JP3423906B2 - Voice operation characteristic detection device and detection method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION Voice characteristic detector (vo)
An ice activity detector is a device supplied with a signal whose purpose is to detect the duration of a conversation, or the duration of noise alone. The invention is not limited to these applications, and a particular embodiment of the invention for such detectors is a mobile radiotelephone system in which speech is spoken by a speech coder.
der) to improve the effective utilization of the radio spectrum, and in these systems the noise level (from the onboard units) is generally high.

[0002]

The essence of voice behavior characteristic detection is to look for different amounts between speech and non-speech periods. In a device containing a conversational coder, many parameters can easily be used from one coder, or from another stage, and by utilizing such parameters, It is desirable to economically simplify the processing required. In many situations, the dominant noise occurs within a limited region of some frequency spectrum. For example, noise (for example, engine noise) of a moving car (automobile) has a low frequency band spectrum. Where knowledge of such positions in the noise spectrum is available, it is desirable to base the decision on whether speech is present on the measure obtained from the spectral portion containing relatively little noise. Of course, it is actually possible to filter the signal before detecting and analyzing the behavioral characteristics of the speech, but if the speech behavioral characteristic detector relies on the output of the speech coder, then this Filtering interferes with the coded audio signal.

[0003]

According to the present invention, the embodiment of which is shown in FIG. 3, an input signal component is generated in order to generate an output signal indicating whether or not conversation is present in the input signal. A first operative to form a value of spectral similarity between the components of the input signal determined to be speech-free
A voice activity characteristic detector; a memory for storing data derived from the non-speech portion; and an auxiliary voice activity characteristic detector, the auxiliary voice activity characteristic detector updating the memory. And the auxiliary behavioral characteristic detector operates to form a spectral similarity value between the most recent component of the input signal and the earlier component of the input signal. Provided.

Further, it is a method for detecting a conversational operation characteristic of an input signal, in which the input signal is received, a noise signal portion of the input signal is roughly estimated, data representing the noise signal portion is stored, and one of the input signal Form a value M of the spectral similarity between the part and the noise signal part, the threshold value (thr
The parameter derived from the above value M is compared with a first threshold value T, in order to create a first voice behavior characteristic indication indicating whether a conversation is present or absent according to whether it is higher than the threshold value), The step of approximating in the step of generating an auxiliary voice behavioral characteristic indication includes the generation of the auxiliary voice behavioral characteristic indication of the similarity between the latest component of the input signal and the earlier component of the input signal. The spectral distortion value of the second threshold value to generate an indication of whether speech is present or absent, corresponding to whether the spectral distortion value is exceeded or not exceeded. A method is provided which includes the step of updating the stored data with an input signal only during a period when the active voice behavior characteristic indication indicates the absence of speech.

The above value M is preferably a distortion value by Itakura and Saito.

Other aspects of the invention are within the scope of the following claims.

[0007]

Some embodiments of the present invention will now be described with reference to the accompanying drawings.

The general principles which characterize the first embodiment of the speech motion characteristic detector according to the invention will now be given.

The n signal samples (s ₀ , s ₁ , s ₂ , s ₃ , s ₄ ... s _n-1 ) are
When passed through a conceptual 4th order finite pulse response (FIR) digital filter of pulse response (1, h ₀ , h ₂ , h ₃ ), it becomes a filtered signal (ignoring the sample from the previous frame. ),

[0010]

[Equation 1]

The zero-order autocorrelation coefficient is the sum of the squares of each term, which is normalized, that is, divided by the total number of terms (for a constant frame length, omitting that division. Easy), so the sum of the filtered signals is

[0012]

[Equation 2]

This is therefore the logically filtered signal s
′ Power, ie the signal s in the passband of the conceptual filter
Is the electric energy of the part.

Ignoring the first four terms and expanding,

[0015]

[Equation 3]

[0016] Thus, _R'0 is obtained by coupling of the autocorrelation coefficients R _i, weighted by constants in parentheses to determine the frequency band to a value _R'0 responds. In fact, the bracketed terms are the autocorrelation coefficients of the pulse response of the logic filter, so the above expression can simply be expressed as:

[0017]

[Equation 4]

Where N is the order of the filter and H _i is the (unnormalized) autocorrelation coefficient of the pulse response of the filter.

That is, the effect of the signal filtering on the signal autocorrelation coefficient is to simulate by using the pulse response of the required filter and generating the sum of the autocorrelation coefficients of the (unfiltered) signal. (Simulat
e) can be done.

Thus, a relatively simple algorithm involving a small number of multiply operations can perform a digital filter simulation that generally requires this number of 100 multiply operations.

On the other hand, in this filtering operation, the signal spectrum is matched with the reference spectrum (match).
d) in the state (negative phase response of the logic filter), it can be seen as a form of spectral comparison. Since the logical filter in this application is chosen to approximate the inverse of the noise spectrum, this behavior is done with the spectrum of speech and noise, such as values that indicate dissimilarities between the spectra, and the 0th order generated. It can be viewed as a spectral comparison with the autocorrelation coefficient (ie the energy of the back-filtered signal). The distortion values by Itakura and Saito are predicted by the prediction filter (pre
Used in the LPC to evaluate the matching of the input filter and the input filter, one format is shown as follows.

[0022]

[Equation 5]

Here, A _{0 and the} like are autocorrelation coefficients of the LPC parameter set. It turns out that this is very similar to the relationship obtained above, where the LPC coefficients are the taps of the FIR with the inverse spectral response of the input signal.
Therefore, considering that the LPC coefficient set is the pulse response of the inverse LPC filter,
Saito's distortion value is simply one form of Equation 1, where the filter response H is the all-pol model (all-p) of the input signal.
ole model).

In fact, the LPC coefficient of the test spectrum and the autocorrelation coefficient of the reference spectrum can be used to transform and obtain different values of spectral similarity.

The distortion value due to the IS is "encoding speech based on vector quantization"("SpeechCodin").
g based upon Vector Quant
isation ”by A Buzo, A H Gra
y, RM Gray and J D Marke
l, IEEE Trans on ASSP, Vol A
SSP-28, No. 5, October 1980).

The above results are only approximations, since the frame of the signal simply has a finite value length and the number of terms (N, where N is the filter order) is ignored. However, it very well indicates if there is a conversation and is therefore used as the value M in the conversation report. The noise spectrum is known,
If it is static noise, it is quite possible to apply fixed h ₀ , h ₁ etc coefficients to the inverse noise filter.

However, a device that can adapt to different noise situations is more beneficial.

FIG. 1 shows a first embodiment of the present invention in which a signal s from a microphone (not shown) is input 1
Received by the analog-to-digital converter 2 and the appropriate sampling rate (sampling r).
ate) and converted to digital samples. LPC analysis unit 3 (general LPC coder)
Obtains a set of N (eg, 8 or 12) LPC filter coefficients L _i that are transmitted to indicate the input conversation for a continuous frame of n (eg, 160) samples. The speech signal s also has a correlation unit (corre).
later unit) 4 (usually this is the LPC coder 3
Is part of. Because here, the separating correlator [se
, but the autocorrelation vector R _i of the speech is usually LP.
(Since it is generated as one step of C analysis). The correlator 4 generates an autocorrelation vector R _i ,
The vector R _i comprises a zero-order correlation coefficient R ₀ and at least two further autocorrelation coefficients R1, R2, R3. These are multiplier units.
it) 5.

The second input 11 is connected to a second microphone located away from the speaker and receives only background noise. The input from this microphone is converted into a digital input sample string by the AD converter 12,
LPC analysis is performed by the LPC analyzer 13. The "noise" LPC coefficients generated by the analyzer 13 pass through the correlation unit 14 and the autocorrelation vector generated thereby is term-wise multiplied by the autocorrelation coefficient R _i of the input signal from the speech microphone of the multiplier 5. , The weighting factors produced thereby are added by the adder 6 according to Equation 1, thereby providing a filter having the anti-phase shape of the noise spectrum from a noise-only microphone (actually in a signal / pulse / noise microphone). It has the same shape as the noise spectrum) and thus filters out most noise. The resulting measurement value M is compared with a threshold value by a thresholder 7 to produce a logic output 8 indicating whether speech is present. If M is large here, it is considered that a conversation exists.

This embodiment uses two microphones and two LPC analyzers, which adds cost and complexity, but these can be increased if desired.

On the other hand, another embodiment uses the autocorrelation from the noise microphone 11 and the corresponding value formed using the LPC coefficients from the main microphone 1. In that case, another autocorrelator is needed instead of the LPC analyzer.

Thus, these embodiments are capable of operating in different situations with different frequencies of noise, or in a given situation, in the presence of varying noise spectra.

In the preferred embodiment of FIG. 2, a buffer 15 is provided which stores a set (or set of autocorrelation vectors) of LPC coefficients, the values of which are "non-speech". ) (I.e., noise only) ". These values are used to obtain the value according to Equation 1, which of course corresponds to the distortion measurement method of Itakura and Saito, but not the current frame of LPC coefficients, but an approximation of the anti-phase noise spectrum. The difference is that a single frame of LPC coefficients that matches the values is used.

The LPC coefficient vector L _i output by the analyzer 3 is also directed to the correlator 14 and thereby produces the autocorrelation vector of the LPC coefficient vector.
Buffer memory 15 is controlled by the speech / non-speech output of thresholder 7 to provide a "speech" frame.
During the frame, the buffer holds the "noise" autocorrelation coefficient, but during the "noise" frame, a new set of LPC coefficients is used to update the buffer, for example by the composite switch 16. The output of the correlator 14 carrying each autocorrelation coefficient is connected to the buffer 15 via this switch 16. The correlator 14 may be placed after the buffer 15. Furthermore, the determination of the speech / non-speech for updating the coefficients need not be from output 8,
It can be obtained (preferably) in other ways.

Since a period without conversation often occurs,
The LPC coefficients stored in the buffer are occasionally updated, which allows the device to follow changes in the noise spectrum. If the noise spectrum is relatively stable in time (as is often the case), then updating such a buffer may only be necessary infrequently or only for the initial operation of the detector. Though conceivable, it is often desirable to update in situations like mobile (car) radio.

As a modification of this embodiment, the system applies Equation 1 with coefficient terms that match a simple fixed highpass filter, and then switches by using the "noise period" LPC coefficients. Start fitting. If speech detection fails for some reason, the system can re-use a simple high pass filter.

The expression that can be normalized by dividing the above value by R ₀ and compared to the threshold is:

[0038]

[Equation 6]

This value is independent of the frame's total signal power and is therefore compensated for total signal level changes, but does not give a significant contrast between "noise" and "speech" levels, and thus It is not suitable for use in a noisy environment.

When the noise spectrum changes gradually (as described below) (obtained from a noise microphone or noise only period in the various embodiments described above).
Instead of using LPC analysis to obtain the inverse filter coefficients of the noise signal, a general adaptive filter (adapt) is used.
The ive filter) can be used to generate a prototype of the anti-phase noise spectrum, and a relatively slow precision common to such filters can be obtained. In the embodiment corresponding to FIG. 1, the LPC analysis unit 13 is easily adapted to the adaptive filter (eg transversal (t
FIR) or lattice filter
(Lattice filter), whose filter is connected to the system to convert the noise input into white noise by generating an inverse filter prototype, the coefficients of which are autocorrelated as described above. Is supplied to the container 14.

In the second embodiment shown in FIG. 2, L
The PC analysis means 3 is replaced by such a fitness filter and the buffer means 15 is omitted. However, switch 16 operates to prevent the adaptive filter from matching its coefficients during the conversation period.

A second speech motion characteristic detector used in another embodiment of the invention will now be described.

In the following description, it is clear that the LPC coefficient vector is simply the pulse response of the FIR filter, which is the anti-phase spectral shape of the input signal. When the distortion value by Itakura and Saito is formed between adjacent frames, it is actually equal to the power of the signal because it has been filtered by the LPC filter of the previous frame. Thus, if there is little difference in the spectra of adjacent frames, the corresponding little spectral power of the frames will escape filtering and its value will be small.
At the same time, the large spectral difference between the frames produces large Itakura-Saito distortion values, which reflect the spectral similarity of adjacent frames. Speed
With respect to the chillers, it is desirable to maximize the frame length by minimizing the data rate.
That is, if the frame length is long enough, the speech signal exhibits significant spectral changes from frame to frame (otherwise coded is redundant). On the other hand, noise has a spectral shape that gradually changes from frame to frame, and during periods when speech is not present in the signal, the anti-phase LPC filter is applied from the previous frame, and most noise power is removed. "Filter out"
Therefore, the distortion values due to Itakura and Saito are correspondingly small.

The Itakura-Saito distortion value between adjacent frames of a noisy signal including an intermittent conversation is generally larger in the conversation period than in the noise period, and the degree of change (indicated by the standard deviation). Is large, and there are few intermittent changes.

Here, the standard deviation (standard) of M
deviation) is also a reliable value, and the effect of taking each standard deviation is essentially to smooth the value.

In this second form of speech motion characteristic detector, the measured parameter used to determine whether speech is present is preferably the standard deviation of the Itakura-Saito distortion value. However, other methods of measuring change and other methods of measuring spectral distortion (eg, based on FFT analysis) can be applied.

There is also an advantage in using adaptive thresholds for voice performance characteristic detection. Such thresholds should not be adjusted for the duration of the speech, and if so, the speech signal will be threshold out.
t) is performed. Therefore, it is necessary to control the threshold adapter using the speech / non-speech control signal,
This control signal is preferably independent of the output of the threshold adapter. The threshold T is adjusted so that it is kept above the level of the value M when only noise is present. Since the value generally changes randomly in the presence of noise, the threshold is changed by determining the average level for many blocks and setting the threshold to a level proportional to this average level. However, this is generally not sufficient in noisy situations, and parameters for some blocks
The assessment regarding the degree of change of the data is considered.

Therefore, the threshold value T is calculated according to the following equation.

[0049]

[Equation 7]

Here, M is the average value of the measured values for many consecutive frames, d is the standard deviation of the measured values for those frames, and K is a constant (typically). Is 2.)

In practice, the adaptation action should not be started again immediately after the absence of speech is indicated, and the descent is stable (to avoid repeated rapid switching between conforming and nonconforming states). You should wait until you know what you have done.

FIG. 3 shows a preferred embodiment of the present invention having the above-mentioned matters.
This is a preferred embodiment, and input 1 is an analog / digital converter.
Sampled by an ADC 2 and digitized
Received signal and send it to the input of the anti-phase filter analyzer 3.
Signal, and the anti-phase filter analyzer 3 actually
Operating characteristic part of the speech coder where the detector operates
And a filter that matches the opposite phase of the input signal spectrum
Coefficient of_i(Typically 8). Digital signal
Also supplied to the autocorrelator 4 (which is part of the analyzer 3)
And the autocorrelator 4 receives the input signal (or at least those
Are as many low order terms as LPC coefficients).
Cutle R_iTo occur. The operation of these parts of the device is
It is shown in FIGS. 1 and 2. Autocorrelation coefficient R_iIs suitable
And several continuous flavor frames (typically 5 to
20 ms) averaged and their confidence
Be improved. This averaging is based on the autocorrelation in the buffer 4a.
Storing each set of autocorrelation coefficients output by the device 4,
Using the averager 4b, the current self
Correlation coefficient R_i, And the buffer 4 stored in the buffer 4a
Weighting of coefficients from the previous frame supplied by a
This is accomplished by producing a summed value that is calculated. It
Averaged autocorrelation coefficient Ra obtained by _iIs heavy
The means for adding and adding 5, 6 are also provided, which means also
The data stored in the autocorrelator 14 via the buffer 15
Anti-phase filter coefficient L during noise period_iAutocorrelation vector A of
_iAnd receive Ra_iAnd A_iIs defined by
Form the value M.

[0053]

[Equation 8]

This value is compared with the threshold value by the thresholder 7 and a logical result is produced at the output 8 indicating whether or not a conversation is present.

In order for the anti-phase filter coefficients L _i to match a good approximation of the anti-phase of the noise spectrum, it is desirable to update these coefficients during the noise period (of course not during the conversation). However, the speed based on that update
Each non-speech decision is unaffected by the result of the update, or due to a single frame of falsely identified signal, the voice performance detector will eventually "out of lock". ", And the next frame is erroneously recognized. Therefore, a control signal generating circuit 20, that is to say an auxiliary operating characteristic detector for the separated voice, is provided, which forms an independent control signal which indicates whether speech is present, the anti-phase filter analyzer 3 (or the buffer 8). ), And therefore the anti-phase filter autocorrelation coefficient A _i used to form the value M is updated only during the "noise" period. The control signal generation circuit 20 is an LPC analyzer 21.
(Which is again part of the speech coder and is especially implemented by the analyzer 3), which matches the input signal and the autocorrelator 21a (which can be implemented by the autocorrelator 3a). A set of LPC coefficients M
_i , and the autocorrelator 21a outputs the autocorrelation coefficient B of M _i.
get _i . If the analyzer 21 is implemented by the analyzer 3, then M _i = L _i and B _i = A _i . These autocorrelation coefficients are weighted and added by means 22 and 23.
(Equal to 5 and 6), and this means is also autocorrelator 4
Receives an autocorrelation vector R _i of the input signal from Therefore, the spectral similarity between the input and the previous speech frames is calculated. As described above, this is R _i of the current frame and B of the previous frame.
The Itakura-Saito distortion value between _i and the Itakura-Saito distortion value related to Ri and B _i of the current frame, or the corresponding value stored in the buffer 24. Resulting in a spectrally distinct signal (in each case
Its value is preferably energy-normalized by dividing by Ro). Of course, the buffer 24 is updated here. This spectrally distinct signal, when compared to the threshold by thresholder 26, indicates whether speech is present, as described above. Although this method is excellent for distinguishing noise from speech that is not voice (a task (ta
sk)), it has been discovered that the ability to distinguish noise from spoken speech is generally low. Therefore, the circuit 20 includes a pitch analyzer.
r) 27 (actually capable of operating as part of a speech coder, in particular capable of measuring the long delay value of a calculator produced in a multi-pulse LPC coder) It would be desirable to provide a voice speech detection circuit. Pitch analyzer 27 generates a logic signal that is "true" when a spoken conversation is detected, which signal is threshold 2.
Combined with the threshold value obtained from 6 (generally "truth" when there is a non-speech conversation),
It is applied to the input of NOR gate 28 and produces a signal that is "false" when speech is present and "true" when noise is present. This signal is supplied to the buffer 8 (or the anti-phase filter analyzer 3), whereby the anti-phase filter coefficient Li is updated only during the noise period.

The threshold adapter 29 is also connected,
The non-speech signal control output of the control signal generation circuit 20 is received. The output of the threshold adapter 29 is supplied to the thresholder 7. The output of the threshold adapter 29 is supplied to the thresholder 7. The threshold adapter increments the threshold in steps proportional to the instantaneous threshold level until the threshold approaches the noise power level (which is easily obtained, for example, by adding and weighting circuits 22, 23). Decrement
ment). When the input signal is very small, the threshold is preferably set to low level automatically. This is because at low signal levels, the amount of signal produced by ADC2 cannot produce reliable results.

Furthermore, "hangover (hangover
r) "generating means 30 is provided, which is the threshold 7
After that, the duration of the conversation is measured and the output is short when the presence of the conversation is indicated for a period exceeding a predetermined time constant. "
It remains high during the "hangover." In this way, the clip in the middle of the low-level speech burst.
ping) is avoided, and by selecting an appropriate time constant,
A short spike noise that is erroneously displayed during a conversation can prevent the hangover generator 30 from being activated. Of course, all the functions described above may be implemented as part of a properly programmed single digital processing means, such as an LPC codec (which is the desired configuration), or an associated memory device. It can be implemented by means such as a suitably programmed micro-computer or a digital signal processing chip (DSP) configured as a micro-controller chip.

As mentioned above, the voice detector can be easily constructed as part of the LPC codec.
On the other hand, if the autocorrelation coefficient of the signal, or its related value (partial correlation or "parcor" coefficient) is transmitted to the distant station, the speech detection will be done away from the codec. Be seen.

[Brief description of drawings]

FIG. 1 is a block diagram showing a first embodiment of the present invention.

FIG. 2 is a block diagram showing a second embodiment of the present invention.

FIG. 3 shows a preferred third embodiment of the present invention.

[Explanation of symbols]

1 ... Input 2… AD converter 3… Analyzer 4… AFC 5 ... Multiplier 6 ... Adder 7 ... Threshold 8… Output 11… Noise microphone 12… AD converter 13… Analyzer 14… AFC 15… Buffer memory 16… Switch

Continued front page    (72) Inventor Daniel Kenneth Freeman               UK IP 4,4 Etch               Tay, Suffolk, Ipswich,               Huinchray Road 42 (72) Inventor Even Boyd               British country IP 9,2 Etsuk               Su Yi, Suffolk, Ips Wits               Chi, Capel S. T. Marie, Ho               Moufield 5                (56) References JP-A-63-98700 (JP, A)                 Japanese Patent Publication Sho 62-50837 (JP, B2)

Claims (7)

(57) [Claims] 1. (i) Spectral similarity between an input signal component and a component of the input signal that is determined to be non-speech to produce an output signal that indicates the presence or absence of speech in the input signal. A first voice motion characteristic detector (3-6, 14) that operates to form a sex value; (ii) a memory (15) for storing data derived from the non-speech portion; (iii) ) An auxiliary voice operation characteristic detector (20), the auxiliary voice operation characteristic detector (20) independently controls the update of the memory (15) to detect the auxiliary operation characteristic detector. The container (20) is used to determine whether there is a conversation in the input signal.
Generate an independent control signal from the output signal indicating
In order to do so, a speech motion characteristic detection device is operative to form a value of the spectral similarity between the latest component of the input signal and the earlier component of the input signal. 2. (i) means for receiving an input signal (1); (ii) a memory (15) for storing a noise expression signal representing an estimated noise portion of the input signal; and (iii) the above. Means (3-6, periodically forming from the input signal and the noise representation signal a value of the spectral similarity between the portion of the input signal and the estimated noise portion.
14), (iv) means (7) for comparing said value with a threshold value to produce an output signal indicating whether speech is present or absent, and (v) an auxiliary voice behavior characteristic detector (20). ), And (vi) a memory updating means for updating the memory with the input signal, wherein the auxiliary voice behavior characteristic detector is configured to detect the presence or absence of speech from the output signal. To generate an independent control signal, which operates in response to the value of the spectral similarity between the latest part of the input signal and the previous part of the input signal, and said control signal being present in the input signal You
Indicates no Luke or present, and the memory update unit voice activity characteristic detection, characterized in that the operation to update the memory by the input signal only when indicating that the control signal is not present conversations apparatus. 3. The apparatus of claim 2, further comprising means for adjusting the threshold value during a period when the control signal indicates that no conversation is present. 4. The auxiliary voice activity characteristic detector further comprises means (27) for detecting spoken speech including pitch analyzer means for generating a signal indicating the presence of the spoken speech. Device according to claim 2 or 3, characterized in that the control signal produced by the auxiliary voice behavior characteristic detector (20) is further followed. 5. A device for encoding a speech signal, comprising a device according to any one of claims 1 to 4. 6. An automobile telephone device including the device according to claim 1. 7. A method for detecting a conversational operation characteristic of an input signal, comprising: receiving the input signal, estimating a noise signal portion of the input signal, storing data representing the noise signal portion, Forming a value M of the spectral similarity between a portion of the signal and the noise signal portion and providing a first voice behavior characteristic indication indicating whether speech is present or absent depending on whether it is above a threshold value or not. To produce, a parameter derived from the value M is compared to a first threshold value T, wherein the estimating step comprises producing an ancillary voice behavior characteristic indication, and To form a spectral distortion value of the similarity between the most recent component of the input signal and the earlier component of the input signal, and whether or not there is a conversation depending on whether or not that value is exceeded. The spectral distortion value is compared to a second threshold value to create the ancillary audio behavioral characteristic indication indicating absent, and the ancillary audio instruction is compared to the first threshold value .
How to the independent of the speech operation characteristics indicated, comprising the step of said auxiliary voice activity characteristic instruction to update the data said stored by an input signal only during a period for indicating that there is no conversation .