FI110726B - Detection of voice activity - Google Patents

Abstract

The first aspect provides a voice activity detection appts. for receiving an input signal, estimating the noise signal component of the input signal and continually forming a measure M of the spectral similarity between a portion of the input signal and the noise signal. A circuit is provided to compare a parameter derived from the measure M with a threshold value T to produce an output to indicate the presence, or absence, of speech depending on whether, or not, that value is exceeded. A second aspect covers voice activity detection appts. which continually forms a spectral distortion measure and carries out a comparison.

Description

1 110726 Sound Activity Detection - Detection Av ustactivities The Sound Activity Detector is a device that is supplied with a signal for detecting speech or noise-only episodes. While not limited to the present invention, one particularly interesting application of such detectors is mobile radiotelephone systems where the speech encoder can use speech presence or absence information to improve radio spectrum utilization and where the noise level (from the unit mounted on the vehicle) is likely to be high.

10 The essential content of the expression of voice activity is to find a measure that clearly differs between speech periods and speechless periods. In a device containing a speech encoder, various parameters can easily be obtained from different degrees of the encoder, and it is therefore desirable to reduce the processing required by using one of these parameters. In many environments, the main noise sources occur in certain known areas of the frequency spectrum. For example, in a moving car, much of the noise (e.g. engine noise) is concentrated in the low-frequency areas of the spectrum. When such information about the noise spectral position is available, it is advantageous to base the decision on the presence or absence of speech on the measurements made in that part of the spectrum that contains the relative ··; its a bit of noise. In practice, it would of course be possible to filter · ;;; the signal beforehand prior to analysis for detecting speech activity, but when the voice activity detector follows the output of the speech encoder, pre-filtering would distort the audio signal to be encoded.

25

The invention thus relates to a sound activity detector device according to the preamble of claim 1, characterized in that the device includes analysis means for producing coefficients of a filter which has an inverse spectral response to the frequency of one of said two signals; and the means for forming the dimension are capable of forming a dimension M 1; ., Which is proportional to the zero-order autocorrelation (R'0) of a signal,> 2,110,726, which signal is obtained by filtering the remaining signal of said two signals as a filter having said coefficients.

Preferably, the measure is the Itakura-Saito distortion measure.

5

The invention also relates to a method for detecting voice activity, a device for coding speech signals and a mobile phone device. Characteristic features of these inventions are apparent from the independent claims 15, 16 and 17.

10

Other embodiments of the present invention are as defined in the claims.

Some embodiments of the invention will now be described, by way of example, with reference to the accompanying drawings, in which:

Figure 1 is a block diagram of a first embodiment of the invention.

Fig. 2 shows another embodiment of the invention.

: 20 '·; Figure 3 illustrates a third preferred embodiment of the invention.

! "The general principle underlying the first tone activity detector according to the first embodiment of the invention is as follows.

25

A frame having n signal samples (i.e., S1, S2, s3, S4 ... Sn-1) is obtained when it is passed through a fourth order finite impulse response (FIR) digital computation filter having an impulse response of (1, ho, hi, h2, h3), 'the resulting filtered signal (when samples from previous frames are ignored) s' = (so),; (si + hoSo),> 3 110726 (s2 + hosi + Hiso), (S3 + hoS2 + h1Sl + h2So) (S4 + hoS3 + h1S2 + h2Sl + Hiso), (S5 + hoS4 + h1S3 + h2S2 + h3Sl), (S 6 + hoS5 + hlS4 + h2S3 + h3S2) / (S7. ··)

The zero autocorrelation coefficient is a squared sum of terms that can be normalized, i.e. divided by an integer number of terms (for frames of constant length, the division is easiest to omit). The coefficient of the filtered signal is thus n-1 R'o = (s'i) 2 i = 0 and thus forms a measure of the power of a portion of the signal s that falls within the calculated bandwidth of the calculated filtered signal, that is, within the passband of the calculator.

When solving an expression, the first 4 terms are ignored '1. s'o '(S4 * »0 * 3 *» 1 * 2 * V: * Vo) 2 + (* 5 * V »+" Λ * V: * »3si) 2 + ...

* * 4 * »0 * 4 * 3 +» 1 * 4 * 2 * V4 * l * »3 * 4 * 0 *» 0 * 4 * 3 * 4 * 0 * »0» 1 * 3 * 2 * » 0 »2 * 3 * 1 * ty> 3 * 3 * 0 *» 1 * 4 * 2 * »0» iS3 * 2 + »1 * 2 *» 1 »2 * 2S1 +» AVa * ”2 * 4 * 1 * Vl * j * l * »lV2 * l *» 2 * 1 *!> 2 · »3 * 1 * 0 4 110726 + il3Vo + V3V0 + * 1l'13S2S0 + V3S Λ + h3S0 + · '« 3 S0 (1 + h0 + "r h2 'h3} + sx (2» 0 + 2Vl + 2tllh2 + ^ 31 + R2 (2hj_ * »J & 3 + 21 ^) + s3 (2¾ * a0n3) r (21l3) R'o can thus be obtain the combination of autocorrelation coefficients Ri weighted by the constants in parentheses that determine the frequency band in which the value of the coefficient R'o is affected. The terms in parentheses are in fact the autocorrelation coefficients of the impulse response of the computational filter.

B

; 'I! * · Ο * * oV * Σ! * Α .............. 111 ··· i »1 ♦« 4 «where N is the order of the filter and Hi is the impulse of the filter, ·· . response (non-normalized) autocorrelation coefficients.

In other words, the effect of signal filtering on the signal autocorrelation coefficients can be simulated by generating a weighted sum of the (unfiltered) signal autocorrelation coefficients using the impulse response that the required filter would have had.

Thus, a relatively simple algorithm with little multiplication can simulate the effect of a digital filter, which typically requires 100 times this number of multiplications.

Alternatively, the filtering operation can be viewed as a spectral comparison in its form comparing the signal spectrum with the reference spectrum (inverse of the response of the computation filter). Because the computational filter in this embodiment is chosen to approximate the inverse of the noise spectrum, this operation can be considered as a spectral comparison of speech and noise spectra and the resulting zero autocorrelation coefficient (i.e., inverse-filtered signal energy) can be considered as a measure of spectral difference. The Itakura-Saito measure is used in linear prediction coding to evaluate LPC predictor filter compatibility with the result spectrum and can be expressed in one form

OF

M = Vo + ^ 8iAi 'ui where Ao, etc. are the autocorrelation coefficients of the LPC parameter set. It is found that the expression is very similar to the dependency derived above, and remembering that the LPC coefficients are the pins of an FIR filter having a ··· inverse spectral response of the input signal such that the LPC coefficient 1> · ·. "the set is the impulse response of the LPC inverse filter, it is obvious /: ·. that the Itakura-Saito distortion measure is in fact only a form of equation 1, where the filter response H is the inverse of the spectral form of the input-only model.

In fact, it is also possible to convert the spectra using LPC coefficients of the test spectrum and autocorrelation coefficients of the reference spectrum to obtain a different measure of the same spectra; particular aspect.

6 110726

Vector Quantisation, "IEEE Trans on ASSP, Vol. ASSP-28, No. 5, October 1980.

Since the signal frames have only a finite length and a certain number of terms (N, where N is the order of the filter) are ignored, the result above is only an approximation. However, it provides an astonishingly good indication of the presence or absence of speech and can thus be used as a measure of M speech expression. In an environment where the noise spectrum is well known and constant, it is perfectly possible to simply use fixed coefficients ho, hi, etc. to model the inverse noise filter.

However, a device that can adapt to different noise environments can be used more generally.

As shown in Figure 1, in the first embodiment, the signal from the microphone (not shown) is received at input 1 and converted to digital samples s at a suitable sampling rate by an analog-to-digital converter 2. The LPC · · · analysis unit 3 (included in LPC encoder of known type) 160) for sequential frames of the sample, a set of N (e.g., 8 or 12) LPC filter coefficients L1, which are transmitted to represent incoming speech. The speech signal s is input ««! also for the correlator unit 4 (normally included as part of the LPC-1 in the encoder 3, since the speech autocorrelation vector R1 is also developed as a step in the LPC analysis, although it is clear that a separate correlator could also be used). The correlator 4 generates an autocorrelation vector R1 which includes a zero order correlation coefficient R0 and at least two other autocorrelation coefficients R1, R2, R3. After this, they are fed to the multiplier unit 5.

>

The second input 11 is connected to a second microphone away from the speaker so that this microphone receives only background noise. The input from this microphone is converted by the AD converter 12 to a digital input sample queue and is LPC analyzed by another LPC analyzer 13. The "noise ,," generated from the analyzer 13 is fed to the correlator unit 14 and the resulting autocorrelator vector is multiplied by the term the weighted coefficients developed in the multiplier 5 and so combined in the adder 6 according to equation 1 to obtain a filter effect which is inverse to the noise spectrum of the noise-detecting microphone (which is virtually the same as the noise spectrum in the signal and noise receiving microphone) . The resulting dimension M is compared to a threshold value in a threshold circuit 7 to generate a logic output 8 which indicates the presence or absence of speech. If M is large, speech is considered to occur.

However, this embodiment requires two microphones and two LPC analyzers, which increases the cost and complexity of the equipment required.

Alternatively, in the second embodiment, a corresponding dimension formed using the autocorrelations from the noise microphone 11 and the LPC coefficients obtained from the main microphone 1 is used, so that an additional autocorrelator is required instead of an additional LPC analyzer.

These embodiments may thus operate in different environments where noise occurs at different frequencies or when the noise spectrum changes in a given environment.

As shown in Figure 2, in a preferred embodiment of the invention, there is a buffer 15 storing a set of LPC coefficients (or set of autocorrelation vectors) derived from microphone input 1 during a period recognized as a "speechless" period (i.e., merely). kohinajaksoksi). These coefficients are then used to derive the measure using equation 1, which of course also corresponds to the Itakura-Saito distortion measure, except that one stored LPC coefficient frame corresponding to the inverse noise spectrum approximation is used instead of the current LPC coefficient frame.

The LPC coefficient vector provided by the analyzer 3 is also provided to the correlator 14 which forms the autocorrelation vector of the LPC coefficient vector. The speech / speech output of the threshold circuit 7 controls the buffer memory 15 in such a way that the buffer retains "noise" autocorrelation actions during "speech frames", but a new set of LPC coefficients can be used during "noise frames" to update the buffer, e.g. the outputs, each of which has an autocorrelation coefficient, are coupled to the buffer 15. It is clear that the correlator 14 could be located after the buffer 15. Further, a speech / speechless decision to update the coefficients does not need to be made at output 8, but could be derived (and preferably derived) in some other way.

Because of the frequent occurrence of speechless periods, the _ LPC coefficients stored in the buffer are updated from time to time so that the device is able to monitor changes in the noise spectrum. It is clear that such a buffer update may be necessary only occasionally, or may occur only once at the start of the detector operation if (as is often the case) the noise spectrum is relatively unchanged over time, but in the mobile radio communication environment.

In a variant of this embodiment, the system initially uses equation 1 as the coefficients correspond to a simple fixed high pass filter, and then the system begins to adapt by switching to the "noise period" LPC-9 110726 coefficients. If speech detection fails for any reason, the system may revert to using a simple high-pass filter.

The above dimension can be normalized by dividing by Ro so that the expression compared to the threshold is of the form N RjA, M = A0 + 2Σ-1 = 1 Rq This dimension is independent of the total signal energy of the frame and thus compensated for, but gives a weaker signal. contrast between "noise" and "speech levels" and is therefore most advantageously not used in highly disturbed environments.

Instead of using LPC analysis to derive the inverse filter coefficients of the noise signal (either from noise microphone or from noise-only sequences, as in the various examples described above), the inverse noise spectrum •. it is possible to model using a known type of adaptive filter. Because the noise spectrum changes only slowly (as will be explained below), the usual relatively slow coefficient adaptation rate for such filters can be accepted. In one embodiment corresponding to Figure 1, the LPC analysis unit 13 is simply replaced by an adaptive filter (e.g., an FIR transverse filter or a network filter) coupled to render incoming noise whiter by modeling the inverse filter, and its coefficients are input as above. autocorrelation function. to the stand 14.

In another embodiment corresponding to the embodiment of FIG. 2 10 110726, the LPC analysis means 3 is replaced by such an adaptive filter, and the buffer means 15 is omitted, but the switch 16 operates to prevent the adaptive filter from adapting its coefficients during speech periods.

Hereinafter, another voice activity detector for use with another embodiment of the invention will be described.

It is clear from the above that the LPC coefficient vector is simply the impulse response of an FIR filter whose response approximates the inverse spectral form of the input signal. When generating an Itakura-Saito distortion measure between adjacent frames, this is actually equal to the signal power filtered by the LPC filter of the previous frame. Thus, if the spectra of adjacent frames differ slightly, the corresponding small amount of spectral power of the frame is not filtered and the dimension is small. Similarly, the large difference between frames produces a large Itakura-Saito distortion measure so that the measure reflects the spectral similarity of the adjacent frames. In a speech encoder, it is desirable to minimize the data frequency so that the frame length is made as large as possible. In other words, if the length of the frame is large enough, there should be a significant spectral shift between frames in the speech signal: * (if not, this is an excess coding). On the other hand, noise has a spectral form that varies slowly from frame to frame, and thus, during a period of no speech, the Itakura-Saito distortion measure is correspondingly small - since using the reverse frame reverse LPC filter "filters out" most of the noise power.

The Itakura-Saito distortion measure between adjacent frames of an intermittent speech-containing noise signal is typically greater during speech periods than during noise periods. The degree of variation (as described by the standard deviation) is also greater than 11,110,726 and less variable at times.

Note that the standard deviation of the standard deviation of dimension M is also a reliable measure. The effect of generating each standard deviation is in fact smoothing the measure.

The parameter measured here in another form of the voice activity detector used to determine whether speech is present is preferably the standard deviation of the Itakura-Saito distortion measure, but other measures of variation and other spectral distortion measurements (based, for example, on FFT analysis) could be used.

The use of an adaptive threshold for detecting voice activity has been found to be advantageous. Such thresholds must not be set during speech periods or else the speech signal will be cut. Therefore, the threshold adaptation circuit must be controlled using a speech / speech control signal, and this control signal should preferably be independent of the output of the threshold adaptation circuit.

The threshold T is adaptively set such that the threshold, just above the level of dimension M, is kept by noise only; occurs. Because the dimension generally varies randomly: when noise occurs, the threshold is changed by determining the average level over several blocks, and the threshold is set at a level proportional to this average. However, this is usually not sufficient in a noisy environment and thus the determination of the degree of parameter variation over several blocks is also considered.

The threshold value T is thus preferably calculated according to the following expression [· T = M ′ + K.d 12 110726 where M ′ is the average of the measure over several successive frames, d is the standard deviation of the measure during these frames and K is a constant (typically 2).

In practice, it is preferable that the adaptation is not restarted immediately after the speech is detected, but is expected to ensure that the drop is stable (to avoid a rapid repetitive switching between the adaptive and non-adaptive modes).

As shown in Fig. 3, in a preferred embodiment of the invention having the above features, the input 1 receives a sampled signal converted to digital by an analog-to-digital converter (ADC) 2 and is input to the input of the inverse filter analyzer 3, with which the sound activity detector is intended to operate, and which generates the coefficients Li (typically 8) of the filter corresponding to the inverse of the input signal spectrum. The digitized signal is also fed to an autocorrelator 4 (included as part of the analyzer 3), which generates an autocorrelation vector Ri of the input signal (or at least as many orders of magnitude lower than the LPC coefficients). The operation of these parts of the device is as described in Figures 1 and 2. · Average values of the autocorrelation coefficients R1 over a plurality of consecutive speech frames (typically 5 to 20 msec) * to improve their reliability are generated. This can be accomplished by storing each set of autocorrelation coefficients provided by the autocorrelator 4 in buffer 4a and using the averaging factor 4b to form a weighted sum of the current autocorrelation coefficients R 1 and the coefficients of previous frames stored in and input to the buffer 4a. The thus derived average autocorrelation coefficients Rai are fed by weighting and summing means 5, 6, which also receive the stored noise period inverse filter filter coefficients and 110726

Li from the autocorrelator vector Ai from the autocorrelator 14 via buffer 15 and forming a dimension M of Rai and Ai, preferably defined as: S * & 0 ΜΑ ·

Ro This measure is then compared to the threshold level in the threshold circuit 7 and the logical result gives an indication of the presence or absence of speech at the output 8.

In order for the inverse filter coefficients L i to correspond to a reasonable estimate of the inverse of the noise spectrum, it is desirable to update these coefficients during the noise periods (and of course not to update during the speech periods). However, it is preferred that the speech / speechless decision upon which the update is based does not depend on the result of the update or otherwise a single misidentified signal frame may cause the "activity lock" indicator of the audio activity thereafter and the following frames ·. false identification. Therefore, it is preferable to use a control signal generating circuit 20 which is in fact a separate voice activity detector which provides an independent '; a control signal indicating the presence of speech or; absence, to control the inverse filter analyzer 3 (i.e., buffer 8) so that the inverse filter autocorrelation coefficients Ai used to form the dimension M are updated only during "noise periods". The control signal generating circuit 20 includes an LPC analyzer 21 (which may also be part of a speech coder and may be implemented in particular by analyzer 3) generating a set of LPC coefficients Hi corresponding to the input signal, and an autocorrelator 21a (implemented by auto correlator 3a) autocorrelation coefficients Bi. If the analyzer 21 is implemented by the analyzer 3, then Mi = Li and Bi = Ai. These autocorrelation coefficients 110726 14 are then applied to the weighting and summing means 22, 23 (corresponding to the elements 5, 6), which also receive the result signal autocorrelator vector R1 from the autocorrelator 4. Thus, a measure of spectral similarity between the incoming speech frame I and the previous speech frame. This measure may be the Itakura-Saito distortion measure between the current frame coefficients Ri and the previous frame coefficients Bi, as described above, or it may instead be derived by calculating the Itakura-Saito distortion measure for the current frame coefficients Ri and Bi and subtracting (in the subtractor 25 a corresponding previous measure stored in buffer 24 to generate a spectrum difference signal (in each case the energy of the measure is normalized by dividing by Ro). The buffer 24 is then naturally updated. This spectral difference signal, after the threshold comparison performed in the threshold circuit 26 described above, provides a detector for the presence or absence of speech. However, we have found that, although this measure is excellent for distinguishing noise from silent speech (a task known systems do not usually), it is generally somewhat less capable of distinguishing noise from verbal speech. In accordance with this * '1 · I ·, the circuit 20 preferably further utilizes an audio speech detector circuit having a pitch analyzer 27 (which may in practice function as part of the speech encoder and' specifically measure the long-term predictor delay value generated in the multipulse LPC encoder ). The pitch analyzer 27 generates a logical signal that is "true" when the voiced speech is detected, and this signal is coupled with a threshold value derived from the threshold circuit 26 (which is usually "true" in the presence of silent speech) to the inputs of the OR port 28. , which is "false" when speech occurs and "true" noise occurs.

This signal is supplied to buffer 8 (or inverse filter analyzer 3) so that the inverse filter coefficients Li> 'are updated only during noise periods.

»'» 15 110726

The threshold adapting circuit 29 is also coupled to receive a speech output control of a control signal generator circuit 20. The output of the threshold adaptation circuit 29 is supplied by a threshold circuit 11. The threshold adaptation circuit increases or decreases the threshold in steps proportional to the current threshold value until the threshold approximates the noise power level (which can be practically derived from the weighting and summing circuit 23).

When the input signal is very low, it may be advantageous! that the threshold is automatically set to a fixed low level, since the signal quantization effect generated by the analog-to-digital converter 2 may produce unreliable results at low signal levels.

Additionally, "overrun" generating means 30 can be used to measure the duration of speech detection after threshold circuit 7, and when the occurrence of speech is detected over a predetermined time constant, the output is maintained in a higher state for a short "overrun" period. This avoids': clipping of low-level speech bursts in the middle, and '' right selection of the time constant prevents triggering of the crossing generator 30 by the presence of short, incorrectly pronounced noise peaks.

♦ t • It is, of course, clear that all of the above functions can be performed by one suitably programmed digital processing device, such as a digital signal processing circuit (DSP), which is part of the implemented LPC codec (this is preferred), or microcontroller circuit with associated memory devices.

·. · As explained above, the audio detector can be practically implemented as part of an LPC codec. Alternatively, when the autocorrelation coefficients of a signal, or related dimensions 110726 16 (partial correlation, or "parcor" coefficients), are transmitted to a remote station, voice detection can occur far from the codec.

Claims (17)

A voice activity detecting device, comprising: (i) means (1) for receiving a first input signal; (Ii) means (14,15) for periodically adaptively generating a second signal representing an estimate of the noise signal component of the first signal; (iii) means (4, 5, 6) for periodically forming, from the first and second signals, a spectral similarity of a portion of the input signal M and said estimated cavity signal component; and (iv) means (7) for comparing the dimension M with a threshold T for generating an output to detect the presence or absence of speech; characterized in that (v) the apparatus includes analysis means (13, 3) for generating coefficients of a filter 15 having a spectral response inverse of one of said two signals to the frequency spectrum; and (vi) the means (4, 5, 6) for generating the dimension are capable of generating a dimension M relative to the zero-order autocorrelation (R'o) of the signal obtained by filtering the remaining signals from said two 20 signals. as a filter having said coefficients; activities. Device according to Claim 1, characterized in that the analyzing means (13, 3) comprises an adaptive filter. 25 Device according to Claim 1, characterized in that the development means (14,15) are capable of calculating the impulse response of said coefficients: "'the autocorrelation coefficients A and the dimensionalizing means (4, 5, 6) comprise ..' means for calculating said residual auto-correlation coefficients for the signal .... 30 Ri, and means (5, 6) coupled to receive the values R j and A and calculate the dimension M of them. Apparatus according to claim 2, characterized in that the means (4) for calculating the autocorrelation coefficients R * of said remaining signal are arranged (4a, 4b) to do so in dependence on the autocorrelation coefficients of the plurality of consecutive portions of the signal. 5 Device according to claim 3 or 4, characterized in that M = RoAo + 2 Σ, RtAj 10 where A 1 represents the ith autocorrelation coefficient of the impulse response of said filter. Device according to claim 3 or 4, characterized in that Device according to one of claims 1 to 6, characterized in that '·. said one signal being the second noise representing signal and said remaining signal being the first input signal. 25 Device according to Claim 7, characterized in that in addition: '·. includes an input (11) configured to receive a second input signal, respectively, exposed to noise, where speech is missing, wherein the development means comprises LPC analysis means (13) for deriving Ai values from the second input signal. 110726 19 A device according to any one of claims 1 to 7, further comprising a buffer (15) coupled to store data from which autocorrelation coefficients A of said filter response can be obtained, wherein said filter response is periodically calculated from a signal by the LPC analysis tool. (3) wherein the device is so coupled and controlled that the dimension M is calculated using said stored data, and said stored data is updated only from periods in which speech is reported to be absent. The device according to claim 9, characterized in that it further comprises means (20) for detecting the absence of speech to control the updating of stored data, the means (20) for detecting the absence of speech being other means for detecting voice activity (20). Device according to one of the preceding claims, characterized in that it further comprises means (29) for adjusting the threshold value T during periods in which speech is detected missing. A device according to claim 11, characterized in that it further comprises: • a second voice activity detecting means (20) arranged to prevent the threshold being adjusted when speech occurs. The device of claim 10, further comprising: means (20) for adjusting said threshold value T during periods of speech absence, wherein said second voice activity detecting device (20) is arranged to prevent adjusting the threshold when there is speech. The device according to claim 11, 12 or 13, characterized in that; The threshold T, when adjusted, is adjusted to be equal to the mean of the measure, plus a term which is a fraction of the standard deviation of the measure. 20 110726 A method of detecting voice activity in a first input signal, comprising the steps of: (a) periodically adaptively generating a second signal representing an estimated noise signal component of the first signal; (B) periodically forming, from the first and second signals, a measure of the spectral similarity of a portion of the input signal and said estimated noise signal component; and (c) comparing the measure M with a threshold T to produce an output indicating the presence or absence of speech; Characterized by: (d) forming the coefficients of a filter having a spectral response which is the inverse of one of the two frequencies spectrum; and (e) the dimension M is relative to the zero-order autocorrelation R'o of the signal obtained by filtering the remaining filter of said two signals having said coefficients. 15 R / A, M = A0 + 2 £ ------- Ro where A, represents the ith autocorrelation-20 thi multiplier of the impulse response of said filter. Device for encoding speech signals, characterized in that it comprises a device according to any one of claims 1 to 14. '20 A mobile telephone device, characterized in that it comprises a device according to any one of claims 1 to 14. 21 110726