JP6596833B2 - Noise suppression device and program, noise estimation device and program, and SNR estimation device and program - Google Patents

Description

  The present invention relates to a noise suppression device and program, a noise estimation device and program, and an SNR estimation device and program, for example, a communication terminal and an audio device that desire to suppress a noise component included in an input signal and emphasize a speech component. It can be applied to a voice recognition device or the like.

  Since noise exists everywhere in the natural environment, generally, when recording voice in the real world, noise from various sources is mixed in the observed signal. Those noises reduce the intelligibility of speech even when a person listens, and reduce the accuracy of speech processing (for example, speech recognition rate) even when input to a speech processing device such as a speech recognition device. For this reason, there is a great demand for a technique for enhancing a speech component by suppressing a noise component mixed in an input signal, and various noise suppression methods (sometimes called a speech enhancement method) have been developed so far.

  In a general noise suppression method, a noise power spectrum (noise power spectrum) or an SNR (Signal-to-Noise Ratio) for each frequency band is required. In addition, since the power spectrum (input power spectrum) and the noise power spectrum of the input signal are used for the calculation of the SNR, it is only necessary to know the noise power spectrum after all. In general noise power spectrum estimation, an approach is adopted in which a noise interval is first specified by voice activity detection (VAD), and then an input power spectrum in the noise interval is averaged.

  In the above method, the estimation accuracy of the noise power spectrum depends on the performance of VAD. However, it is difficult to realize highly accurate VAD without information about noise components, and even if it can be realized, a large amount of calculation is required.

  On the other hand, the noise suppression method described in Patent Document 1 does not require VAD to estimate the noise power spectrum. Hereinafter, the algorithm of the noise suppression method described in Patent Document 1 will be briefly described. In the following, elements of a frequency band having an input power spectrum and a noise power spectrum are referred to as input power and noise power, respectively.

  The input power is multiplied by an appropriate weighting factor, the obtained weighted input power is stored for a predetermined time (T seconds), and the average value of the stored weighted input power is used as the estimated noise power. The appropriate weighting factor is calculated by the predicted post-SNR obtained by dividing the current input power by the previous estimated noise power. Specifically, the weighting factor is set to 1 when the predicted posterior SNR is equal to or smaller than a predetermined value G1, and the weighting factor is inversely proportional to the predicted posterior SNR when the predicted posterior SNR is larger than the value G1 and equal to or smaller than the predetermined value G2 (G2> G1). Is set, and the weighting factor is set to 0 when the predicted a posteriori SNR is larger than the value G2. When the weight coefficient is 0, the weighted input power is not stored. The posterior SNR is calculated by dividing the input power by the estimated noise power obtained in this way, and noise suppression (also referred to as speech enhancement) is performed based on the obtained posterior SNR. For noise suppression, a method described in Non-Patent Document 1 called MMSE-STSA is used.

JP 2002-204175 A

Y. Ephraim and D.C. Malah, “Speech enhancement using a minimum mean-square error short-time spectral Amplitude Estimator”, IEEE ASSP, vol. ASSP-32, no. 6, p. 1109-1121, Dec. 1984

  The technique described in Patent Document 1 uses fixed threshold values G1 and G2, and therefore, particularly when the noise level is high, the sound component is averaged as a noise component, and the estimation of the noise power spectrum becomes inaccurate. There's a problem. As a result, the accuracy of the posterior SNR is lowered and the accuracy of noise suppression is also lowered.

  Therefore, a noise suppression device and program, a noise estimation device and program, and an SNR estimation device and program with high accuracy in suppression processing and estimation processing are desired.

According to a first aspect of the present invention, in a noise suppression apparatus that suppresses a noise component included in an input signal and emphasizes a target sound component, (1) a frequency analysis unit that calculates an input spectrum by performing frequency analysis of the input signal; 2) a plurality of band-specific noise suppression units that correspond to the frequency band of any input spectrum calculated by the frequency analysis unit and suppress noise components in the signal of the frequency band; The noise suppression means (2-1) compares the first feature amount based on the first input power calculated for the input frequency band signal with the first threshold value generated internally, A first parameter calculation unit for obtaining a detection result of the first target sound section in the frequency band signal, and (2-2) a second input power based on the second input power calculated for the input frequency band signal. 2 A second parameter calculation unit that compares the collected amount with a second threshold value generated internally to obtain a detection result of the second target sound section in the input frequency band signal; (2-3) Noise that suppresses a noise component included in the input frequency band signal based on the first parameter obtained by the first parameter calculation unit and the second parameter obtained by the second parameter calculation unit (2-1a) the first parameter calculation unit outputs at least a detection result of the second target sound section output by the second parameter calculation unit before a predetermined unit time. The first threshold value is generated using a second parameter that includes (2-2a) the second parameter calculation unit, and the first parameter calculation unit outputs the first threshold value in the same unit time. The detection result of the target sound section of And generates the second threshold value using a second parameter including.

According to a second aspect of the present invention, in the noise estimation apparatus for estimating the noise power in the input signal, (1) a frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum, and (2) the frequency analysis unit calculates A plurality of noise estimation means for each band corresponding to a frequency band of any one of the input spectra and estimating the noise power in the signal in that frequency band; and (3) the frequency bands obtained by the noise estimation means for each band. And a noise power integration unit for each band that obtains a final noise power estimation value by integrating a plurality of noise power estimation values, and (2) each of the noise estimation means for each band is (2-1) a first feature amount based on the first input power calculated for the input the frequency band signal, by comparing the first threshold value generated internally in the input the frequency band signals A first parameter calculating section that detects a target sound section, and (2-2) second feature amount based on the second input power calculated for the input the frequency band signal, generated internally A second parameter calculation unit that compares the second threshold value and detects a second target sound section in the input frequency band signal, and (2-1) the first parameter calculation unit. (2-1-1) The first input power is smoothed by controlling the execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. A first smoothing unit that calculates one smoothing power; and (2-1-2) a first threshold value calculation unit that calculates the first threshold value by applying at least the first smoothing power. And (2-1-3) the first input power as the first feature amount A first target sound section determination unit that determines whether or not the target sound section is compared with the first threshold value, and obtains a detection result of the first target sound section, (2-2) The second parameter calculation unit (2-2-1) controls the execution and stop of the smoothing based on the detection result of the first target sound section of the same unit time, while the second input power And (2-2-2) calculating the second threshold value by applying at least the second smoothing power to the second smoothing unit that calculates the second smoothing power by smoothing And (2-2-3) comparing the second input power with the second threshold value as the second feature amount to determine whether the target sound section is the second sound amount. A second target sound section determination unit that obtains the detection result of the target sound section of (4), (4) the first smoothing unit or the second smoothing unit, Smoothing is performed when the detection result of the second target sound section before a predetermined unit time or the detection result of the first target sound section of the same unit time is not the target sound section, and smoothed when the detection result is the target sound section. And the first smoothing power or the second smoothing power is obtained as an estimated value of noise power for each band.

According to a third aspect of the present invention, in the SNR estimation apparatus for estimating the SNR in the input signal, (1) a frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum, and (2) any of the frequency analysis units calculated by the frequency analysis unit A plurality of SNR estimation means for each band corresponding to the frequency band of the input spectrum and estimating the SNR in the signal of the frequency band, and (3) a plurality of frequency band-specific SNR estimation means obtained by the respective SNR estimation means for each band. Band-specific SNR integrating means for integrating final SNR estimated values to obtain a final SNR estimated value, and (2) each of the band-specific SNR estimating means is (2-1) the input frequency band a first feature amount based on the first input power calculated for the signal, by comparing the first threshold value generated inside, the first target sound Ward in the input the frequency band signals A first parameter calculation unit that detects (2), a second feature amount based on the second input power calculated for the input frequency band signal, and a second threshold value generated internally And a second parameter calculation unit for detecting a second target sound section in the input frequency band signal,
(2-1) The first parameter calculation unit (2-1-1) controls the execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. A first smoothing unit that smoothes the first input power to calculate a first smoothing power; and (2-1-2) applies at least the first smoothing power, and A first threshold value calculation unit that calculates a threshold value of 1, and (2-1-3) whether the first input power is the first feature amount and is compared with the first threshold value to determine whether or not the target sound section And (2-2) the second parameter calculation unit is (2-2-1) having a first target sound section determination unit that obtains a detection result of the first target sound section. An SNR calculation unit for calculating an estimated value of SNR based on the second input power and the first smoothing power of the same unit time; (2-2-2) Based on the detection result of the first target sound section in the same unit time, the smoothed SNR value is obtained by smoothing the SNR estimated value while controlling the execution and stop of smoothing. A second smoothing unit to calculate, (2-2-3) a second threshold value calculating unit to calculate the second threshold value by applying at least the SNR smooth value, and (2-2-4) The SNR estimation value is compared with the second threshold value as the second feature amount to determine whether or not the target sound section is present, and second target sound section determination for obtaining a detection result of the second target sound section (4) The SNR estimation value from the SNR calculation unit is obtained as an SNR estimation value of the frequency band in the input frequency band signal.

A fourth aspect of the present invention is a noise suppression program that suppresses a noise component included in an input signal and emphasizes a target sound component, and (1) a frequency at which an input spectrum is calculated by frequency analysis of the input signal. An analysis unit, and (2) function as a plurality of band noise suppression units that correspond to the frequency band of any input spectrum calculated by the frequency analysis unit and suppress noise components in the signal of the frequency band. Yes, (2) the band-specific noise suppression means includes (2-1) a first feature amount based on the first input power calculated for the input frequency band signal, and a first generated internally. A first parameter calculation unit that compares the threshold value and obtains a detection result of the first target sound section in the input frequency band signal; and (2-2) calculates the input frequency band signal. The second feature value based on the input second input power is compared with the internally generated second threshold value to obtain a detection result of the second target sound section in the input frequency band signal. Based on the second parameter calculation unit, (2-3) the first parameter obtained by the first parameter calculation unit, and the second parameter obtained by the second parameter calculation unit A noise suppression unit that suppresses a noise component included in the frequency band signal. (2-1a) The first parameter calculation unit is output by the second parameter calculation unit a predetermined unit time before Generating the first threshold value using a second parameter including at least the detection result of the second target sound section, and (2-2a) the second parameter calculation unit calculates the first parameter calculation. part is output in the same unit time And generates the second threshold value using the second parameters including at least a detection result of the first target sound period.

The fifth aspect of the present invention is a noise estimation program for estimating noise power in an input signal, the computer comprising: (1) a frequency analysis unit for frequency analysis of the input signal to calculate an input spectrum; and (2) the frequency described above. A plurality of band-specific noise estimation means for estimating a noise power in a signal of the frequency band corresponding to a frequency band of any one of the input spectra calculated by the analysis unit; In addition, a plurality of noise power estimation values for each frequency band are integrated to function as band-specific noise power integration means for obtaining a final noise power estimation value. (2) Each band-specific noise estimation means (2-1) Compare the first feature quantity based on the first input power calculated for the input frequency band signal and the first threshold value generated internally. A first parameter calculating section for detecting a first target sound section in the input the frequency band signals, based on the second input power calculated for the frequency band signal inputted (2-2) A second parameter calculation unit that compares the second feature amount with a second threshold value generated internally and detects a second target sound section in the input frequency band signal; and (2 -1) The first parameter calculation unit (2-1-1) controls the execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. A first smoothing unit that smoothes the first input power and calculates a first smoothing power; and (2-1-2) applies at least the first smoothing power, A first threshold value calculation unit for calculating a threshold value, and (2-1-3) above The first target sound section determination is performed by comparing the first input power with the first threshold value as the first feature value to determine whether or not the target sound section is present, and obtaining a detection result of the first target sound section. (2-2) The second parameter calculation unit (2-2-1) executes smoothing based on the detection result of the first target sound section of the same unit time, A second smoothing unit that calculates the second smoothing power by smoothing the second input power while controlling the stop; and (2-2-2) at least the second smoothing power is applied. A second threshold value calculation unit for calculating the second threshold value, and (2-2-3) comparing the second input power with the second threshold value as the second feature value. A second target sound segment determination unit that determines whether or not the current segment is a sound segment and obtains a detection result of the second target sound segment; (4) The first smoothing unit or the second smoothing unit uses the detection result of the second target sound section before a predetermined unit time or the detection result of the first target sound section of the same unit time as the purpose. Smoothing is performed when it is not a sound section, smoothing is stopped when it is a target sound section, and the first smoothing power or the second smoothing power is obtained as an estimated value of noise power for each band. And

A sixth aspect of the present invention is an SNR estimation program for estimating an SNR in an input signal, comprising: (1) a frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum; and (2) the frequency analysis described above. A plurality of band-specific SNR estimation means for estimating the SNR of a signal in the frequency band corresponding to the frequency band of any input spectrum calculated by the section; and (3) the band-specific SNR estimation means obtained above. A plurality of SNR estimation values for each frequency band are integrated to function as band-specific SNR integration means for obtaining a final SNR estimation value. (2) Each of the band-specific SNR estimation means is (2- 1) The first feature value based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, A first parameter calculating section for detecting a first target sound section in the serial frequency band signals, (2-2) second feature based on a second input power calculated for the input the frequency band signals A second parameter calculation unit that compares the amount with a second threshold value generated internally and detects a second target sound section in the input frequency band signal, (2-1) The first parameter calculation unit (2-1-1) controls the execution and stop of the smoothing based on the detection result of the second target sound section of a predetermined unit time before the first input. A first smoothing unit that calculates the first smoothed power by smoothing the power; and (2-1-2) calculates the first threshold value by applying at least the first smoothed power. A first threshold value calculation unit; and (2-1-3) the first input power Is compared with the first threshold value as the first feature value to determine whether or not the target sound section is present, and a first target sound section determining unit for obtaining a detection result of the first target sound section is provided. (2-2) The second parameter calculation unit (2-2-1) calculates an SNR estimate based on the second input power and the first smoothing power in the same unit time. Based on the SNR calculation unit to calculate and (2-2-2) the detection result of the first target sound section of the same unit time, smoothing the SNR estimated value while controlling smoothing execution and stop A second smoothing unit that calculates the smoothed value of the SNR, and (2-2-3) a second threshold value calculating unit that calculates the second threshold value by applying at least the SNR smoothed value; (2-2-4) The SNR estimated value is compared with the second threshold value as the second feature value, and the target sound section And a second target sound section determination unit for determining whether or not the second target sound section is detected, and (4) the SNR estimated value from the SNR calculation section is input It is obtained as an estimated value of SNR of the frequency band signal in the frequency band signal.

  According to the present invention, it is possible to provide a noise suppression device and program, a noise estimation device and program, and an SNR estimation device and program with high accuracy in suppression processing and estimation processing.

It is a block diagram which shows the structure of the noise suppression apparatus of 1st Embodiment. It is a block diagram which shows the detailed structure of the 1st parameter calculation part in the noise suppression apparatus of 1st Embodiment. It is a block diagram which shows the detailed structure of the 2nd parameter calculation part in the noise suppression apparatus of 1st Embodiment. It is a block diagram which shows the detailed structure of the 2nd parameter calculation part in the noise suppression apparatus which deform | transformed 1st Embodiment by the hangover surface. It is a block diagram which shows the detailed structure of the 1st parameter calculation part in the noise suppression apparatus of 2nd Embodiment. It is a block diagram which shows the detailed structure of the 2nd parameter calculation part in the noise suppression apparatus of 2nd Embodiment. It is a block diagram which shows the detailed structure of the 2nd parameter calculation part in the noise suppression apparatus of 3rd Embodiment. It is a block diagram which shows the structure of the noise estimation apparatus of embodiment. It is a block diagram which shows the structure of the SNR estimation apparatus of embodiment.

(A) First Embodiment A noise suppression device and a first embodiment of the present invention will be described below with reference to the drawings.

(A-1) Configuration of the First Embodiment FIG. 1 is a block diagram showing the configuration of the noise suppression device of the first embodiment.

  The noise suppression apparatus according to the first embodiment can be configured by hardware as the components shown in FIG. 1, and can also be realized by software (noise suppression program) executed by the CPU and the CPU. However, even if any realization method is adopted, it can be functionally represented in FIG.

  In FIG. 1, the noise suppression device 100 according to the first embodiment includes a frequency analysis unit 101, band-specific noise suppression units 102-1 to 102 -M, and a waveform restoration unit 103.

  The frequency analysis unit 101 performs frequency analysis on the input signal to calculate a frequency spectrum, and gives the obtained input spectrum to the band-specific noise suppression units 102-1 to 102-M. For example, fast Fourier transform (FFT), wavelet transform, filter bank, or the like can be applied to the frequency analysis, but FFT is preferable. In the following, it is assumed that the input spectrum is given as a complex number.

  The number of noise suppression units 102-1 to 102-M for each band is provided by the number (M) of spectra (frequency bands) obtained by the frequency analysis unit 101. Hereinafter, a signal in the input spectrum to each band noise suppression means 102-1 to 102-M will be referred to as a frequency band signal.

  Each of the noise suppression units 102-1 to 102-M for each band (hereinafter, described with the branch numbers “1” to “M” omitted as appropriate) has a similar configuration, although the input frequency band signals are different. The noise suppression unit 102 for each band performs noise suppression as will be described later on the frequency band signal input to itself, and gives the frequency band signal after noise suppression to the waveform restoration unit 103.

  The waveform restoration unit 103 converts the output spectrum composed of the spectrum (frequency band signal after noise suppression) given from all the noise suppression means 102 for each band into a time domain signal, and the obtained time domain signal is The output of the noise suppression apparatus 100 is output to the next stage apparatus. The conversion to the time domain signal uses a method that is paired with the frequency analysis technique used in the frequency analysis unit 101. For example, if the frequency analysis technique is FFT, the inverse fast Fourier transform is used for the conversion to the time domain signal. (Inverse FFT; IFFT) is used.

  Each band-specific noise suppression unit 102 includes a power calculation unit 110, a first parameter calculation unit 111, a second parameter calculation unit 112, a unit time delay unit 113, and a noise suppression unit 114.

  In the noise suppression unit 102 for each band, the input frequency band signal is supplied to the power calculation unit 110 and the noise suppression unit 114.

  The power calculation unit 110 calculates the power for TP seconds of the input frequency band signal, and provides the obtained input power Pin to the first parameter calculation unit 111, the second parameter calculation unit 112, and the noise suppression unit 114. Is. A known calculation method can be applied as the power calculation method. For example, the sum of squares of absolute values or the sum of absolute values may be calculated as input power, or the maximum amplitude for TP seconds may be calculated as input power.

  The first parameter calculation unit 111 uses the second parameter F2 including the second speech section truth value V2 that is the detection result of the second parameter calculation unit 112 one unit time ago, and the input power Pin. The first parameter F1 including the obtained voice section truth value (first voice section truth value) V1 is given to the second parameter calculation unit 112 and the noise suppression unit 114. is there. The unit time described above is, for example, a frame of 10 milliseconds applied in audio processing or the like.

  The second parameter calculation unit 112 uses the first parameter F1 including at least the first voice segment true / false value V1, which is the detection result of the first parameter calculation unit 111, and the voice segment detection using the input power Pin. And the second parameter F2 including at least the obtained speech section truth value (second speech section truth value) V2 is provided to the first parameter calculation unit 111 via the unit time delay unit 113. The second parameter F2 described above is given to the noise suppression unit 114.

  The unit time delay unit 113 delays the second parameter F2 output from the second parameter calculation unit 112 by one unit time and gives it to the first parameter calculation unit 111.

  The noise suppression unit 114 suppresses the noise component of the input spectrum (input frequency band signal) based on the first parameter F1, the second parameter F2, and the input power Pin, and the obtained output spectrum (after suppression) Frequency band signal) to the waveform restoration unit 103. As a noise suppression method, a spectral subtraction method with a small amount of calculation is suitable, but a Wiener filter method, MMSE-STSA, or the like may be used.

  When using spectral subtraction, the suppression gain is given by subtracting the noise amplitude (or power) from the input amplitude (or power) and then dividing by the input amplitude (or power). However, since the suppression gain becomes negative when the subtraction result becomes a negative value, measures are taken such that the suppression gain does not fall below a predetermined minimum gain value. On the other hand, when a Wiener filter or MMSE-STSA is used, the suppression gain is calculated based on the posterior SNR and the prior SNR. The a priori SNR can be estimated from the a posteriori SNR and the suppression gain of one unit time before using the Decision-Directed method. The posterior SNR is calculated by dividing the input power by the noise power.

  These suppression methods require a first smoothing power P1, which will be described later, included in the first parameter F1. Further, based on the second speech section truth value V2 included in the second parameter F2, if the second speech section truth value V2 is a true value (a value representing the speech section), 1 is set. Noise suppression is a method of suppressing noise by multiplying an input frequency band signal by 0 or a very small value (for example, 0.01) if the voice interval true / false value V2 is a false value (a value representing a noise interval). Part 114 can be used.

  Each of the first parameter calculation unit 111 and the second parameter calculation unit 112 arranges and applies an existing speech segment detection technique. The existing speech segment detection technology applied by the first parameter calculation unit 111 and the existing speech segment detection technology applied by the second parameter calculation unit 112 may be different or the same. There may be.

  In FIG. 1, the first parameter calculation unit 111 and the second parameter calculation unit 112 each apply a voice section detection technique that detects a voice section using input power, and thus can be applied in common. The power calculation unit 110 is described outside the first parameter calculation unit 111 and the second parameter calculation unit 112. However, at least one of the first parameter calculation unit 111 and the second parameter calculation unit 112 may be used. In the case where only the speech section detection technology that detects the speech section using the input power is applied, the common power calculation unit 110 is not necessary. Further, for example, the first power is used so that the input power used by the first parameter calculation unit 111 is the sum of squares of TP seconds, and the input power used by the second parameter calculation unit 112 is the maximum amplitude of TP seconds. The input power used by the parameter calculation unit 111 and the second parameter calculation unit 112 may be different from each other. In such a case, the power calculation unit is used as the first parameter calculation unit 111 and the second parameter calculation unit. It is necessary to provide each unit 112 separately.

  FIG. 2 is a block diagram illustrating a detailed configuration example of the first parameter calculation unit 111. In FIG. 2, the first parameter calculation unit 111 includes a first smoothing unit 201, a first threshold calculation unit 202, and a first speech segment determination unit 203.

  The first smoothing unit 201 smoothes the input power Pin based on the first speech section reference truth value Vr1 (= second speech section truth value V2 before one unit time) and is obtained. 1 smoothing power P <b> 1 is given to the first threshold value calculation unit 202. The first smoothing unit 201 smoothes the input power Pin and updates the first smoothed power P1 when the first speech section reference true / false value Vr1 is a false value (that is, a value representing a noise section). When the first speech section reference truth value Vr1 is a true value (that is, a value representing the speech section), the first smoothing power P1 is not updated. Accordingly, the first smoothing power P1 means a smoothed value of noise power (an average value of noise power). Note that the smoothing method and the smoothing configuration are not limited at all. For example, smoothing is performed using a time constant filter having a time constant of 0.2 seconds.

  The first threshold value calculation unit 202 multiplies the first smoothing power P1 by a predetermined constant coefficient C1 (hereinafter referred to as the first coefficient) that takes a value of 1 or more, and compares it with the input power Pin. The threshold TH1 is formed and given to the first speech segment determination unit 203. The first smoothing power P1 means an average value of the noise power, and by multiplying this, the value of the first coefficient C1 that defines the first threshold TH1 for separating the voice power and the noise power. Although not limited, for example, 2 can be applied.

  The first speech segment determination unit 203 compares the first threshold value TH1 with the input power Pin to determine whether or not it is a speech segment, and outputs a first speech segment truth value V1. The first speech segment determination unit 203 outputs a true value as the first speech segment true / false value V1 if the input power Pin is greater than the first threshold TH1, and otherwise outputs a false value.

  The first parameter calculation unit 111 outputs the first smoothing power P1 and the first speech section truth value V1 obtained as described above as the first parameter F1.

  FIG. 3 is a block diagram illustrating a detailed configuration example of the second parameter calculation unit 112. In FIG. 3, the second parameter calculation unit 112 includes a second smoothing unit 301, a second threshold calculation unit 302, and a second speech segment determination unit 303.

  The second smoothing unit 301 smoothes the input power Pin based on the second speech section reference truth value Vr2 (first speech section truth value V1 in the same unit time), and obtains the second obtained The smoothing power P2 is given to the second threshold value calculation unit 302. The second smoothing unit 301 smoothes the input power Pin and updates the second smoothed power P2 when the second speech section reference truth value Vr2 is a true value (that is, a value representing the speech section). When the second speech section reference true / false value Vr2 is a false value (that is, a value representing a noise section), the second smoothing power P2 is not updated. Therefore, the second smoothing power P2 means the smoothed value of the audio power (average value of the audio power). Note that the smoothing method and the smoothing configuration are not limited at all. For example, smoothing is performed using a time constant filter having a time constant of 0.8 seconds.

  The time constant is determined by a trade-off between the followability of the target signal and the stability of the smoothed value, and the first smoothing unit 201 described above smoothes the input power Pin in the noise section. On the other hand, since the second smoothing unit 301 smoothes the input power Pin of the speech section, the latter is selected so as to give a weight to stability and to make the latter time constant longer.

  The second threshold value calculation unit 302 multiplies the second smoothing power P2 by a predetermined constant coefficient C2 (hereinafter referred to as the second coefficient) having a value greater than 0 and 1 or less, and compares the result with the input power Pin. The second threshold TH2 is formed and given to the second speech segment determination unit 303. The second smoothing power P1 means an average value of the voice power, and by multiplying this, the value of the second coefficient C2 that defines the second threshold TH2 for separating the voice power and the noise power. Although not limited, for example, 0.5 can be applied.

  The second speech segment determination unit 303 compares the second threshold value TH2 with the input power Pin to determine whether or not the speech segment is present, and outputs a second speech segment truth value V2. The second speech segment determination unit 303 outputs a true value as the second speech segment true / false value V2 if the input power Pin is greater than the second threshold value TH2, and otherwise outputs a false value.

  The first parameter calculation unit 111 and / or the second parameter calculation unit 112 described above may perform hangover frequently used in voice segment detection. The hangover will be clarified in the description of the operation described later.

  The second parameter calculation unit 112 outputs the second smoothing power P2 and the second speech interval truth value V2 obtained as described above as the second parameter F2.

(A-2) Operation of the First Embodiment Next, the operation of the noise suppression device 100 of the first embodiment having the above-described configuration will be described. First, after describing the overall operation of the noise suppression apparatus 100 of the first embodiment, the operations of the first parameter calculation unit 111 and the second parameter calculation unit 112 will be described in order, and further the hangover operation will be described. To do.

  In FIG. 1, an input signal to the noise suppression apparatus 100 of the first embodiment is given to a frequency analysis unit 101. In the frequency analysis unit 101, the input signal is subjected to frequency analysis to calculate a frequency spectrum, and the obtained input is obtained. The spectrum is given to the noise suppression unit 102 (102-1 to 102-M) for each band.

  In each noise suppression unit 102 for each band, the input frequency band signal is supplied to the power calculation unit 110 and the noise suppression unit 114, respectively.

  The power calculation unit 110 calculates the power of the input frequency band signal for TP seconds, and gives the obtained input power Pin to the first parameter calculation unit 111, the second parameter calculation unit 112, and the noise suppression unit 114. It is done.

  In the first parameter calculation unit 111, the second parameter F2 including the second speech section truth value V2 that is the detection result of the second parameter calculation unit 112 one unit time ago, and the input power Pin are obtained. The voice section detection operation is applied and the first parameter F1 including the obtained first voice section truth value V1 is given to the second parameter calculation unit 112 and the noise suppression unit 114.

  Also in the second parameter calculation unit 112, the first parameter F1 including at least the first voice section true / false value V1, which is the detection result of the first parameter calculation unit 111, and the input power Pin are applied and the voice is applied. The section detection operation is executed, and the second parameter F2 including at least the obtained second voice section truth value V2 is given to the first parameter calculation section 111 via the unit time delay section 113, and The noise suppression unit 114 is provided.

  In the noise suppression unit 114, the noise component of the input spectrum (input frequency band signal) is suppressed based on the first parameter F1, the second parameter F2, and the input power Pin, and the obtained output spectrum (suppression) The later frequency band signal) is given to the waveform restoration unit 103.

  Then, in the waveform restoration unit 103, the output spectrum composed of the spectrums (frequency band signals after noise suppression) given from all the noise suppression units 102 (102-1 to 102-M) for each band is converted into a signal in the time domain. The time domain signal obtained by the conversion is output as an output of the noise suppression apparatus 100 to the next stage apparatus.

  Next, the operation of the first parameter calculation unit 111 will be described with reference to FIG.

  In the first smoothing unit 201, the input power Pin is smoothed based on the first speech section reference truth value Vr1 (= second speech section truth value V2 one unit time before). That is, when the first speech section reference truth value Vr1 is a false value, the input power Pin is smoothed and the first smoothing power P1 is updated, while the first speech section reference truth value Vr1 is updated. When the value is true, the first smoothing power P1 is not updated and the immediately preceding first smoothing power P1 is maintained.

  The first smoothing power P1 obtained as described above is given to the first threshold value calculation unit 202 and the first speech segment determination unit 203. In the first threshold value calculation unit 202, the first smoothing power P1 is multiplied by a first coefficient C1 having a value of 1 or more. Then, in the first speech section determination unit 203, the first threshold value TH1 as the multiplication result is compared with the input power Pin, and when the input power Pin is larger than the first threshold value TH1, the first true value is obtained. Is output from the first speech segment determination unit 203 to the second parameter calculation unit 112 and the noise suppression unit 114, and when the input power Pin is less than or equal to the first threshold value TH1, The first speech segment truth value V1 is output from the first speech segment determination unit 203 to the second parameter calculation unit 112 and the noise suppression unit 114. The first smoothing power P1 described above is also output to the noise suppression unit 114.

  Next, the operation of the second parameter calculation unit 112 will be described with reference to FIG.

  In the second smoothing unit 301, the input power Pin is smoothed based on the second speech section reference truth value Vr2 (= the first speech section truth value V1 in the same unit time). That is, when the second speech section reference truth value Vr2 is a true value, the input power Pin is smoothed and the second smoothing power P2 is updated, while the second speech section reference truth value Vr2 is updated. When it is a false value, the second smoothing power P2 is not updated and the immediately preceding second smoothing power P2 is maintained.

  The second smoothing power P2 obtained as described above is provided to the second threshold value calculation unit 302 and the second speech segment determination unit 303. In the second threshold value calculation unit 302, the second smoothing power P2 is multiplied by a second coefficient C2 that is greater than 0 and less than or equal to 1. Then, in the second speech section determination unit 303, the second threshold value TH2 which is the multiplication result is compared with the input power Pin, and when the input power Pin is larger than the second threshold value TH2, the second value of the true value is obtained. Is output from the second speech segment determination unit 303 to the unit time delay unit 113 (and thus the first parameter calculation unit 111) and the noise suppression unit 114, and the input power Pin is set to the second threshold value TH2. In the following case, the false second speech section true / false value V2 is output to the unit time delay unit 113 (therefore, the first parameter calculation unit 111) and the noise suppression unit 114. The above-described second smoothing power P2 is also output to the noise suppression unit 114.

  In the above description, the hangover operation is not performed. However, at least one of the first speech segment determination unit 203 and the second speech segment determination unit 303 may perform the hangover operation.

  Hereinafter, the hangover operation executed by at least one of the first speech segment determination unit 203 and the second speech segment determination unit 303 will be described. Note that the hangover operation may be performed by both the first speech segment determination unit 203 and the second speech segment determination unit 303, or may be performed by one (however, the first speech segment determination unit). An embodiment in which both the 203 and the second speech segment determination unit 303 do not execute the hangover operation is also an embodiment of the present invention).

  Since the hangover operation in the first speech segment determination unit 203 and the hangover operation in the second speech segment determination unit 303 are the same, only the hangover operation in the first speech segment determination unit 203 will be described below. A description of the hangover operation in the second speech section determination unit 303 will be omitted.

  With respect to the first speech section determination unit 203, a predetermined hangover time Thn1 with respect to the first elapsed time Te1 since the last true value was output is determined in advance. When comparing the first threshold value TH1 and the input power Pin, the first speech segment determination unit 203 outputs a true first speech segment truth value V1 if (i) Pin> TH1. In addition, the first elapsed time Te1 is cleared to 0, and (ii) if Pin ≦ TH1 and Te1 ≦ Thn1, the first true voice interval truth value V1 is output and the first elapsed time is output. Te1 is incremented by one unit time, and (iii) if Pin ≦ TH1 and Te1> Thn1, a false first voice interval true / false value V1 is output.

  Here, when the hangover operation is performed in both the first speech segment determination unit 203 and the second speech segment determination unit 303, the hangover time Thn1 in the first speech segment determination unit 203 and the second speech segment are determined. The hangover time Thn2 in the section determination unit 303 may be the same or different. Below, the example made to make it differ is demonstrated. In the first embodiment, since the first speech section true / false value V1 is used for the estimation of the average value of the speech power in the second parameter calculation unit 112, the noise section is not erroneously determined as the speech section. In addition, the hangover time Thn1 of the first speech segment determination unit 203 is set short. On the other hand, since the second speech section truth value V2 is used for estimating the average value of the noise power in the first parameter calculation unit 111, the second speech section truth value V2 is used to prevent the voice section from being erroneously determined as the noise section. The hangover time Thn2 of the voice section determination unit 303 is set longer. For example, it is preferable to set the hangover time Thn1 in the first speech segment determination unit 203 to 0.1 seconds and the hangover time Thn2 in the second speech segment determination unit 303 to 0.2 seconds.

(A-3) Concept of reaching the configuration of the first embodiment Next, the concept of reaching the configuration of the noise suppression detection apparatus 100 of the first embodiment (the same applies to the embodiments described later) will be described.

  The approach of the prior art is to detect the speech section after estimating the noise power, or to estimate the noise power after detecting the speech section. The weighting factor in the technique described in Patent Document 1 is almost the same as detecting a speech section because it is 0 if the voice is strong and dominant and is 1 if the noise is strong and dominant. In such an approach, the estimation or detection performed earlier is inaccurate, and the estimation or detection performed later is also inaccurate.

  On the other hand, the first embodiment has a first parameter calculation unit and a second parameter calculation unit, and exchanges information with each other. That is, for example, assuming that both parameter calculation units are configured to detect the speech section after estimating the noise power, when the noise power is estimated by one parameter calculation unit, it is output from the other parameter calculation unit. By using the detection result of the speech section, it is possible to estimate the noise power with high accuracy. Practically, for example, the noise power is estimated by one parameter calculation unit and the sound power is estimated by the other parameter calculation unit (the first embodiment corresponds to this case). The accuracy is further improved by estimating and detecting the parameter calculation unit from different viewpoints.

  In this way, two parameter calculation units are prepared for each frequency band, and the parameters are organically calculated while exchanging parameters with each other, whereby noise power for each band (third embodiment to be described later). Then, estimation of SNR) and speech segment detection can be realized with high accuracy. Then, by applying these results to noise suppression, it is possible to obtain output speech in which speech is enhanced with less distortion.

(A-4) Effect of First Embodiment According to the first embodiment, two parameter calculation units supplement each other's update of parameters (estimated values of audio power and noise power). Therefore, since the parameter calculation accuracy and the speech interval detection accuracy can be improved, noise suppression with high naturalness and clarity can be realized.

(A-5) Modified Embodiment of First Embodiment In the description of the first embodiment described above, the second parameter calculation unit 112 (in other words, the second speech segment determination unit 303) performs a hangover operation. As described above, it has been explained that the hangover operation is not necessary. In the case of performing the hangover operation, the speech section truth value fed back to the first parameter calculation unit 111 and the speech section truth value output to the noise suppression unit 114 are also subjected to the hangover operation. If the over operation is not performed, neither the noise suppression unit 114 fed back to the first parameter calculation unit 111 nor the speech interval truth value output to the next stage device is subjected to the hangover operation. .

  FIG. 4 is a block diagram showing a configuration of the second parameter calculation unit 112A in the band-by-band suppression unit 102A obtained by modifying the first embodiment with a hangover surface.

  In the second parameter calculation unit 112A, in addition to the second voice segment determination unit 303 that does not execute the hangover operation, the second voice segment truth value V2 output from the second voice segment determination unit 303 is used. Is provided with a hangover unit 304 for executing a hangover operation. The second speech segment truth value V2 output from the second speech segment determination unit 303 is provided to the first parameter calculation unit 111 via the unit time delay unit 113 and also via the hangover unit 304. It is given to the next stage device.

  In the hangover unit 304, a predetermined hangover time Thn0 with respect to an elapsed time Te0 from when the true value is set to the voice section true / false value V0 output by itself is determined in advance. The hangover unit 304 (i) outputs the true speech interval truth value V0 and clears the elapsed time Te0 to 0 when the input second speech interval truth value V2 is a true value. (Ii) When the second speech section truth value V2 is a false value and Te0 ≦ Thn0, the true speech section truth value V0 is output and the elapsed time Te0 is incremented by one unit time. (Iii) If the second speech interval truth value V2 is a false value and Te0> Thn0, the false speech interval truth value V0 is output. The optimum value of the hangover time Thn0 varies depending on the use of the voice section true / false value V0. However, for example, 0.5 seconds is preferable when used for voice recognition.

(B) Second Embodiment Next, a noise removal apparatus and program according to a second embodiment of the present invention will be described with reference to the drawings.

  In the first embodiment, when calculating the smoothing power (particularly, the first smoothing power that means the average value of the noise power) and determining the speech section, a predetermined value is used for comparison between the smoothing power and the input power. Although constant coefficients C1 and C2 were used, the optimum coefficient differs depending on the power balance between voice and noise. Therefore, the parameters sent and received by the first parameter calculation unit and the second parameter calculation unit are only the speech section truth values in the first embodiment, but in the second embodiment, in addition to the speech section truth values. Thus, the threshold value is updated using the smoothing power including the smoothing power.

(B-1) Configuration of Second Embodiment The overall configuration of the noise removal apparatus (hereinafter, “100B” is used) of the second embodiment can also be expressed in FIG. 1 described above. However, the detailed configurations of the first parameter calculation unit (hereinafter referred to as “111B”) and the second parameter calculation unit (hereinafter referred to as “112B”) are different from those of the first embodiment. Therefore, hereinafter, the detailed configuration of the first parameter calculation unit 111B and the second parameter calculation unit 112B will be mainly described.

  FIG. 5 is a block diagram showing a detailed configuration of the first parameter calculation unit 111B in the second embodiment, which is the same as in FIG. 2 according to the first embodiment described above, the same corresponding parts, the same reference numerals Is shown.

  In FIG. 5, the first parameter calculation unit 111B includes a first smoothing unit 201, a first threshold value calculation unit 202B, and a first speech segment determination unit 203. Since the first smoothing unit 201 and the first speech segment determination unit 203 are the same as those in the first embodiment, their functional descriptions are omitted.

  The first threshold calculation unit 202B according to the second embodiment is configured to output the first smoothing power P1 output from the first smoothing unit 201 and the second parameter calculation unit 112B via the unit time delay unit 113. The first threshold value TH1B to be compared with the input power Pin is formed based on the first reference smoothing power Pr1 (= second smoothing power P2 before one unit time) given by This is given to the voice section determination unit 203.

Of the two values used to form the first threshold value TH1B, the first smoothing power P1 means an average value of the noise power, and the first reference smoothing power Pr1 is the sound one unit time before. Since it means an average value of power, it is preferable to apply an average value of the first smoothing power P1 and the first reference smoothing power Pr1 as the first threshold TH1B. The average value may be an arithmetic average (P1 + Pr1) / 2 or a geometric mean (P1 × Pr1) ^1/2 . When a value other than the average value is applied as the first threshold value TH1B, the first threshold value TH1B is set to a value larger than the first smoothing power P1 and smaller than the average value, and the first speech segment determination unit 203 uses the noise segment. It is preferable to increase the number of opportunities to be determined as voice segments. From the viewpoint of ease of calculation, the arithmetic average (P1 + Pr1) / 2 is preferable as the value of the first threshold value TH1B.

  In the case of the second embodiment, the first parameter calculation unit 111B includes the first smoothing power P1 output from the first smoothing unit 201 and the first output from the first speech segment determination unit 203. The first parameter F1 including the voice section true / false value V1 is supplied to the second parameter calculation unit 112B and the noise suppression unit 114.

  FIG. 12 is a block diagram showing a detailed configuration of the second parameter calculation unit 112B in the second embodiment, which is the same as in FIG. 3 according to the first embodiment described above, corresponding parts are the same, corresponding codes Is shown.

  In FIG. 12, the second parameter calculation unit 112 </ b> B includes a second smoothing unit 301, a second threshold calculation unit 302 </ b> B, and a second speech segment determination unit 303. Since the second smoothing unit 301 and the second speech segment determination unit 303 are the same as those in the first embodiment, their functional descriptions are omitted.

  The second threshold value calculation unit 302B in the second embodiment includes the second smoothing power P2 output from the second smoothing unit 301 and the second reference given from the first parameter calculation unit 111B. Based on the smoothing power Pr2 (= first smoothing power P1 of the same unit time), a second threshold value TH2B to be compared with the input power Pin is formed and given to the second speech section determination unit 303. is there.

Of the two values used to form the second threshold TH2B, the second smoothing power P2 means the average value of the voice power, and the second reference smoothing power Pr2 is the average of the noise power. Since it means a value, it is preferable to apply an average value of the second smoothing power P2 and the second reference smoothing power Pr2 as the second threshold TH2B. The average value may be an arithmetic average (P2 + Pr2) / 2 or a geometric mean (P2 × Pr2) ^1/2 . When a value other than the average value is applied as the second threshold value TH2B, the second threshold value TH2B is set to a value larger than the first smoothing power P1 and smaller than the average value. It is preferable to increase the number of opportunities to be determined as voice segments. From the viewpoint of ease of calculation, the arithmetic mean (P1 + Pr1) / 2 is preferable as the value of the second threshold TH2B.

  In the case of the second embodiment, the second parameter calculation unit 112B includes the second smoothing power P2 output from the second smoothing unit 301 and the second output from the second speech segment determination unit 303. The second parameter F2 including the voice section true / false value V2 is given to the first parameter calculation unit 111B via the unit time delay unit 104, and the second parameter F2 is given to the noise suppression unit 114.

(B-2) Operation of Second Embodiment Next, the operation of the noise suppression device 100B of the second embodiment will be described. Since the overall operation of the noise suppression device 100B of the second embodiment is also the same as the overall operation of the noise suppression device 100 of the first embodiment, description of the overall operation is omitted, and in the following, the second embodiment is the first embodiment. The operations of the first parameter calculation unit 111B and the second parameter calculation unit 112B, which are different from the first embodiment, will be described in order.

  First, the operation of the first parameter calculation unit 111B will be described with reference to FIG.

  In the first smoothing unit 201, the input power Pin is smoothed and obtained based on the first speech section reference truth value Vr1 (= second speech section truth value V2 before one unit time). The first smoothing power P1 is given to the first threshold value calculation unit 202B. The first threshold value calculation unit 202B is also given a first reference smoothing power Pr1, which is the second smoothing power P2 one unit time before. In the first threshold value calculation unit 202B, the first threshold value TH1B to be compared with the input power Pin is formed by the above-described method based on the first smoothing power P1 and the first reference smoothing power Pr1. To the first speech segment determination unit 203. Then, in the first speech segment determination unit 203, the first threshold value TH1B is compared with the input power Pin, and when the input power Pin is greater than the first threshold value TH1B, the true first speech segment true When the false value V1 is formed and the input power Pin is less than or equal to the first threshold value TH1B, the false first voice interval true / false value V1 is formed. Then, the first parameter F1 including the first smoothing power P1 output from the first smoothing unit 201 and the first speech section truth value V1 output from the first speech section determination unit 203. Is provided to the second parameter calculation unit 112B and the noise suppression unit 114.

  Next, the operation of the second parameter calculation unit 112B will be described with reference to FIG.

  In the second smoothing unit 301, the input power Pin is smoothed and obtained based on the second speech section reference truth value Vr2 (= first speech section truth value V1 of the same unit time). The second smoothing power P2 is given to the second threshold value calculation unit 302B. The second threshold value calculation unit 302B is also given a second reference smoothing power Pr2, which is the first smoothing power P1 of the same unit time. In the second threshold value calculation unit 302B, the second threshold value TH2B to be compared with the input power Pin is formed by the above-described method based on the second smoothing power P2 and the second reference smoothing power Pr2. To the second speech segment determination unit 303. Then, in the second speech segment determination unit 303, the second threshold TH2B is compared with the input power Pin, and when the input power Pin is greater than the second threshold TH2B, the true second speech segment true When the false value V2 is formed and the input power Pin is less than or equal to the second threshold value TH2B, the false second voice interval true / false value V2 is formed. Then, the second parameter F2 including the second smoothing power P2 output from the second smoothing unit 301 and the second speech section truth value V2 output from the second speech section determination unit 303. Is provided to the first parameter calculation unit 111B via the unit time delay unit 113, and the second parameter F2 is provided to the noise suppression unit 114.

  Also in the second embodiment, as in the first embodiment, a hangover operation may be performed on at least one of the first speech segment determination unit 203 and the second speech segment determination unit 303. In the case where both hangover operations are performed, the hangover time in the first speech segment determination unit 203 and the hangover time in the second speech segment determination unit 303 may be the same or different. Also in the second embodiment, it is preferable that the hangover time in the first speech segment determination unit 203 is 0.1 seconds and the hangover time in the second speech segment determination unit 303 is 0.2 seconds. is there.

  Similarly to the first embodiment, the second embodiment can be modified with the hangover portion 15 as shown in FIG. The hangover time in the hangover portion 15 is preferably 0.5 seconds.

(B-3) Effect of Second Embodiment Also according to the second embodiment, the first parameter calculation unit and the second parameter calculation unit supplement each other's parameter updates, thereby stably updating the parameters. Therefore, it is possible to improve the parameter calculation accuracy and the speech interval detection accuracy, so that noise suppression with high naturalness and clarity can be realized.

  In addition, according to the second embodiment, even when the power balance between voice and noise is unknown or when this power balance fluctuates with time, the threshold value to be compared with the input power can be updated appropriately. Also from this point, the parameter calculation accuracy and the speech interval detection accuracy can be improved, and noise suppression with high naturalness and clarity can be realized.

(C) Third Embodiment Next, a third embodiment of the noise suppression device and program according to the present invention will be described with reference to the drawings.

  The third embodiment is different from the first and second embodiments in that the second parameter calculation unit estimates the SNR (here, the posterior SNR) and provides it to the noise suppression unit.

(C-1) Configuration of the Third Embodiment The overall configuration of the noise suppression device (hereinafter, using the symbol “100C”) of the third embodiment can also be represented in FIG. 1 described above.

  However, in the third embodiment, the first parameter calculation unit 111 performs parameter estimation and voice segment detection based on the input power Pin, as in the first embodiment, but the second parameter calculation. The unit (hereinafter, the code “112C” is used) estimates an SNR (here, a posteriori SNR), and detects a speech section based on the estimated SNR. In addition, the noise suppression unit (hereinafter referred to as “114C”) performs noise suppression also using the SNR output from the second parameter calculation unit 112C.

  Therefore, the detailed configuration of the second parameter calculation unit 112C and the function of the noise suppression unit 114C will be mainly described below. The first parameter calculation unit 111 has the detailed configuration shown in FIG. 2 described above.

  FIG. 7 is a block diagram showing a detailed configuration of the second parameter calculation unit 112C in the third embodiment, and is the same as in FIG. 3 according to the first embodiment described above, corresponding parts are the same, Is shown.

  In FIG. 7, the second parameter calculation unit 112C of the third embodiment includes an SNR calculation unit 305 in addition to the second smoothing unit 301C, the second threshold value calculation unit 302C, and the second speech segment determination unit 303C. Have

  The SNR calculation unit 305 uses the input power Pin (corresponding to S of SNR) as a second reference smoothing power Pr2 (= first smoothing power P1 before one unit time; SNR) that is an estimated value of noise power; SNR estimated value Ri is obtained by dividing by R), and the obtained SNR estimated value Ri is given to second smoothing section 301C and second speech section determining section 303C.

  Unlike the first and second embodiments, the second smoothing unit 301C in the third embodiment smooths not the input power Pin but the SNR estimated value Ri. The second smoothing unit 301C smoothes the SNR estimated value Ri based on the second speech section reference truth value Vr2 (= first speech section truth value V1 before one unit time), and is obtained. The SNR smoothing value Rs is given to the second threshold value calculation unit 302C. The second smoothing unit 301C smoothes the SNR estimated value Ri and updates the SNR smoothed value Rs when the second speech section reference truth value Vr2 is a true value (that is, a speech section), When the speech section reference true / false value Vr2 is a false value (that is, a noise section), the SNR smoothing value Rs is maintained without being updated. Therefore, the SNR smoothing value Rs means the average SNR of the speech section. The smoothing method is not limited at all. For example, a time constant filter with a time constant of 0.8 seconds is suitable.

  Since the SNR smoothing value Rs means the SNR of the speech section, the second threshold value calculation unit 302C in the third embodiment takes a constant value greater than 0 and 1 or less as the SNR smoothing value Rs. Is multiplied by the coefficient C2C to form a second threshold value TH2C to be compared with the SNR estimated value Ri and provided to the second speech segment determination unit 303C. By multiplying the SNR smoothing value Rs, the value of the second coefficient C2C that defines the second threshold TH2C for separating the SNR estimated value of the speech section and the SNR estimated value of the noise section is not limited. For example, 0.5 can be applied.

  The speech segment determination unit 303C in the third embodiment compares the SNR estimated value Ri and the second threshold value TH2C to form a second speech segment truth value V2 that indicates whether or not the speech segment is a speech segment. The second speech segment determination unit 303C outputs a true value as the second speech segment truth value V2 if the SNR estimated value Ri is larger than the second threshold value TH2C, and otherwise outputs a false value.

  From the second parameter calculation unit 112C, the SNR estimated value Ri and the second speech interval truth value V2 are output.

  The noise suppression unit 114C of the third embodiment suppresses the noise component of the input spectrum (frequency band signal) based on the first parameter F1 and the second parameter F2C and the input power Pin, and the suppressed spectrum is obtained. The waveform restoration unit 103 is provided. In the third embodiment, as described above, the SNR estimated value Ri is included in the second parameter F2C, and the noise suppression unit 114C does not need to calculate the a posteriori SNR necessary for calculating the suppression gain, The given SNR estimated value Ri is used as it is. The noise suppression method applied by the noise suppression unit 114C is not limited, but the Wiener filter method and MMSE-STSA are preferable from the viewpoint that the SNR estimation value Ri for each frequency band can be used.

(C-2) Operation of the Third Embodiment Next, the operation of the noise suppression device 100C of the third embodiment will be described. Since the overall operation of the noise suppression device 100C of the third embodiment is also the same as the overall operation of the speech segment detection device 100 of the first embodiment, description of the overall operation is omitted. In addition, since the operation of the first speech section detection unit 102 in the third embodiment is the same as that of the first embodiment, description of the operation is omitted, and the second operation in the third embodiment is described below. The operation of the speech segment detection unit 103C will be described, and the operation of the noise suppression unit 114C to which the output of the second speech segment detection unit 103C is given will also be described.

  In FIG. 7, the SNR calculation unit 305 is given an input power Pin and a second reference smoothing power Pr2 (first smoothing power P1 one unit time before) that is an estimated value of noise power. The power Pin is divided by the second reference smoothing power Pr2 to obtain an estimated SNR value Ri, and the obtained SNR estimated value Ri is given to the second smoothing unit 301C and the second speech segment determination unit 303C. It is done.

  The SNR estimated value Ri is smoothed by the second smoothing unit 301C with reference to the second speech section reference truth value Vr2 (= first speech section truth value V1 one unit time before). That is, when the second speech section reference truth value Vr2 is a true value (that is, speech section), the SNR estimated value Ri is smoothed and the SNR smoothed value Rs is updated, and the second speech section reference truth value is updated. When Vr2 is a false value (that is, a noise interval), the SNR smoothed value Rs is maintained without being updated, and the SNR smoothed value Rs thus obtained is given to the second threshold value calculation unit 302C. Then, in the second threshold value calculation unit 302C, the SNR smoothed value Rs is multiplied by a second coefficient C2C having a constant value greater than 0 and equal to or less than 1, and the second threshold value is compared with the SNR estimated value Ri. TH2C is formed and provided to the second speech segment determination unit 303C.

  The SNR estimated value Ri and the second threshold value TH2C are compared in the second speech segment determination unit 303C, and the true second speech segment truth value V2 when the SNR estimated value Ri is greater than the second threshold value TH2C. Is output, and when the SNR estimated value Ri is equal to or smaller than the second threshold value TH2C, the false second speech section true / false value V2 is output.

  From the second parameter calculation unit 112C, the second parameter F2C including the SNR estimated value Ri and the second speech interval truth value V2 is output, and the second parameter F2C is provided to the noise removal unit 114C. The second speech section truth value V2 is given to the first parameter calculation unit 111 via the unit time delay unit 113.

  In the noise suppression unit 114C, the noise component of the input spectrum (frequency band signal) is suppressed based on the first parameter F1 and the second parameter F2C and the input power Pin, and the suppressed spectrum is sent to the waveform restoration unit 103. Given. Here, since the SNR estimation value Ri is included in the second parameter F2C described above, the given SNR estimation is performed without calculating the a posteriori SNR necessary for calculating the suppression gain in the noise suppression unit 114C. The value Ri is used as it is.

  Also in the third embodiment, as in the first embodiment, the hangover operation may be performed on at least one of the first voice segment determination unit 203 and the second voice segment determination unit 303C. In the case of performing both hangover operations, the hangover time in the first speech segment determination unit 203 and the hangover time in the second speech segment determination unit 303C may be the same or different. Also in the third embodiment, it is preferable that the hangover time in the first speech segment determination unit 203 is 0.1 second and the hangover time in the second speech segment determination unit 303C is 0.2 seconds. is there.

  Similarly to the first embodiment, the third embodiment can be modified to have the hangover portion 15 as shown in FIG. The hangover time in the hangover portion 15 is preferably 0.5 seconds.

(C-3) Effect of Third Embodiment According to the third embodiment, estimation and determination based on the power of the input signal by the first parameter calculation unit, and SNR in the input signal by the second parameter calculation unit Estimation and determination based on each parameter and determination result are used to estimate and determine each other, so that parameter calculation accuracy and speech segment detection accuracy can be improved, and noise with high naturalness and clarity Repression can be realized.

(C-4) Modified Embodiment of Third Embodiment In the above description, the first parameter calculation unit detects a voice section based on input power, and the second parameter calculation unit detects a voice section based on SNR. However, the first parameter calculation unit may detect the voice interval based on the SNR, and the second parameter calculation unit may detect the voice interval based on the input power. Both the first parameter calculation unit and the second parameter calculation unit may detect a speech section based on the SNR.

(D) Other Embodiments In the description of each of the above embodiments, various modified embodiments have been mentioned, and further modified embodiments as exemplified below can be given.

  In each of the above-described embodiments, the noise suppression unit for each band functions independently. However, it may be configured to influence each other statically or dynamically. For example, the first parameter calculation unit and the second parameter calculation unit of some band-specific noise suppression units include only a smoothing unit, and other band-specific noise suppression units including a threshold calculation unit and a voice section detection unit. You may make it operate | move by taking in a speech section truth value. Further, for example, the first speech section true / false values from the first parameter calculation unit in all the noise suppression means for each band are integrated by majority vote, logical product, or logical sum, and the second in the noise suppression means for all the bands. The second speech interval truth value from the second parameter calculation unit in all the noise suppression means for all bands is integrated by majority vote, logical product or logical sum, etc., and further for one unit time You may make it delay and give it to the 1st parameter calculation part in all the noise suppression means classified by band.

  In each of the above embodiments, the first parameter calculation unit updates the predetermined feature amount when the detection result of one unit time before the second parameter calculation unit indicates the noise period, and the second parameter calculation The unit has shown the case where the predetermined feature amount is updated when the detection result of the same unit time of the first parameter calculation unit indicates the voice period, but the combination of the periods for updating the feature amount is limited to this. It is not something. For example, the first parameter calculation unit may update the feature amount in the voice period, and the second parameter calculation unit may update in the noise period. The first parameter calculation unit and the second parameter calculation unit may Both may be updated in the noise interval, and both the first parameter calculation unit and the second parameter calculation unit may be updated in the voice interval. Depending on the selection of the update period, the first parameter calculation unit may detect the speech section based on the SNR.

  In each of the above embodiments, the speech section detection result (second speech section true / false value) output by the second parameter calculation unit is output to the noise suppression unit. However, the speech section output to the noise suppression unit The detection result is not limited to this. For example, the detection result of the speech section (first speech section truth value) by the first parameter calculation unit may be output to the noise suppression unit, and the detection result of the speech section by the first parameter calculation unit You may make it output the logical product and logical sum of the detection result of the audio | voice area by a 2nd parameter calculation part to a noise suppression part.

  In the first and second embodiments, the case where the first parameter calculation unit and the second parameter calculation unit have completely different configurations has been described. (A conversion unit, a threshold value calculation unit, and a speech segment determination unit) may be applied in a time division manner within one unit time so as to function as a first parameter calculation unit and a second parameter calculation unit. In this case, when functioning as the first parameter calculation unit, the data related to the second parameter calculation unit (for example, the second smoothing power P2, the second coefficient C2, etc.) is saved and the second When functioning as a parameter calculation unit, an auxiliary configuration such as a memory for saving data related to the first parameter calculation unit (for example, the first smoothing power P1, the first coefficient C1, etc.) is provided. Cost. The expression of a claim shall include the case where such an identical structure is used by time division.

  In each of the above-described embodiments, the threshold value used for speech segment detection is determined based on the smooth value of the feature amount. However, the threshold value may be determined by another method. For example, a predetermined multiple of the minimum value of the input power in a predetermined period (for example, 3 seconds; when the noise period is intermittently generated, 3 seconds in the past) determined as the noise period is set as a threshold value. Alternatively, the threshold value may be a predetermined multiple of the maximum value of the input power in a predetermined period (for example, 3 seconds) immediately before the determination of the voice section. Further, when SNR is used as a feature quantity as in the third embodiment, a predetermined multiple of the maximum value of SNR in a predetermined period (for example, 3 seconds) immediately before the determination as a speech section is used as a threshold value. Anyway.

  In each of the embodiments described above, the feature amount obtained by the first parameter calculation unit and / or the second parameter calculation unit is used for noise removal. However, the feature amount may be used for other purposes. good.

  FIG. 8 is a block diagram showing the configuration of an embodiment of the noise estimation apparatus according to the present invention, in which the same and corresponding parts as in FIG.

  In FIG. 8, the noise estimation apparatus 400 of this embodiment includes a frequency analysis unit 101, band-specific noise power estimation units 402-1 to 402-M, and a noise power integration unit 403 similar to those of the first embodiment. Each band-specific noise power estimation unit 402 (402-1 to 402-M) includes a power calculation unit 110 similar to that in the first embodiment, a first parameter calculation unit 111 similar to that in the first embodiment, and a first parameter. The second parameter calculation unit 112 similar to that of the first embodiment is provided, and the first smoothing power P1 obtained inside the first parameter calculation unit 111 is integrated with the noise power from each band noise power estimation unit 402. To the unit 403. The noise power integration unit 403 integrates the first smoothed powers P1-1 to P1-M from all the band-specific noise power estimation units 402-1 to 402-M into a noise power spectrum. Integration at this time may be creation of a vector in which values of each frequency band are assigned to vector elements, summation, or calculation of an average value (may be a weighted average value). May be.

  FIG. 9 is a block diagram showing a configuration of an embodiment of the SNR estimation apparatus according to the present invention, in which the same and corresponding parts as in FIG.

  In FIG. 9, the SNR estimation apparatus 500 of this embodiment includes a frequency analysis unit 101, band-specific SNR estimation units 502-1 to 502-M, and an SNR integration unit 503 similar to those of the first embodiment. Each SNR estimation unit 502 (502-1 to 502-M) for each band includes a power calculation unit 110 similar to that in the first embodiment, a first parameter calculation unit 111 similar to that in the first embodiment, and a third parameter calculation unit. A second parameter calculation unit 112C similar to that of the embodiment is provided, and the SNR estimation value Ri obtained inside the second parameter calculation unit 112C is given to the SNR integration unit 503 from each band-specific SNR estimation unit 502. The SNR integration unit 503 forms SNR estimation values that are output by integrating the SNR estimation values Ri from the band-specific SNR estimation units 502-1 to 502-M. As a method for integrating the SNR estimated values Ri, for example, a method of taking an average value of the SNR estimated values Ri of the entire frequency band can be cited. The method of obtaining the average value may be a method of obtaining an average value of the SNR estimated value Ri, or a method of obtaining the average value after converting the SNR estimated value Ri into a decibel (logarithmic scale). (In this case, the output may be left in decibels or may be restored to the original scale).

  In each of the above embodiments, the case where the target sound to be compared with the noise is a voice is shown, but the present invention is not limited to this. For example, the technical idea of the present invention can also be applied when the motor sound of a machine is a target sound compared with noise.

  DESCRIPTION OF SYMBOLS 100, 100A, 100B, 100C ... Noise suppression apparatus, 101 ... Frequency analysis part, 102-1 to 102-M ... Noise suppression means classified by band, 103 ... Waveform restoration part, 110 ... Power calculation part, 111, 111B ... 1st 112, 112A, 112B, 112C ... second parameter calculation unit, 113 ... unit time delay unit, 114 ... noise suppression unit, 201 ... first smoothing unit, 202, 202B ... first threshold Calculation unit, 203... First speech section determination unit, 301, 301C, second smoothing unit, 302, 302B, 302C, second threshold calculation unit, 303, 303C, second speech section determination unit, 304 Hangover unit, 305 ... SNR calculation unit, 400 ... noise estimation device, 402-1 to 402-M ... noise power estimation unit for each band, 403 ... noise power integration unit, 50 ... SNR estimation device, 502-1 to 502-M ... band domain SNR estimator, 503 ... SNR integration unit.

Claims (12)

In the noise suppression device that suppresses the noise component contained in the input signal and emphasizes the target sound component,
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
Corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit, comprising a plurality of band-specific noise suppression means for suppressing noise components in the signal of the frequency band,
The noise suppression means for each band is as follows:
The first feature quantity based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency in the input frequency band signal is compared. A first parameter calculation unit for obtaining a detection result of the target sound section;
The second feature amount based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second feature value in the input frequency band signal is compared. A second parameter calculation unit for obtaining a detection result of the target sound section;
Noise that suppresses a noise component included in the input frequency band signal based on the first parameter obtained by the first parameter calculation unit and the second parameter obtained by the second parameter calculation unit A repressor,
The first parameter calculation unit uses the second parameter output by the second parameter calculation unit before a predetermined unit time and includes at least a detection result of the second target sound section. Generate a threshold,
The second parameter calculation unit uses the second parameter that is output by the first parameter calculation unit in the same unit time and includes at least the detection result of the first target sound section, and then uses the second parameter. A noise suppressor characterized by generating The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
A second smoothing power is calculated by smoothing the second input power while controlling the execution and stop of the smoothing based on the detection result of the first target sound section of the same unit time. A smoothing section of
A second threshold value calculation unit for calculating the second threshold value by applying at least the second smoothing power;
The second target sound is obtained by comparing the second input power as the second feature amount with the second threshold value to determine whether the target sound section is present, and obtaining a detection result of the second target sound section. The noise suppression device according to claim 1, further comprising: an interval determination unit. The first smoothing unit performs smoothing of the first input power when a detection result of the second target sound section before a predetermined unit time is not the target sound section, and performs a predetermined unit time before When the detection result of the second target sound section is the target sound section, the smoothing of the first input power is stopped and the first smoothed power is maintained.
The second smoothing unit performs the second input power smoothing when the detection result of the first target sound section of the same unit time is the target sound section, and the second smoothing unit performs the above-mentioned of the same unit time. 3. The smoothing of the second input power is stopped when the detection result of the first target sound section is not the target sound section, and the second smoothing power is maintained. 4. Noise suppression device. One of the first smoothing unit and the second smoothing unit performs smoothing when the detection result of the target sound section input to itself is the target sound section, and the other Smoothing is performed when the input detection result of the target sound section is not the target sound section.
The first threshold value calculation unit calculates the first threshold value by applying the first smoothing power and the second smoothing power before a predetermined unit time,
The second threshold value calculation unit calculates the second threshold value by applying the first smoothing power and the second smoothing power in the same unit time. The noise suppressor described.   The first threshold value calculating unit calculates an arithmetic average or a geometric average of the first smoothing power and the second smoothing power before a predetermined unit time as the first threshold value. The noise suppression device according to claim 4.   The second threshold value calculation unit calculates an arithmetic average or a geometric average of the first smoothing power and the second smoothing power in the same unit time as the second threshold value. The noise suppression device according to claim 4 or 5. The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
An SNR calculator that calculates an estimated value of SNR based on the second input power and the first smoothed power of the same unit time;
Second smoothing for smoothing the SNR estimated value and calculating the smoothed value of the SNR while controlling the execution and stop of smoothing based on the detection result of the first target sound section of the same unit time And
A second threshold value calculation unit for calculating the second threshold value by applying at least the SNR smooth value;
The SNR estimation value is compared with the second threshold value as the second feature amount to determine whether or not the target sound section is present, and second target sound section determination for obtaining a detection result of the second target sound section The noise suppression device according to claim 1, further comprising: In a noise estimation device that estimates noise power in an input signal,
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
Corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit, a plurality of band-specific noise estimation means for estimating the noise power in the signal of the frequency band,
The noise estimation means for each band obtained by the noise estimation means for each band obtained by integrating the noise noise estimation values for each frequency band to obtain a final noise power estimation value,
The noise estimation means for each band is respectively
The first feature amount based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency value in the input frequency band signal is compared . A first parameter calculation unit for detecting a target sound section of
The second feature value based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second frequency value in the input frequency band signal is compared . A second parameter calculation unit for detecting the target sound section of
The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
A second smoothing power is calculated by smoothing the second input power while controlling the execution and stop of the smoothing based on the detection result of the first target sound section of the same unit time. A smoothing section of
A second threshold value calculation unit for calculating the second threshold value by applying at least the second smoothing power;
The second target sound is obtained by comparing the second input power as the second feature amount with the second threshold value to determine whether the target sound section is present, and obtaining a detection result of the second target sound section. A section determination unit,
The first smoothing unit or the second smoothing unit receives a detection result of the second target sound section before a predetermined unit time or a detection result of the first target sound section of the same unit time. Smoothing if it is not the target sound section, stopping smoothing if it is the target sound section, and obtaining the first smoothing power or the second smoothing power as an estimate of noise power for each band. A characteristic noise estimation apparatus. In an SNR estimation device for estimating an SNR in an input signal,
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
A plurality of band-specific SNR estimation means for estimating the SNR of a signal in the frequency band corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit;
Band-specific SNR integration means for obtaining a final SNR estimate value by integrating a plurality of SNR estimation values for each frequency band obtained by the band-specific SNR estimation means,
Each of the band-specific SNR estimation means is respectively
The first feature amount based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency value in the input frequency band signal is compared . A first parameter calculation unit for detecting a target sound section of
The second feature value based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second frequency value in the input frequency band signal is compared . A second parameter calculation unit for detecting the target sound section of
The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
An SNR calculator that calculates an estimated value of SNR based on the second input power and the first smoothed power of the same unit time;
Second smoothing for smoothing the SNR estimated value and calculating the smoothed value of the SNR while controlling the execution and stop of smoothing based on the detection result of the first target sound section of the same unit time And
A second threshold value calculation unit for calculating the second threshold value by applying at least the SNR smooth value;
The SNR estimation value is compared with the second threshold value as the second feature amount to determine whether or not the target sound section is present, and second target sound section determination for obtaining a detection result of the second target sound section And
The SNR estimation apparatus, wherein the SNR estimation value from the SNR calculation unit is obtained as an SNR estimation value of the frequency band in the input frequency band signal. A noise suppression program that suppresses a noise component contained in an input signal and emphasizes a target sound component,
Computer
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
Corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit, to suppress the noise component in the signal of the frequency band, to function as a plurality of band-specific noise suppression means,
The noise suppression means for each band is as follows:
The first feature quantity based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency in the input frequency band signal is compared. A first parameter calculation unit for obtaining a detection result of the target sound section;
The second feature amount based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second feature value in the input frequency band signal is compared. A second parameter calculation unit for obtaining a detection result of the target sound section;
Noise that suppresses a noise component included in the input frequency band signal based on the first parameter obtained by the first parameter calculation unit and the second parameter obtained by the second parameter calculation unit A repressor,
The first parameter calculation unit uses the second parameter output by the second parameter calculation unit before a predetermined unit time and includes at least a detection result of the second target sound section. Generate a threshold,
The second parameter calculation unit uses the second parameter that is output by the first parameter calculation unit in the same unit time and includes at least the detection result of the first target sound section, and then uses the second parameter. A noise suppression program characterized by generating A noise estimation program for estimating noise power in an input signal,
Computer
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
Corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit, a plurality of band-specific noise estimation means for estimating the noise power in the signal of the frequency band,
The noise estimation means for each band obtained by each of the above bands functions as a noise power integration means for each band to obtain a final noise power estimation value by integrating a plurality of noise power estimation values for each frequency band,
The noise estimation means for each band is respectively
The first feature amount based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency value in the input frequency band signal is compared . A first parameter calculation unit for detecting a target sound section of
The second feature value based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second frequency value in the input frequency band signal is compared . A second parameter calculation unit for detecting the target sound section of
The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
A second smoothing power is calculated by smoothing the second input power while controlling the execution and stop of the smoothing based on the detection result of the first target sound section of the same unit time. A smoothing section of
A second threshold value calculation unit for calculating the second threshold value by applying at least the second smoothing power;
The second target sound is obtained by comparing the second input power as the second feature amount with the second threshold value to determine whether the target sound section is present, and obtaining a detection result of the second target sound section. A section determination unit,
The first smoothing unit or the second smoothing unit receives a detection result of the second target sound section before a predetermined unit time or a detection result of the first target sound section of the same unit time. Smoothing if it is not the target sound section, stopping smoothing if it is the target sound section, and obtaining the first smoothing power or the second smoothing power as an estimate of noise power for each band. Feature noise estimation program. An SNR estimation program for estimating an SNR in an input signal,
Computer
A frequency analysis unit that performs frequency analysis of the input signal to calculate an input spectrum;
A plurality of band-specific SNR estimation means for estimating the SNR of a signal in the frequency band corresponding to the frequency band of any input spectrum calculated by the frequency analysis unit;
It functions as a band-specific SNR integration unit that obtains a final SNR estimation value by integrating a plurality of SNR estimation values for each frequency band obtained by each band-specific SNR estimation unit,
Each of the band-specific SNR estimation means is respectively
The first feature amount based on the first input power calculated for the input frequency band signal is compared with the first threshold value generated internally, and the first frequency value in the input frequency band signal is compared . A first parameter calculation unit for detecting a target sound section of
The second feature value based on the second input power calculated for the input frequency band signal is compared with the second threshold value generated internally, and the second frequency value in the input frequency band signal is compared . A second parameter calculation unit for detecting the target sound section of
The first parameter calculation unit includes:
A first smoothing power is calculated by smoothing the first input power while controlling execution and stop of smoothing based on the detection result of the second target sound section before a predetermined unit time. 1 smoothing unit;
A first threshold value calculation unit for calculating the first threshold value by applying at least the first smoothing power;
The first target sound that obtains the detection result of the first target sound section by comparing the first input power with the first threshold value as the first feature amount to determine whether the target sound section is present or not. A section determination unit,
The second parameter calculation unit includes:
An SNR calculator that calculates an estimated value of SNR based on the second input power and the first smoothed power of the same unit time;
Second smoothing for smoothing the SNR estimated value and calculating the smoothed value of the SNR while controlling the execution and stop of smoothing based on the detection result of the first target sound section of the same unit time And
A second threshold value calculation unit for calculating the second threshold value by applying at least the SNR smooth value;
The SNR estimation value is compared with the second threshold value as the second feature amount to determine whether or not the target sound section is present, and second target sound section determination for obtaining a detection result of the second target sound section And
The SNR estimation program characterized in that the SNR estimation value from the SNR calculation unit is obtained as an estimated value of SNR in the frequency band in the input frequency band signal.

Images