer (AREA)
- Filters That Use Time-Delay Elements (AREA)
- Cable Transmission Systems, Equalization Of Radio And Reduction Of Echo (AREA)
Abstract
一种有限冲激响应滤波器系数矢量的可持续更新方法及装置。有限冲激响应FIR滤波器系数矢量的可持续自适应更新方法,包括:获取采用FIR滤波器系数矢量对预设信号处理时用于进行FIR滤波器系数矢量迭代更新的时变正则化因子(21);根据时变正则化因子,对FIR滤波器系数矢量进行更新(22)。
Description
相关申请的交叉引用
本申请主张在2017年9月7日在中国提交的中国专利申请号No.201710800778.4的优先权,其全部内容通过引用包含于此。
本公开涉及信号处理技术领域,特别涉及有限冲激响应滤波器系数矢量的可持续更新方法及有限冲激响应滤波器系数矢量的可持续更新装置。
声学回声抵消器(AEC)是全双工语音通信系统中一个关键部件,其主要作用是移去由扬声器(喇叭)耦合到麦克风(话筒)中的远端信号的回声信号。在AEC中,用一个有限冲激响应(FIR)线性滤波器对回声路径进行自适应地学习建模,并由此合成一个回波的有效估计,然后在麦克风的接收信号中减去该估计,从而完成回波抵消的任务。当近端发话语音信号出现时,由于它与远端语音信号统计上不相关,因而其行为犹如一个突发的噪声,使得这一滤波器的系数更新将偏离实际回波路径所对应的真值而发生发散现象,这便相应地增大了回波残留量,使AEC的性能恶化。鉴于此,人们首先需要用“双讲”检测器(DTD)对麦克风接收信号中是否含有近端发话语音信号(即“双讲”情形)进行及时而准确地检测,在麦克风接收信号中无近端发话语音信号(即“单讲”)情况下,线性滤波器系数的自适应学习持续进行;而在麦克风接收信号中含有近端发话语音信号(即“双讲”)时,线性滤波器系数的自适应学习必须停止进行,以避免在该情况下滤波器系数持续学习所致的发散现象。然而DTD的处理延时及其可能的误判均会严重地影响AEC滤波器的自适应学习行为,进而影响AEC的性能。为此,人们提出许多采用自适应 变步长的学习技术对FIR滤波器的系数进行持续的迭代更新。
自适应变步长类学习技术的缺点是对其参数初始化十分敏感;此外,该类技术不适用处理回声抵消应用中有近端(非平稳)发话语音的“双讲”情况。而相关技术中的其他方案,例如广义归一化梯度下降(GNGD)算法,也不能很好地处理有近端(非平稳)发话语音的“双讲”情况。
发明内容
本公开的一些实施例提供一种有限冲激响应滤波器系数矢量的可持续更新方法及装置,以解决相关的FIR滤波器的自适应学习方式无法保证FIR滤波器性能的稳定性,进而影响信号处理可靠性的问题。
第一方面,本公开的一些实施例提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新方法,包括:
获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;
根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
可选地,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
可选地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,所述获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:
分别获取麦克风接收信号的功率和耦合因子的有效估计值;
根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更 新的时变正则化因子。
可选地,所述麦克风接收信号的功率的获取方式为:
根据公式:
得到麦克风接收信号的功率;
可选地,所述耦合因子的有效估计值的获取方式为:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
可选地,所述根据互相关方法获取耦合因子的有偏估计值的步骤,包括:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的步骤,包括:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
进一步地,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的步骤,包括:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方的步骤,包括:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
可选地,所述根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值的步骤,包括:
可选地,所述根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:
根据公式:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样本时间索引序号。
可选地,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为X
state(t)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
具体地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,所述获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:
分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;
根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
可选地,所述麦克风接收信号的子带功率谱的获取方式为:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述子带域耦合因子的有效估计值的获取方式包括:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦 合因子的有效估计值。
可选地,所述根据互相关方法获取子带域耦合因子的有偏估计值的步骤,包括:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的步骤,包括:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
可选地,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的步骤包括:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方的步骤,包括:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
可选地,所述根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值的步骤,包括:
可选地,所述根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子的步骤,包括:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
可选地,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进 行更新的步骤,包括:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维 状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
第二方面,本公开的一些实施例还提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置,包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序;其中,所述处理器执行所述计算机程序时实现以下步骤:
获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;
根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
具体地,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
可选地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,计算机程序被处理器执行时所述处理器还可实现如下步骤:
分别获取麦克风接收信号的功率和耦合因子的有效估计值;
根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
得到麦克风接收信号的功率;
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样本时间索引序号。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR 滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为X
state(t)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,计算机程 序被处理器执行时所述处理器还可实现如下步骤:
分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;
根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个 数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧 时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
第三方面,本公开的一些实施例还提供一种计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时所述处理器实现以下步骤:
获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;
根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
第四方面,本公开的一些实施例还提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置,包括:
获取模块,用于获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;
更新模块,用于根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
可选地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,所述获取模块包括:
第一获取单元,用于分别获取麦克风接收信号的功率和耦合因子的有效估计值;
第二获取单元,用于根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
可选地,所述第一获取单元中所述麦克风接收信号的功率的获取方式为:
根据公式:
得到麦克风接收信号的功率;
可选地,所述第一获取单元中所述耦合因子的有效估计值的获取方式为:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
可选地,所述根据互相关方法获取耦合因子的有偏估计值的方式为:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的方式为:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
可选地,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的方式为:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号 样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方的方式为:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
可选地,所述根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值的方式为:
可选地,所述第二获取单元用于:
根据公式:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样本时间索引序号。
可选地,所述更新模块用于:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述更新模块用于:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为X
state(t)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
可选地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,所述获取模块包括:
第三获取单元,用于分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;
第四获取单元,用于根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
可选地,所述第三获取单元中所述麦克风接收信号的子带功率谱的获取方式为:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述第三获取单元中所述子带域耦合因子的有效估计值的获取方式包括:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
可选地,所述根据互相关方法获取子带域耦合因子的有偏估计值的方式为:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的方式为:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
可选地,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的方式为:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数 矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方的方式为:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
可选地,所述根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值的方式为:
可选地,所述第四获取单元用于:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子 带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
可选地,所述更新模块用于:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
可选地,所述更新模块用于:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
本公开的有益效果是:
上述方案,通过获取用于进行FIR滤波器系数矢量迭代更新的时变正则化因子,依据该时变正则化因子对FIR滤波器系数矢量进行持续自适应更新,以此,保证了FIR滤波器性能的稳定性,提高了信号处理的可靠性。
图1表示AEC的工作原理示意图;
图2表示本公开的一些实施例的FIR滤波器系数矢量的可持续自适应更新方法的流程图;
图3表示时域LAEC的结构示意图;
图4表示子带域AEC的结构示意图;
图5表示子带k的LAEC结构示意图;
图6表示本公开的一些实施例的FIR滤波器系数矢量的可持续自适应更新装置的结构示意图;以及
图7表示本公开的一些实施例的FIR滤波器系数矢量的可持续自适应更新装置的模块示意图。
为使本公开的目的、技术方案和优点更加清楚,下面将结合附图及具体实施例对本公开进行详细描述。
AEC的工作原理如图1所示,它通常包含线性回声抵消器(LAEC)、“双讲”检测器(DTD)和非线性残留回声抑制器(RES)3个处理模块,LAEC利用FIR线性滤波器和下行链路信号(即参考信号,用x(t)表示)来对回声信号(用d(t)表示)进行估计,然后把该估计(用
表示)从麦克风接收信号(用y(t)表示)中减去;FIR滤波器的系数通常仅在“单讲”(即无近端发话信号(用s(t)表示))情况下由误差信号(用e(t)表示)按回声消除(NLMS)算法来自适应地更新,用公式一表示为:
在“双讲”(即有近端发话信号(s(t)))的情况下,FIR滤波器的系数更新将被冻结以避免NLMS学习算法发散;“双讲”的检测由DTD来完成。由于 下行链路扬声器和上行链路麦克风的非线性,使得LAEC仅能抵消麦克风接收信号(y(t))中回声信号的线性分量,而其残留的非线性分量将由后续的残留回声抑制(RES)模块来做抑制处理。
上述指出,AEC的FIR滤波器系数自适应更新必须仅在无近端发话语音信号的条件下(即“单讲”模式)进行;在有近端发话语音信号的条件下(即“双讲”模式),其系数的自适应更行必须停止以免滤波器系数的发散而造成对近端发话语音信号的损伤。“双讲”模式的检测通常由DTD来完成,然而DTD的处理延时及其可能的误判均会严重地影响AEC滤波器的自适应学习行为,进而影响AEC的性能。为此,人们提出许多采用自适应变步长的学习技术对FIR滤波器的系数进行持续的迭代更新,这类技术均使用一个学习率变量(即变步长参数),该学习率变量是根据AEC的代价函数对其梯度来自适应调节的,其缺点是对算法相关参数的初始化十分敏感,并且仅适用于平稳输入信号,因而在实际应用特别是回声抵消的应用中难以获得良好的性能。于是人们又提出了一种称之为广义归一化梯度下降(GNGD)算法以克服上述的不足,如公式三至公式五所示:
其中,η(t)为学习率;δ(t)为学习率控制参数;ρ
1为δ(t)的自适应迭代步长参数。
尽管GNGD算法对其相关参数的初始化不敏感,但其学习率(η(t))的控制参数(δ(t))是基于NLMS随机梯度的行为特征来自适应调节的,因而它仅能进行相对较慢的自适应调节(通常需要数十或数百个样本)。正因如此,GNGD算法也不能很好地处理有近端(非平稳)发话语音的“双讲”情况。
鉴于上述情况,本公开针对相关的FIR滤波器的自适应学习方式无法保 证FIR滤波器性能的稳定性,影响信号处理可靠性的问题,提供一种有限冲激响应滤波器系数矢量的可持续更新方法及装置。
如图2所示,本公开的一些实施例提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新方法,包括步骤21-22。
步骤21,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
需要说明的是,该预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
步骤22,根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
上述方案,通过获取用于进行FIR滤波器系数矢量迭代更新的时变正则化因子,依据该时变正则化因子对FIR滤波器系数矢量进行持续自适应更新,以此,保证了FIR滤波器性能的稳定性,提高了信号处理的可靠性。
下面以对AEC中输入的远端参考语音信号和近端麦克风接收的语音信号进行处理为例,分别从时域和子带域出发,对本公开说明如下。
一、时域
对图2所示的时域LAEC,本公开主要讨论其对“双讲”和回声路径变化具有良好鲁棒性(Robustness)的持续性自适应学习问题,提出了一种用于时域LAEC自适应学习的时变正则化因子实时在线计算方法,用该时变正则化因子构成的LAEC中FIR滤波器系数矢量自适应迭代更新算法对“双讲”和回声路径变化具有良好鲁棒性。首先以时域LAEC自适应学习的归一化最小均方(NLMS)算法为例,来详细阐述实时在线求解该时变正则化因子的方法,然后将该时变正则化因子应用到LAEC持续性自适应学习中,并分别给出了 一种时变正则化因子的NLMS算法和一种时变正则化因子的仿射投影(AP)算法。
其中,w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;L为FIR滤波器系数的个数;T为转置运算符;t为数字信号样本时间索引序号。
那么FIR滤波器系数更新的NLMS学习算法可用公式七和公式八表述为:
其中,e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,符号*代表取复共轭运算;H为复共轭转置运算符;0<μ<2为预定的系数更新步长参数,δ>0为一个称为“正则化因子”的小正实常数以避免公式八的分母为零。
其中,x(t-t2)为在t-t2信号样本时刻的远端参考信号。
y(t)由公式十定义:y(t)=d(t)+s(t);
其中,d(t)为t时刻回声信号;s(t)为t时刻近端发话语音信号。
鉴于自适应滤波器系数的迭代步长可以通过一个“时变”正则化因子δ(t)参数来加以控制,那么这一“时变”正则化因子参数δ(t)必须能够实时在线地计算求解出来。事实上,设子FIR滤波器理想复系数矢量为
其估计为
则滤波器系统适配矢量利用公式十一定义为:
那么LAEC的输出的误差e(t)可由公式十二表达为:
那么δ(t)的选择必须使得下述变量最小化,即有公式十三:
假设s(t)与x(t)不相关,则公式十五两边取统计平均可得公式十七:
进一步地,定义公式十八和公式十九:
置公式二十为零,则得到最优的正则化因子δ
opt(t)用公式二十一表示为:
进一步假设L>>1,并且x(t)为白化的激励信号(如不是的话,可通过白化处理将它变为白化的激励源),那么有如下公式:
不妨取μ=1,则公式二十一可用公式二十四表示为:
作为一个滤波器系数矢量更新迭代步长的控制变量δ
opt(t),可用公式二十五估计得到,虽然在有回声的情况下略微地过于抑制(over-suppress)了滤波器系数的自适应迭代,但在较高电平的近端发话语音信号出现时,仍能达到减慢滤波器系数矢量更新迭代的期望效果。
的工程化计算可以通过一阶递归模型来实时在线地实现,考虑到
必须及时地跟踪近端发话语音信号电平的变化,那么本公开用下述“快升慢降”(fast attack/slow decay)的方式实时在线地估计变量
即有公式二十六:
其中,其中0≤α
a<α
d<1为预设的递归常数。
尽管公式二十八利用近端信号s(t)与参考信号x(t)间的统计不相关性来去除近端语音s(t)的影响,但这种估计是一个有偏估计,因而存在偏差。为此,我们需要对其作修正补偿,以提高耦合因子β(t)的估计精度。
其中r
ex(t)为LAEC输出信号e(t)与远端参考信号x(n)之间在“单讲”情况下的(归一化)相关系数。公式二十九表明:在“单讲”情况下,用互相关技术估计的β(t)|
互相关法与用直接法估计的β(t)|
直接法之比正好是时域LAEC输出信号e(t)与远端参考信号x(t)之间的(归一化)相关系数幅度平方。如果我们能有效地估计出这个相关系数幅度平方,用它来修正补偿互相关法估计的β(t)|
互相关法,便可提高耦合因子β(t)的估计精度,即
的估计可用公式三十表达为:
显然,用于估计耦合因子的公式三十不仅适用于“单讲”情况,而且也适用于“双讲”情况,只要能有效地估计出e(t)与x(t)之间的相关系数幅度平方。如何正确而有效地估计出该相关系数幅度平方,可通过下述的分析讨论来获得相关的启示。注意到这样一个事实:在“单讲”情况下,|r
ex(t)|
2可以按二十九中的定义来计算,即有公式三十一:
而在“双讲”情况下,记此时相关系数幅度平方的备选值为r(t),则r(t)可用公式三十二表示为:
式中T
s<<L为一个正整数,它是用来估计公式三十四至公式三十五所使用的样本数目。公式三十、公式三十三至公式三十五构成了一种既可用于“双讲”场景又可用于回声路径变化场景的
的高精度估计算法,用该算法求得的
估计值代入公式二十五便得到公式三十六:
其中ρ和ρ
0>0为很小的实常数以避免公式三十六的分母为零,
从公式三十六可以看出:在“单讲”场景下,
若此时发生回声路径的变化情况,那么
会迅速增大进而导致
相应地
也会迅速增大(因为此时残留回声信号∈(t)与远端参考信号x(t)强相关),而
相对而言变化不是太大,因此导致δ
opt(t)变小,有利于加速子带FIR滤波器系数矢量的更新迭代;在“双讲”场景下,
近端发话语音信号的
出现,使得
迅速变大,而
和
的估计(参见公式三十三和公式三十四)对“双讲”具有鲁棒性,所以其估计的参数值变化不大,因而导致δ
opt(t)迅速变大,从而极大地减慢了滤波器系数矢量更新迭代,避免了滤波器系数矢量更新在“双讲” 情况下发生发散现象。
另一方面,在近端发话语音信号电平较低或无近端发话语音信号的条件下,为使滤波器系数矢量更新迭代算法稳定,其正则化因子必须限定为一个小实常数δ
min>0。因此一个正则化因子的合理在线计算式如公式三十七所示:
分别获取麦克风接收信号的功率和耦合因子的有效估计值;
根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
进一步地,所述麦克风接收信号的功率的获取方式为:
根据公式二十六:
得到麦克风接收信号的功率;
进一步地,所述耦合因子的有效估计值的获取方式为:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
进一步地,所述根据互相关方法获取耦合因子的有偏估计值的步骤,包括:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1。
进一步地,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的步骤,包括:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
进一步地,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的步骤,包括:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为 x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<< L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1。
进一步地,所述根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方的步骤,包括:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
进一步地,所述根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值的步骤,包括:
进一步地,所述根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于FIR滤波器系数矢 量迭代更新的时变正则化因子的步骤,包括:
根据公式三十六和公式三十七:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0。
具体地,在使用NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新时,步骤22的具体实现方式为:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1。
将公式三十七定义的时变正则化因子δ
opt(t)应用到时域LAEC的NLMS自适应学习算法中,可得到一种时域LAEC带时变正则化因子的NLMS持续 自适应学习算法,该算法对“双讲”和回声路径变化具有良好的鲁棒性;具体的算法实现流程为:
第一步:初始化
1.1、预置参数0<μ<2、0≤α
a<α
d<1、τ≥0、ρ>0、
δ
min、L和T
s的值;
1.2、初始化相关变量
1.3、设置信号时间索引变量t为零,即t=0;
第二步:在线计算相关变量
2.2、用公式七计算LAEC的输出e(t);
3.1、用公式三十七计算δ
opt(t);
第四步:更新迭代FIR滤波器系数矢量
第五步:更新信号帧索引变量t,即:t=t+1,并跳转至第二步。
那么根据公式十一定义的系统适配矢量,存在公式四十,即:
根据公式三十九容易得到公式四十三,即:
假设参考信号x(t)为白化激励源,并且LAEC的FIR滤波器系数的个数L很大即L>>1,那么公式四十三中的逆矩阵可近似由公式四十四表示:
因而有公式四十六:
对比公式四十七与公式二十四,即知由公式三十七确定的最优时变正则化因子同样适用于AP算法;具体地,在使用AP算法对所述FIR滤波器系数 矢量进行持续的自适应更新时,步骤22的具体实现方式为:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为
的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
将公式三十七定义的时变正则化因子δ
opt(t)应用到时域LAEC的AP自适应学习算法中,可得到一种时域LAEC带时变正则化因子的AP持续自适应学习算法,该算法对“双讲”和回声路径变化具有良好的鲁棒性;具体的算法实现流程为:
第一步:初始化
1.1、预置参数0<μ<2、0≤α
a<α
d<1、τ≥0、ρ>0、
δ
min、L和T
s的值;
1.2、初始化相关变量
1.3、设置信号时间索引变量t为零,即t=0;
第二步:在线计算相关变量
2.2、用公式七计算LAEC的输出e(t);
第三步:在线计算正则化因子δ
opt(t)
3.1、用公式三十七计算δ
opt(t);
第四步:更新迭代FIR滤波器系数矢量
第五步:更新信号帧索引变量t,即:t=t+1,并跳转至第二步。
二、子带域
众所周知,声学路径通常随时间变化较快,由于语言信号的相关性,致使时域LAEC的NLMS学习算法将收敛很慢,因而不能实时地跟踪其回声路径的变化;此外,在某些应用场合下(例如免提扬声电话),声学回声的路径通常较长,这便加剧了上述的缺陷。尽管时域递归最小二乘(RLS)算法具有很快的收敛速度,但因其计算复杂度很大而无法在目前的商用DSP上实现。时域仿射投影(AP)算法在算法收敛速度和其计算复杂度方面进行了一定程度上的折中,但在仿射投影阶数P值较大(以便使AP算法收敛加快)的情况下,其计算复杂度也难以在目前的商用DSP上实现。这便促使人们研究和探索AEC的频域实现技术,特别是子带域的现技术。事实上,在子带域中,语音信号和声学回波路径传递函数的某些特性(例如每个子带中回波路径传递函数系数个数及其权值更新率的降低、子带信号动态范围相对其时域信号而言变小等)将有益于自适应学习算法对声学回波路径传递函数进行快速估计和实时跟踪。下面所叙述的内容是将上述讨论的时域LAEC时变正则化因 子自适应学习算法的思想应用到子带域LAEC的自适应学习问题,提出了子带域LAEC自适应学习的时变正则化因子NLMS算法和AP算法。
子带域AEC的结构如图4所示,它由分析滤波器组(AFB)、子带域线性回声抵消器(LAEC)、子带域残留回声抑制器(RES)和合成滤波器组(SFB)构成,其中t、n和k分别表示数字信号样本时间、信号帧时间索引变量和子带索引序号(子带索引序号变量);AFB首先分别把来自远端的参时域考信号x(t)和近端的时域麦克风信号y(t)变换成子带域信号X(k,n)和Y(k,n)输入给子带域LAEC进行线性回声抵消,其输出信号E(k,n)经子带域RES处理后再由SFB变换成时域信号
子带k的LAEC结构如图5所示,其中k子带中在n信号帧时刻的远端参考信号子带谱为X(k,n)、麦克风接收信号子带谱为Y(k,n)、近端语音信号子带谱为S(k,n)、回声信号子带谱为D(k,n),FIR滤波器的系数个数为L
s,其输出
为回声信号D(k,n)的一个线性估计值,k子带中在n信号帧时刻的FIR滤波器系数矢量由公式四十八表示为:
其中,k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数。
对比图3所示的时域LAEC结构和图5所示的子带域LAEC结构,可以得到如表1表示的对应关系:
表1时域LAEC和子带域LAEC的参数对应关系表
表1表明,子带域LAEC在一个固定的k子带上,它实际上等价于一个时域LAEC(沿着信号帧n轴向)。由于每个子带之间是互不相关的,那么对于一个给定的子带k,我们可以直接应用上节讨论的有关时变正则化因子的自适应算法来解决子带k中的FIR自适应滤波的学习问题。根据表1的对应关系,易知时域LAEC中的时变正则化因子计算公式三十七在子带k中应对应于公式四十九:
其中,
和
分别为子带功率谱、子带域耦合因子以及子带域LAEC输出信号E(k,t)与远端参考信号X(k,t)之间的幅度平方相干系数(MSC),它们分别由时域LAEC中的相关公式二十六、公式 三十三至公式三十五作变量替换而得,分别由以下公式表达:
公式五十:
这里0≤α
a<α
d<1为预设的递归常数。
其中,N
s为计算平均时使用的信号帧数目,且N
s<<L
s。
那么子带k中的FIR滤波器系数更新的时变正则化因子NLMS学习算法可用公式五十四和公式五十五表述为:
有上述推导可知,在所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子时,本公开中的步骤21的具体实现方式为:分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子 带域时变正则化因子。
进一步地,所述麦克风接收信号的子带功率谱的获取方式为:
根据公式五十:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述子带域耦合因子的有效估计值的获取方式包括:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
进一步地,所述根据互相关方法获取子带域耦合因子的有偏估计值的步骤,包括:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的 共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的步骤,包括:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
进一步地,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的步骤包括:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻 的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方的步骤,包括:
根据公式五十三:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
进一步地,所述根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值的步骤,包括:
进一步地,所述根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子的步骤,包括:
根据公式四十九:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小时常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
具体地,在使用NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新时,步骤22的具体实现方式为:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
将公式四十九定义的时变正则化因子δ
opt(k,n)应用到子带域LAEC的NLMS自适应学习算法中,可得到一种子带域LAEC带时变正则化因子的NLMS持续自适应学习算法,该算法对“双讲”和回声路径变化具有良好的鲁棒性;具体的算法实现流程为:
第一步:初始化
1.1、预置参数0<μ<2、0≤α
a<α
d<1、τ≥0、ρ>0、
δ
min、L
S和N
S的值;
1.2、初始化相关变量,对所有子带k,k=0,...,K-1,做如下初始化:
1.3、设置信号帧索引变量n为零,即n=0;
第二步:所有子带k,k=0,...,K-1,做如下处理:
2.1、在线计算相关变量
2.1.2、用公式五十四计算LAEC的输出E(k,n);
2.2、在线计算正则化因子δ
opt(k,n)
2.2.1、用公式四十九计算δ
opt(k,n);
2.3、更新迭代子带域FIR滤波器系数矢量
第三步:更新信号帧索引变量n,即:n=n+1,并跳转至第二步。
具体地,在使用AP算法对所述FIR滤波器系数矢量进行持续的自适应更新时,步骤22的具体实现方式为:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3 信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
将公式四十九定义的时变正则化因子δ
opt(k,n)应用到子带域LAEC的AP自适应学习算法中,可得到一种子带域LAEC带时变正则化因子的AP持续自适应学习算法,该算法对“双讲”和回声路径变化具有良好的鲁棒性;具体的算法实现流程为:
第一步:初始化
1.1、预置参数0<μ<2、0≤α
a<α
d<1、τ≥0、ρ>0、
δ
min、L
S和N
S的值;
1.2、初始化相关变量,对所有子带k,k=0,...,K-1,做如下初始化:
1.3、设置信号帧索引变量n为零,即n=0;
第二步:所有子带k,k=0,...,K-1,做如下处理:
2.1、在线计算相关变量
2.1.2、用公式五十四计算LAEC的输出E(k,n);
2.2、在线计算正则化因子δ
opt(k,n)
2.2.1、用公式四十九计算δ
opt(k,n);
2.3、更新迭代子带域FIR滤波器系数矢量
第三步:更新信号帧索引变量n,即:n=n+1,并跳转至第二步。
需要说明的是,本公开的一些实施例具有如下优点:
A、本公开提出的时域和子带域LAEC时变正则化因子自适应学习算法对初始化过程不敏感;
B、本公开提出的时域和子带域LAEC时变正则化因子自适应学习算法对回声抵消应用中的“双讲”模式具有很好的鲁棒性;
C、本公开提出的时域和子带域LAEC时变正则化因子自适应学习算法对回声抵消应用中的“回声路径变化”具有很好的鲁棒性;
D、本公开提出的子带域LAEC时变正则化因子自适应学习算法与相关相应的时域LAEC学习算法相比,其复杂度降低了M
2/K,(这里M为抽取因子,K为子带总数);考虑到实信号的子带谱满足共轭对称特性,那么子带域LAEC的学习算法仅需运行于前(K/2+1)个子带,从而其算法复杂度可进一步地降低(约减少一半);这种低计算复杂度子带域LAEC学习算法易于在商用DSP芯片上实现;此外,鉴于子带域处理结构的并行性,本公开提出的子带域算法更易于专用集成电路(ASIC)实现。
如图6所示,本公开的一些实施例还提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置,包括存储器61、处理器62及存储在所述存储器61上并可在所述处理器62上运行的计算机程序;且所述存储器61通过总线接口63与所述处理器62连接;其中,所述处理器62执行所述计算机程序时实现以下步骤:获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
具体地,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号 和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
进一步地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,计算机程序被处理器62执行时处理器62还可实现如下步骤:分别获取麦克风接收信号的功率和耦合因子的有效估计值;根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
得到麦克风接收信号的功率;
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计 时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为 转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样 本时间索引序号。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为X
state(t)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦 克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
具体地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,计算机程序被处理器62执行时处理器62还可实现如下步骤:
分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;
根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅 度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器62执行时处理器62还可实现如下步骤:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
本领域技术人员可以理解,实现上述实施例的全部或者部分步骤可以通 过硬件来完成,也可以通过计算机程序来指示相关的硬件来完成,所述计算机程序包括执行上述方法的部分或者全部步骤的指令;且该计算机程序可以存储于一可读存储介质中,存储介质可以是任何形式的存储介质。
如图7所示,本公开的一些实施例还提供一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置70,包括:获取模块71,用于获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;更新模块72,用于根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
具体地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,所述获取模块71包括:第一获取单元,用于分别获取麦克风接收信号的功率和耦合因子的有效估计值;第二获取单元,用于根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
进一步地,所述第一获取单元中所述麦克风接收信号的功率的获取方式为:
根据公式:
得到麦克风接收信号的功率;
进一步地,所述第一获取单元中所述耦合因子的有效估计值的获取方式为:
根据互相关方法获取耦合因子的有偏估计值;
获取用于所述耦合因子的有偏估计值补偿的修正因子;
根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
进一步地,所述根据互相关方法获取耦合因子的有偏估计值的方式为:
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的方式为:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;
根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
进一步地,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的方式为:
其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,所述根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方的方式为:
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
进一步地,所述根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值的方式为:
进一步地,所述第二获取单元用于:
根据公式:
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
ppt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样本时间索引序号。
进一步地,所述更新模块72用于:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,所述更新模块72用于:
根据公式:
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为X
state(t)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
具体地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,所述获取模块71包括:第三获取单元,用于分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;第四获取单元,用于根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
进一步地,所述第三获取单元中所述麦克风接收信号的子带功率谱的获取方式为:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述第三获取单元中所述子带域耦合因子的有效估计值的获取方式包括:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
进一步地,所述根据互相关方法获取子带域耦合因子的有偏估计值的方式为:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的方式为:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
进一步地,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的方式为:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方的方式为:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
进一步地,所述根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值的方式为:
进一步地,所述第四获取单元用于:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
进一步地,所述更新模块72用于:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变 量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,所述更新模块72用于:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为X
state(k,n)的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
需要说明的是,该装置的实施例是与上述方法实施例一一对应的装置,上述方法实施例中所有实现方式均适用于该装置的实施例中,也能达到相同的技术效果。
本公开的一些实施例还提供一种计算机可读存储介质,其上存储有计算机程序,所述计算机程序被处理器执行时实现以下步骤:获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的 时变正则化因子;根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
具体地,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
进一步地,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,计算机程序被处理器执行时所述处理器还可实现如下步骤:分别获取麦克风接收信号的功率和耦合因子的有效估计值;
根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:
得到麦克风接收信号的功率;
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据互相关方法获取耦合因子的有偏估计值;获取用于所述耦合因子的有偏估计值补偿的修正因子;根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
其中,
为基于互相关技术的耦合因子的有偏估计值;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:
获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;其中,
为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x
*(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T
s-1,T
s为用于进行
估计时所使用的样本数目,且T
s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号,
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为 t信号样本时刻麦克风接收信号;
为
的共轭转置矩阵;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为FIR滤波器系数矢量,
w
t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
获取AEC输出信号与远端参考信号之间相关系数的有效幅度平方;
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:
获取耦合因子的有效估计值;其中,
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;t为数字信号样本时间索引序号。
获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
其中,δ
opt(t)为时变正则化因子;L为FIR滤波器系数的个数;
为麦克风接收信号的功率;
为耦合因子的有效估计值;
为基于互相关技术的耦合因子的有偏估计值;
为AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方;δ
min为预设小实常量,且 δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;t为数字信号样本时间索引序号。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号矢量,且
x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符;
为
的共轭转置矩阵;δ
opt(t)为时变正则化因子;e
*(t)为e(t)的复共轭;
e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的FIR滤波器系数矢量;
为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(t)为时变正则化因子;X
state(t)为L×P维状态矩阵,且
为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数;
为
的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差矢量;
为P维麦克风接收信号矢量,且
y(t-t3)为在t-t3信号样本时刻的麦 克风接收信号;
w
t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
进一步地,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,计算机程序被处理器执行时所述处理器还可实现如下步骤:
分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;
根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
其中,
为麦克风接收信号的子带功率谱;Y(k,n)为麦克风接收信号子带谱;α
a和α
d为预设的递归常量,且0≤α
a<α
d<1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据互相关方法获取子带域耦合因子的有偏估计值;
获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;
根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
其中,
为子带域耦合因子的有偏估计值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的信号帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;
根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅 度平方的备选值;
其中,
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X
*(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N
s-1,N
s为用于进行
估计时所使用的帧数目,且N
s<<L
s,L
s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱;
E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱;
为
的共轭转置矩阵;
为远端参考信号子带谱矢量,且
X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符;
为子带k中FIR滤波器系数矢量,
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L
s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
其中,δ
opt(k,n)为子带域时变正则化因子;
为麦克风接收信号的子带功率谱;
为子带域耦合因子的有效估计值,
为子带域耦合因子的有偏估计值;
为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方;δ
min为预设小实常量,且δ
min>0;ρ
0和ρ分别为预设的小实常数,且ρ>0,ρ
0>0;n为信号帧时间索引变量。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;
为远端参考信号子带谱矢量,
X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数,T为转置运算符;
为
的共轭转置矩阵;E
*(k,n)为E(k,n)的复共轭; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且
Y(k,n)为在n信号帧时刻麦克风接收信号子带谱;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ
opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
进一步地,计算机程序被处理器执行时所述处理器还可实现如下步骤:
根据公式:
应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;
其中,
为更新后的子带k中FIR滤波器系数矢量;
为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ
opt(k,n)为子带域时变正则化因子;X
state(k,n)为子带k中L×P维状态矩阵,且
为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数;
为
的共轭转置矩阵;I
P×P为P×P维单位矩阵;
为
的复共轭,且
为P维误差信号子带谱矢量;
为P维麦克风接收信号子带谱矢量,且
Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号;
W
n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L
s-1,L
s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
本公开提到的计算机可读存储介质可以是易失性的或非易失性的、瞬态 的或非瞬态的。
以上所述的是本公开的可选实施方式,应当指出对于本技术领域的普通人员来说,在不脱离本公开所述的原理前提下还可以作出若干改进和润饰,这些改进和润饰也在本公开的保护范围内。
Claims (72)
- 一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新方法,包括:获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
- 根据权利要求1所述的可持续自适应更新方法,其中,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
- 根据权利要求1所述的可持续自适应更新方法,其中,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,所述获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:分别获取麦克风接收信号的功率和耦合因子的有效估计值;根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
- 根据权利要求3所述的可持续自适应更新方法,其中,所述耦合因子的有效估计值的获取方式为:根据互相关方法获取耦合因子的有偏估计值;获取用于所述耦合因子的有偏估计值补偿的修正因子;根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
- 根据权利要求5所述的可持续自适应更新方法,其中,所述根据互相关方法获取耦合因子的有偏估计值的步骤,包括:其中, 为基于互相关技术的耦合因子的有偏估计值;x *(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求5所述的可持续自适应更新方法,其中,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的步骤,包括:获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求7所述的可持续自适应更新方法,其中,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的步骤,包括:其中, 为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x *(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 为 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求3所述的可持续自适应更新方法,其中,所述根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:根据公式:获取用于FIR滤波器系数矢量迭代更新的时变正则化因子;
- 根据权利要求3所述的可持续自适应更新方法,其中,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
- 根据权利要求3所述的可持续自适应更新方法,其中,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:根据公式:应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的FIR滤波器系数矢量; 为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(t)为时变正则化因子;X state(t)为L×P维状态矩阵,且 为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数; 为X state(t)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差矢量; 为P维麦克风接收信号矢量,且 y(t-t3)为在t-t3信号样本时刻的麦克风接收信号; w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求1所述的可持续自适应更新方法,其中,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,所述获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子的步骤,包括:分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
- 根据权利要求14所述的可持续自适应更新方法,其中,所述子带域耦合因子的有效估计值的获取方式包括:根据互相关方法获取子带域耦合因子的有偏估计值;获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
- 根据权利要求16所述的可持续自适应更新方法,其中,所述根据互相关方法获取子带域耦合因子的有偏估计值的步骤,包括:其中, 为子带域耦合因子的有偏估计值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的信号帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求16所述的可持续自适应更新方法,其中,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的步骤,包括:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系 数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求18所述的可持续自适应更新方法,其中,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的步骤包括:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;其中, 为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求14所述的可持续自适应更新方法,其中,所述根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子的步骤,包括:根据公式:获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子;
- 根据权利要求14所述的可持续自适应更新方法,其中,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:根据公式:其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2; 为远端参考信号子带谱矢量, X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数,T为转置运算符; 为 的共轭转置矩阵;E *(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且 Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求14所述的可持续自适应更新方法,其中,所述根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新的步骤,包括:根据公式:应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(k,n)为子带域时变正则化因子;X state(k,n)为子带k中L×P维状态矩阵,且 为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数; 为X state(k,n)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差信号子带谱矢量; 为P维麦克风接收信号子带谱矢量,且 Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号; W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置,包括存储器、处理器及存储在所述存储器上并可在所述处理器上运行的计算机程序;其中,所述处理器执行所述计算机程序时实现以下步骤:获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
- 根据权利要求25所述的可持续自适应更新装置,其中,所述预设信号包括:声学回声抵消器AEC中输入的远端参考语音信号和近端麦克风接收的语音信号、自适应噪声抵消系统中噪声参考信号和系统输入信号、自适应 干扰抵消系统中干扰参考信号和系统输入信号、自适应系统辨识中的激励输入信号和待辨识的未知系统输出信号的组合对中的一项。
- 根据权利要求25所述的可持续自适应更新装置,其中,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,计算机程序被处理器执行时所述处理器还可实现如下步骤:分别获取麦克风接收信号的功率和耦合因子的有效估计值;根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
- 根据权利要求27所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据互相关方法获取耦合因子的有偏估计值;获取用于所述耦合因子的有偏估计值补偿的修正因子;根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
- 根据权利要求29所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:其中, 为基于互相关技术的耦合因子的有偏估计值;x *(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 为 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求29所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求31所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:其中, 为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x *(t-t1)为 x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 为 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求27所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
- 根据权利要求27所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的FIR滤波器系数矢量; 为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(t)为时变正则化因子;X state(t)为L×P维状态矩阵,且 为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数; 为X state(t)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差矢量; 为P维麦克风接收信号矢量,且 y(t-t3)为在t-t3信号样本时刻的麦克风接收信号; w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求25所述的可持续自适应更新装置,其中,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所述时变正则化因子为子带域时变正则化因子,计算机程序被处理器执行时所述处理器还可实现如下步骤:分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
- 根据权利要求38所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据互相关方法获取子带域耦合因子的有偏估计值;获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
- 根据权利要求40所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:其中, 为子带域耦合因子的有偏估计值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的信号帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求40所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求42所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;其中, 为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求38所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2; 为远端参考信号子带谱矢量, X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数,T为转置运算符; 为 的共轭转置矩阵;E *(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且 Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; W n2(k,n)为n信号帧时刻子带k中FIR 滤波器第n2+1个系数;δ opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求38所述的可持续自适应更新装置,其中,计算机程序被处理器执行时所述处理器还可实现如下步骤:根据公式:应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(k,n)为子带域时变正则化因子;X state(k,n)为子带k中L×P维状态矩阵,且 为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P为AP算法的阶数; 为X state(k,n)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差信号子带谱矢量; 为P维麦克风接收信号子带谱矢量,且 Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号; W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 一种计算机可读存储介质,其上存储有计算机程序,其中,所述计算机程序被处理器执行时所述处理器实现以下步骤:获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
- 一种有限冲激响应FIR滤波器系数矢量的可持续自适应更新装置,包括:获取模块,用于获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子;更新模块,用于根据所述时变正则化因子,对所述FIR滤波器系数矢量进行更新。
- 根据权利要求50所述的可持续自适应更新装置,其中,所述预设信号包括AEC中输入的远端参考语音信号和近端麦克风接收的语音信号,所述获取模块包括:第一获取单元,用于分别获取麦克风接收信号的功率和耦合因子的有效估计值;第二获取单元,用于根据所述麦克风接收信号的功率和耦合因子的有效估计值,获取采用FIR滤波器系数矢量对预设信号处理时用于进行所述FIR滤波器系数矢量迭代更新的时变正则化因子。
- 根据权利要求51所述的可持续自适应更新装置,其中,所述第一获取单元中所述耦合因子的有效估计值的获取方式为:根据互相关方法获取耦合因子的有偏估计值;获取用于所述耦合因子的有偏估计值补偿的修正因子;根据所述耦合因子的有偏估计值和所述修正因子,获取耦合因子的有效估计值。
- 根据权利要求53所述的可持续自适应更新装置,其中,所述根据互相关方法获取耦合因子的有偏估计值的方式为:其中, 为基于互相关技术的耦合因子的有偏估计值;x *(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 为 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求53所述的可持续自适应更新装置,其中,所述获取用于所述耦合因子的有偏估计值补偿的修正因子的方式为:获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;根据所述AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值,获取AEC输出的误差信号与远端参考信号之间相关系数的有效幅度平方,并将此作为耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求55所述的可持续自适应更新装置,其中,所述获取AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值的方式 为:其中, 为AEC输出的误差信号与远端参考信号之间相关系数幅度平方的备选值;x(t-t1)为在t-t1信号样本时刻的远端参考信号;x *(t-t1)为x(t-t1)的复共轭;t1=0,1,2,…,T s-1,T s为用于进行 估计时所使用的样本数目,且T s<<L,L为FIR滤波器系数的个数;e(t-t1)为在t-t1信号样本时刻的AEC输出的误差信号, e(t)为t信号样本时刻AEC输出的误差信号;y(t)为t信号样本时刻麦克风接收信号; 为 的共轭转置矩阵; 为远端参考信号矢量,且 x(t-t2)为在t-t2信号样本时刻的远端参考信号;T为转置运算符; 为FIR滤波器系数矢量, w t2(t)为在t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求51所述的可持续自适应更新装置,其中,所述更新模块用于:应用归一化最小均方NLMS算法对所述FIR滤波器系数矢量进行持续的自适应更新;
- 根据权利要求51所述的可持续自适应更新装置,其中,所述更新模块用于:根据公式:应用仿射投影AP算法对所述FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的FIR滤波器系数矢量; 为更新前的FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(t)为时变正则化因子;X state(t)为L×P维状态矩阵,且 为在t-t3信号样本时刻的远端参考信号矢量,且t3=0,1,…,P-1,P为AP算法的阶数; 为X state(t)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差矢量; 为P维麦克风接收信号矢量,且 y(t-t3)为在t-t3信号样本时刻的麦克风接收信号; w t2(t)为t信号样本时刻FIR滤波器第t2+1个系数,t2=0,1,2,…,L-1;t为数字信号样本时间索引序号。
- 根据权利要求50所述的可持续自适应更新装置,其中,所述预设信号包括AEC中输入的近端麦克风接收的语音信号的子带谱和远端参考语音信号的子带谱,所述FIR滤波器系数矢量为子带域FIR滤波器系数矢量,所 述时变正则化因子为子带域时变正则化因子,所述获取模块包括:第三获取单元,用于分别获取麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值;第四获取单元,用于根据所述麦克风接收信号的子带功率谱和子带域耦合因子的有效估计值,获取采用子带域FIR滤波器系数矢量对预设信号处理时用于进行所述子带域FIR滤波器系数矢量迭代更新的子带域时变正则化因子。
- 根据权利要求62所述的可持续自适应更新装置,其中,所述第三获取单元中所述子带域耦合因子的有效估计值的获取方式包括:根据互相关方法获取子带域耦合因子的有偏估计值;获取用于所述子带域耦合因子的有偏估计值补偿的修正因子;根据所述子带域耦合因子的有偏估计值和所述修正因子,获取子带域耦合因子的有效估计值。
- 根据权利要求64所述的可持续自适应更新装置,其中,所述根据互相关方法获取子带域耦合因子的有偏估计值的方式为:其中, 为子带域耦合因子的有偏估计值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的信号帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求64所述的可持续自适应更新装置,其中,所述获取用于所述子带域耦合因子的有偏估计值补偿的修正因子的方式为:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;根据所述AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值,获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数的有效幅度平方,并将此作为子带域耦合因子的有偏估计值补偿的修正因子。
- 根据权利要求66所述的可持续自适应更新装置,其中,所述获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值的方式为:获取AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;其中, 为AEC输出误差信号子带谱与远端参考信号子带谱之间相关系数幅度平方的备选值;X *(k,n-n1)为X(k,n-n1)的复共轭;n1=0,1,2,…,N s-1,N s为用于进行 估计时所使用的帧数目,且N s<<L s,L s为每个子带中FIR滤波器系数的个数;E(k,n-n1)为在n-n1信号帧时刻的AEC输出误差信号子带谱; E(k,n)为在n信号帧时刻AEC输出误差信号子带谱;Y(k,n)为麦克风接收信号子带谱; 为 的共轭转置矩阵; 为远端参考信号子带谱矢量,且 X(k,n-n2)为在n-n2信号帧时刻的远端参考信号子带谱;T为转置运算符; 为子带k中FIR滤波器系数矢量, W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,2,…,L s-1;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求62所述的可持续自适应更新装置,其中,所述更新模块用于:根据公式:其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2; 为远端参考信号子带谱矢量, X(k,n-n2)为第n-n2信号帧时刻远端参考信号子带谱;n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数,T为转置运算符; 为 的共轭转置矩阵;E *(k,n)为E(k,n)的复共轭;E(k,n)为在n信号帧时刻AEC输出误差信号子带谱,且 Y(k,n)为在n信号帧时刻麦克风接收信号子带谱; W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数;δ opt(k,n)为子带域时变正则化因子;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
- 根据权利要求62所述的可持续自适应更新装置,其中,所述更新模块用于:根据公式:应用仿射投影AP算法对子带域FIR滤波器系数矢量进行持续的自适应更新;其中, 为更新后的子带k中FIR滤波器系数矢量; 为更新前的子带k中FIR滤波器系数矢量;μ为预定的系数更新步长参数,且0<μ<2;δ opt(k,n)为子带域时变正则化因子;X state(k,n)为子带k中L×P维状态矩阵,且 为在n-n3信号帧时刻的远端参考信号子带谱矢量,且n3=0,1,…,P-1,P 为AP算法的阶数; 为X state(k,n)的共轭转置矩阵;I P×P为P×P维单位矩阵; 为 的复共轭,且 为P维误差信号子带谱矢量; 为P维麦克风接收信号子带谱矢量,且 Y(k,n-n3)为在n-n3信号帧时刻的麦克风接收信号; W n2(k,n)为n信号帧时刻子带k中FIR滤波器第n2+1个系数,n2=0,1,…,L s-1,L s为每个子带中FIR滤波器系数的个数;k为子带索引变量,k=0,1,2,…,K-1,且K为总子带数;n为信号帧时间索引变量。
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| US16/645,356 US11295752B2 (en) | 2017-09-07 | 2018-08-21 | Method and device of sustainably updating coefficient vector of finite impulse response filter |
Applications Claiming Priority (2)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201710800778.4 | 2017-09-07 | ||
| CN201710800778.4A CN107749304B (zh) | 2017-09-07 | 2017-09-07 | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| WO2019047710A1 true WO2019047710A1 (zh) | 2019-03-14 |
Family
ID=61255059
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| PCT/CN2018/101491 Ceased WO2019047710A1 (zh) | 2017-09-07 | 2018-08-21 | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 |
Country Status (4)
| Country | Link |
|---|---|
| US (1) | US11295752B2 (zh) |
| CN (1) | CN107749304B (zh) |
| TW (1) | TWI692754B (zh) |
| WO (1) | WO2019047710A1 (zh) |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US12401945B2 (en) | 2020-12-03 | 2025-08-26 | Dolby Laboratories Licensing Corporation | Subband domain acoustic echo canceller based acoustic state estimator |
Families Citing this family (15)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN107749304B (zh) * | 2017-09-07 | 2021-04-06 | 电信科学技术研究院 | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 |
| CN108390663B (zh) | 2018-03-09 | 2021-07-02 | 电信科学技术研究院有限公司 | 一种有限冲激响应滤波器系数矢量的更新方法及装置 |
| CN110136739B (zh) * | 2019-07-03 | 2021-04-09 | 四川大学 | 语音信号中冲激性噪声的去除方法 |
| CN110531149B (zh) * | 2019-08-31 | 2021-06-18 | 广东石油化工学院 | 一种基于波形正则化的功率信号滤波方法和系统 |
| US11804233B2 (en) * | 2019-11-15 | 2023-10-31 | Qualcomm Incorporated | Linearization of non-linearly transformed signals |
| CN110996210B (zh) * | 2019-12-13 | 2021-11-23 | 恒玄科技(上海)股份有限公司 | 一种用于声场均衡的方法和一种耳机 |
| TWI772929B (zh) * | 2020-10-21 | 2022-08-01 | 美商音美得股份有限公司 | 分析濾波器組 及其運算程序、音訊移頻系統 及音訊移頻程序 |
| US11568884B2 (en) | 2021-05-24 | 2023-01-31 | Invictumtech, Inc. | Analysis filter bank and computing procedure thereof, audio frequency shifting system, and audio frequency shifting procedure |
| CN113938113B (zh) * | 2021-07-14 | 2025-07-25 | 沈阳工业大学 | 一种变参数零吸引最小均方稀疏系统辨识方法 |
| US11863702B2 (en) * | 2021-08-04 | 2024-01-02 | Nokia Technologies Oy | Acoustic echo cancellation using a control parameter |
| CN113838474B (zh) * | 2021-11-25 | 2022-02-18 | 全时云商务服务股份有限公司 | 通信系统啸叫抑制方法和装置 |
| CN115209312B (zh) * | 2022-06-21 | 2024-10-22 | 欧仕达听力科技(厦门)有限公司 | 声学设备及其主动反馈抑制方法 |
| CN115527548B (zh) * | 2022-08-31 | 2025-06-27 | 博泰车联网(南京)有限公司 | 车机端功能无关的语音过滤方法、装置、终端和介质 |
| CN119484706B (zh) * | 2025-01-08 | 2025-11-18 | 青岛科技大学 | 一种基于非规范fir结构的回声消除方法及系统 |
| CN120528447B (zh) * | 2025-07-24 | 2025-10-31 | 西安乾景防务技术有限公司 | 一种协议信号处理模块的高速信号接收方法和系统 |
Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009284465A (ja) * | 2008-04-24 | 2009-12-03 | Nec Engineering Ltd | エコーキャンセラ装置 |
| US20110044448A1 (en) * | 2009-08-24 | 2011-02-24 | Leanics Corporation | System for low complexity adaptive ECHO and NEXT cancellers |
| CN104616658A (zh) * | 2015-01-14 | 2015-05-13 | 重庆金美通信有限责任公司 | 一种支持多种语音编码制式的回声抵消实现方法 |
| CN104883462A (zh) * | 2014-02-28 | 2015-09-02 | 中国科学院声学研究所 | 一种用于消除声学回声的自适应滤波器及滤波方法 |
| CN106161823A (zh) * | 2016-05-12 | 2016-11-23 | 西南交通大学 | 基于连续混合p范数的子带结构回声消除方法 |
| CN107483761A (zh) * | 2016-06-07 | 2017-12-15 | 电信科学技术研究院 | 一种回波抑制方法及装置 |
| CN107749304A (zh) * | 2017-09-07 | 2018-03-02 | 电信科学技术研究院 | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 |
Family Cites Families (11)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP3397269B2 (ja) * | 1994-10-26 | 2003-04-14 | 日本電信電話株式会社 | 多チャネル反響消去方法 |
| MY154452A (en) | 2008-07-11 | 2015-06-15 | Fraunhofer Ges Forschung | An apparatus and a method for decoding an encoded audio signal |
| US8285033B2 (en) * | 2010-04-01 | 2012-10-09 | Seiko Epson Corporation | Bi-affinity filter: a bilateral type filter for color images |
| US8675137B2 (en) | 2010-04-13 | 2014-03-18 | Newport Media, Inc. | Apparatus and method for adaptive filtering |
| US9076426B2 (en) * | 2011-12-20 | 2015-07-07 | Avago Technologies General Ip (Singapore) Pte. Ltd. | Training an echo canceller in severe noise |
| US9088336B2 (en) * | 2012-09-06 | 2015-07-21 | Imagination Technologies Limited | Systems and methods of echo and noise cancellation in voice communication |
| EP2738762A1 (en) * | 2012-11-30 | 2014-06-04 | Aalto-Korkeakoulusäätiö | Method for spatial filtering of at least one first sound signal, computer readable storage medium and spatial filtering system based on cross-pattern coherence |
| GB201414352D0 (en) | 2014-08-13 | 2014-09-24 | Microsoft Corp | Reversed echo canceller |
| US9854201B2 (en) | 2015-01-16 | 2017-12-26 | Microsoft Technology Licensing, Llc | Dynamically updating quality to higher chroma sampling rate |
| CN105070295A (zh) * | 2015-07-10 | 2015-11-18 | 西南交通大学 | 一种活性因子成比例子带的自适应方法应用于回声消除 |
| CN106448695B (zh) * | 2016-09-28 | 2019-09-03 | 天津大学 | 一种双端通话鲁棒的变阶变步长仿射投影方法 |
-
2017
- 2017-09-07 CN CN201710800778.4A patent/CN107749304B/zh active Active
-
2018
- 2018-08-21 WO PCT/CN2018/101491 patent/WO2019047710A1/zh not_active Ceased
- 2018-08-21 US US16/645,356 patent/US11295752B2/en active Active
- 2018-08-28 TW TW107129916A patent/TWI692754B/zh active
Patent Citations (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| JP2009284465A (ja) * | 2008-04-24 | 2009-12-03 | Nec Engineering Ltd | エコーキャンセラ装置 |
| US20110044448A1 (en) * | 2009-08-24 | 2011-02-24 | Leanics Corporation | System for low complexity adaptive ECHO and NEXT cancellers |
| CN104883462A (zh) * | 2014-02-28 | 2015-09-02 | 中国科学院声学研究所 | 一种用于消除声学回声的自适应滤波器及滤波方法 |
| CN104616658A (zh) * | 2015-01-14 | 2015-05-13 | 重庆金美通信有限责任公司 | 一种支持多种语音编码制式的回声抵消实现方法 |
| CN106161823A (zh) * | 2016-05-12 | 2016-11-23 | 西南交通大学 | 基于连续混合p范数的子带结构回声消除方法 |
| CN107483761A (zh) * | 2016-06-07 | 2017-12-15 | 电信科学技术研究院 | 一种回波抑制方法及装置 |
| CN107749304A (zh) * | 2017-09-07 | 2018-03-02 | 电信科学技术研究院 | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 |
Cited By (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US12401945B2 (en) | 2020-12-03 | 2025-08-26 | Dolby Laboratories Licensing Corporation | Subband domain acoustic echo canceller based acoustic state estimator |
Also Published As
| Publication number | Publication date |
|---|---|
| US20210035593A1 (en) | 2021-02-04 |
| CN107749304B (zh) | 2021-04-06 |
| CN107749304A (zh) | 2018-03-02 |
| TW201913646A (zh) | 2019-04-01 |
| US11295752B2 (en) | 2022-04-05 |
| TWI692754B (zh) | 2020-05-01 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| WO2019047710A1 (zh) | 有限冲激响应滤波器系数矢量的可持续更新方法及装置 | |
| CN109727604B (zh) | 用于语音识别前端的频域回声消除方法及计算机储存介质 | |
| JP5450567B2 (ja) | クリアな信号の取得のための方法及びシステム | |
| TWI388190B (zh) | 用於回聲抑制的濾波器係數的計算設備和方法 | |
| TWI458331B (zh) | 用於計算回聲抑制濾波器的控制資訊的裝置和方法,以及用於計算延遲值的裝置和方法 | |
| CN109040499B (zh) | 一种抗冲击干扰的自适应回声消除方法 | |
| CN106486135B (zh) | 近端语音检测器、语音系统、对语音进行分类的方法 | |
| CN102739886A (zh) | 基于回声频谱估计和语音存在概率的立体声回声抵消方法 | |
| Albu et al. | New variable step size affine projection algorithms | |
| CN110191245B (zh) | 一种基于时变参数的自适应回声消除方法 | |
| CN104144269A (zh) | 一种基于解相关的成比例自适应电话回声消除方法 | |
| WO2019024621A1 (zh) | 一种声学回声抵消器输出语音信号的后处理方法及装置 | |
| KR102157276B1 (ko) | 비선형 음향 에코 소거 장치 및 그 방법 | |
| CN111868826A (zh) | 一种回声消除中的自适应滤波方法、装置、设备及存储介质 | |
| CN113347536B (zh) | 基于线性预测与子带自适应滤波的声反馈抑制算法 | |
| US9729968B2 (en) | Method and system for acoustic echo cancellation using cascaded kalman filtering | |
| JP2004537941A (ja) | 最小二乗法を用いて導出されるインパルス応答のエラーキャラクタリゼーション推定値を提供するための方法および装置 | |
| EP1055292B1 (en) | Echo canceller | |
| JPH036918A (ja) | 適合非連続フィルタ | |
| Bensouda et al. | A data reuse fast NLMS algorithm for acoustic echo cancellation | |
| Wung et al. | A system approach to residual echo suppression in robust hands-free teleconferencing | |
| Faza et al. | Adaptive regularization in frequency-domain NLMS filters | |
| CN115643342B (zh) | 一种回声消除方法 | |
| Kallinger et al. | Enhanced doubletalk detection based on pseudo-coherence in stereo | |
| Valero et al. | Coherence-aware stereophonic residual echo estimation |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| 121 | Ep: the epo has been informed by wipo that ep was designated in this application |
Ref document number: 18853446 Country of ref document: EP Kind code of ref document: A1 |
|
| NENP | Non-entry into the national phase |
Ref country code: DE |
|
| 122 | Ep: pct application non-entry in european phase |
Ref document number: 18853446 Country of ref document: EP Kind code of ref document: A1 |











































































































































