FI115328B - Expression for sound 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

115328 Sound Activity Detection. - Uttryck for a lot of activity.

The present application is subdivided from application F1904410.

5 A voice activity detector is a device that is fed with a signal for detecting speech sequences or noise-only sequences. While not limited to the present invention, one particularly interesting application of such detectors is mobile radiotelephone systems in which the speech encoder may use information on the presence or absence of speech to improve radio spectrum utilization and where the noise level (from the unit mounted on the vehicle) is likely to be high.

The essential content of the expression of voice activity is to find a measure that is clearly different between speech periods and non-speech periods. In a device containing a speech encoder, several parameters can easily be obtained from different degrees of the encoder, and it is therefore desirable to reduce the processing required by using any such parameter. 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 (eg 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, preferably in the part of the spectrum that contains relatively little noise. Practical: ·. ·. naturally, it would be possible to pre-filter the signal before pu

• I

analysis of human activity, but when audio. The activity detector follows the output of the speech encoder, pre-filtering would distort the audio signal to be encoded.

The invention thus relates to a voice activity detector device comprising: ': (i) a portion of a signal input signal of a first voice activity detector operating by forming' and a portion of the input signal being considered! / being free of speech, for spectral similarity to produce an output signal indicating the presence or absence of speech in the input signal; 115328 2 (ii) data stored in memory obtained from said speech-free portion; and (iii) an additional voice activity detector; Characterized in that the auxiliary voice activity detector controls the memory update, wherein the auxiliary voice activity detector operates by generating a measure of spectral similarity between the current portion of the input signal and the previous portion of the input signal.

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

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

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

Figures 1 and 2 show possible components of an embodiment of the invention, and Figs

Figure 3 illustrates a preferred embodiment of the present invention.

• t: *. ·. The activity of the first voice according to one embodiment of the invention. ·: ·. the general principle underlying the detector is as follows.

• · »25 ....: A frame with n signal samples (s0, sv s2, s3, s4 ... sn-1) is obtained when it is passed through a fourth order finite impulse response (FIR) digital •%» calculator with an impulse response of (1, h0, h1p h2, h3) resulting in t: filtered signal (when samples from previous frames are ignored) 30 s '= (So),: · (s. | + h0s0), j' (s2 + h0s. | + h- | S0), 3 115328 (s3 + h0s2 + h1s1 + h2s0) (s4 + h0s3 + h1s2 + h ^ + h- | S0), (s5 + h0s4 + h1s3 + h2s2 + hgs ^, (s 6 + h0s5 + h1s4 + h2s3 + h3s2), 5 (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 10

R'o = ΣW

i = 0 15 and thus provides a measure of the power of a portion of the signal s that falls within the calculated bandwidth of the calculated filtered signal s', that is, of the calculation filter.

When solving the expression, ignoring the first 4 terms gives 20 R o = (s4 + ^ os3 + h- | S2 + h2s1 + h3s0) + (s5 + h0s4 + h1s3 + h2s2 + h3s / • * * + ...

• «: 25 = S4 + h0s4s3 + h., S4s2 + h2s4s1 + h3s4s0 + hoS4s3 + hgSo + hoh- | S3s2 + h0h2s3s1 + h0h3s3s0

* * h ^ s4s2 + hQh. | S3s2 + h ^ Sj + h ^ h2s2s- | + h ^ h3s2SQ

1 ,,, * ^ h2s4s1 + h0h1s3s1 + h1h2s2s1 ^ h2S ^ 4 · h2h3s ^ SQ

+ h3s4s0 + h0h3s3s0 + h1h3s2s0 + h ^ s ^ o + h23s20 30 + ...

!;> = R0 (1 + ho + h, + hj + hj): + R-, (2h0 + 2h0h1 + 2h1h2 + 2h2h3) + R2 (2h ^ + 2h ^ h3 + 2hgh2),:. + R3 (2h2 + 2h0h3) 35 + R4 (2h3) I i> 115328 4 R'o can thus be obtained from a combination of autocorrelation coefficients Rj, weighted by the constants in parentheses that determine the frequency band in which the value of R'0 affects. The terms in parentheses are actually autocorrelation coefficients for the impulse response of the computational filter, so the above expression can be simplified to

OF

R'o * R0H0 + 2 Σ RA .................................. (1) i = 1 10 where N is the order of the filter and Hj are the autocorrelation coefficients of the impulse response of the filter (normal to no sound).

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 of 20 measures can simulate the effect of a digital filter, which typically requires one hundred times this number of multiplications.

; >: Alternatively, the filtering operation can be viewed as a spectral comparison in the form it compares the signal spectrum with the reference spectrum (inverse of the response of the computational filter v · 25). Since the computational filter is selected in this application: ,, / such that it approximates the inverse of the noise spectrum, this operation can be: "· considered as spectral comparison of speech and noise spectra, and the resulting zero auto, correlation coefficient (i.e. inverse filtered signal energy) The Itakura-Saito measure is used in linear prediction *: 30 coding to evaluate LPC predictor filter compatibility with the result spectrum and can be expressed in one form *

OF

M = RoA0 + 2 Z RA,: · 35 i = i

• 1 I

5 115328 where Aq etc. are the autocorrelation coefficients of the LPC parameter set. It will be seen that the expression is very similar to the dependency derived above, and remembering that the LPC coefficients are pins of an FIR filter having an inverse spectral response of the input signal so that the LPC coefficient set is the impulse response of the LPC filter inverse, that the Itakura-Saito-distortion measure is, in fact, only a form of equation 1 in which the filter response H is the inverse of the spectral form of the model containing only the poles of the input signal.

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 spectral similarity.

The 1-S distortion measure is further discussed in A Buzo, A H Gray, R M Gray, and J D Markel, "Speech Coding Based on Vector Quantization," IEEE 15 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 pu-20 and can thus be used as a measure of M speech expression. In an environment where the noise spectrum is well known and v, constant, it is perfectly possible to simply use fixed coefficients h0, h, etc. to model the inverse noise filter.

$ t • t t * t * * · *, ···, 25 However, a device that can adapt to different noise environments,, ''!; can be used more generally.

In the embodiment of Figure 1, the signal from the microphone (not shown) is received; is input 1 and converted to digital samples s with a suitable sampling (3030 frequency by analog-to-digital converter 2). The LPC analysis unit 3 (included in a known * LPC encoder of a known type) then produces n (e.g. 160) samples in succession »; ;,: for those frames, a set of LPC filter coefficients Lj of N (e.g., 8 or 12) which are transmitted to represent incoming speech. The speech signal s is also provided to the correlator unit!! * * 4 (normally included as part of LPC encoder 3, i 6 115328 correlation vector Rj is developed as one step in the LPC analysis (although it is clear that a separate correlator could be used) .Correlator 4 develops an autocorrelation vector Rj which includes a zero order correlation coefficient R0 and at least two other autocorrelation coefficients R1 (R2, R3. after the multiplier unit 5.

The second input 11 is connected to a second microphone which is far from the speaker such 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 analyzed by the LPC-10 by another LPC analyzer 13. The "noise" -LPC coefficients generated from the analyzer 13 are fed to the correlator unit 14 and the autocorrelation vector the weighted coefficients generated in the multiplier 5 and so combined in the adder 6 according to equation 1 to obtain a filter effect having an inverse shape to the noise spectrum of the noise-only microphone (which is virtually the same as the noise spectrum in the signal and noise receiving microphone) noise. The resulting dimension M is compared to a threshold value in a threshold circuit 7 to generate a logic output 8 indicating the presence or absence of a pu 20. If M is large, speech is considered to occur - *. *. date.

. ·: *. However, this embodiment requires two microphones and two LPC analyzers. ' ". which increases the cost and complexity of the hardware needed.

25th · '·. Alternatively, in another embodiment, a corresponding dimension formed using autocorrelations from the noise microphone 11 and LPC coefficients derived from the microphone 1 is used, so that an additional autocorrelator is needed instead of an extra LPC analyzer.

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

7, 115328

In the embodiment of Figure 2, there is a buffer 15 storing a set of LPC coefficients (or set of autocorrelation vectors) derived from microphone input 1 during a period identified as a "speechless" period (i.e., a mere noise period). 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 approximation. 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 maintains the "noise" autocorrelation coefficients during the "speech frames" but a new set of LPC coefficients can be used during the "noise frames" to update the buffer The outputs 14, each with an autocorrelation coefficient, are coupled to the buffer 15. It is clear that the correlator 14 could be located downstream of the buffer 15. In addition, a speech / speech decoding decision updating of the coefficients does not need to be made at output 8, but could be derived (and preferably derived) in some other way.

20 •. . Because speechless episodes occur frequently, the LPC coefficients stored in the buffer • »: v. Will be updated from time to time so that the device can monitor the noise. : ·. spectrum changes, it is clear that such a buffer update may be necessary. · ·. only occasionally, or it can occur only once when the detector operates ....: 25, if (as is often the case) the noise spectrum is relatively temporal. ' · *; an unchanged but frequent update in a mobile radio environment is »» · cheaper.

. ' * ': In one variant of this embodiment, the system initially uses. '. 30 equations 1 as the multiplication terms correspond to a simple fixed high pass filter and then the system begins to adapt by switching to the "noise time" LPC coefficients. If the speech fails for any reason, the system

I I I

• · ·: Can return to using a simple high pass filter.

»· 8 115328

The above dimension can be normalized by dividing by R0 such that the expression compared to the threshold is of the form n RjAj 5 M = Ao + 2 £ --- i = 1 R0 This dimension is independent of the total signal energy of the frame and is thus compensated for therein , but it provides a weaker contrast between the "noise" and "speech levels" and is therefore most advantageously not used in highly disturbed environments.

Rather than using LPC analysis to derive the inverse filter coefficients of the noise signal (either from a noise microphone or from noise-only cycles, as in the various examples described above), the inverse noise spectrum can be modeled 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 ana-20 lysis unit 13 is simply replaced by an adaptive filter (e.g., an FIR transverse filter or a network filter) coupled to perform; Makes the resulting noise whiter by modeling a reverse filter, and *]: Factors are input as above for autocorrelator 14.

In another embodiment corresponding to the embodiment of Fig. 2, the LPC analysis means 3 is replaced by such an adaptive filter, and the buffer -: ,,, · * means 15 is omitted, but the switch 16 operates so that it prevents the adaptive filter from adapting its coefficients during speech periods.

* * »· 4 ·> 30 The following describes another voice activity detector for use in connection with an embodiment of the invention.

From the foregoing, it is clear 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 establishing between adjacent frames 9,115,328

Itakura-Saito distortion measure, 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 5 large Itakura-Saito distortion metrics so that the measure reflects the spectral similarity of 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, then there should be a significant spectral change in the speech signal between the frames (if not, 10 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 previous frame reverse LPC filter "filters out" most of the noise power.

15

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 and less variable at times.

It should be noted that the standard deviation of the standard deviation of dimension M is also a reliable measure. Effect of generating each standard deviation:. in fact, smooth out the measure.

11 · ,,,,: 25 The parameter measured in this second form of the voice activity detector used to determine whether speech is present is preferably the Itakura-Saito standard deviation of the intersection, but also other fluctuation measurements and others., Spectral distortion measurements (based, for example, on FFT- analysis) butter, · ·. here to use.

30: 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 trimmed. Therefore, the threshold adaptation circuit must be controlled by using a * / * 10 115328 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 so that the threshold is kept just above the level of the measure M5 when noise alone occurs. Because the dimension generally varies randomly with the occurrence of noise, the threshold is altered by determining an average level over a plurality of blocks, and the threshold is set at a level proportional to this average. However, this is generally not sufficient in a noisy environment and thus the determination of the degree of parameter variation over a plurality of blocks 10 is also considered.

The threshold T is thus preferably calculated according to the following expression T = M '+ K.d where M' is the average of the measure over several consecutive frames, d is the standard deviation of the measure during these frames and K is a constant (which can typically be 2).

In practice, it is preferable that the adaptation is not restarted immediately after the speech has been 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 Figure 3, the present invention incorporating the aforementioned features, * · ·. in a preferred embodiment of the invention, the input 1 receives a sampled signal converted to digital by an analog-to-digital converter (ADC) 2 and the signal is input to the input of the inverse filter analyzer 3, which is in practice part of the speech encoder 30 the sound activity detector is intended to operate and generates the coefficients L i (typically 8) of the filter corresponding to the inverse of the input signal spectrum.

• The digitized signal is also fed to autocorrelator 4 (included as part of analyzer 3), which generates an input signal autocorrelation vector R | (or at least as many lower-order terms as LPCs). The operation of these parts of the device 11 115328 is as described in Figures 1 and 2. Hereby, averages of the autocorrelation coefficients R1 are preferably generated over a plurality of consecutive speech frames (typically 5-20 msec) to improve their reliability. This can be accomplished by storing each set of autocorrelation coefficients provided by the autocorrelator 45 in buffer 4a and using the averaging factor 4b to generate a weighted sum of the current autocorrelation coefficients Rj and previous frames stored in and input to the buffer 4a. The average autocorrelation coefficients Ra thus derived are supplied to the weighting and summing means 5, 6, which also receive the stored noise period inverse filter filter coefficients Lf from the autocorrelator 14 of the autocorrelation vector A 1 and preferably form Raj and A, M defined as follows:

RajAj 15 M = A0 + 2 £ ------

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.

20

For the inverse filter coefficients L | would correspond to a reasonable estimate of the inverse of the noise · 'spectrum, these coefficients are desirable to be updated during the noise periods (and of course not updated during the speech periods), however: • it is preferable that the speech / speechless decision on which the update is based does not depend refresh: "25 as a result thereof, or else a single misidentified signal frame may cause the voice activity detector to" unlock "thereafter, and: '': incorrect identification of the following frames. Therefore, it is preferable to use a control signal generating circuit 20 which is a voice i activity detector generating an independent control signal which ': indicates the presence or absence of speech to control the inverse filter analyzer 3 (i.e., buffer 8) such that the inverse filter autocorrelation coefficients A 1 used to generate the dimension M are updated only for "noise periods" "during the signal generating circuit 20 includes an LPC analyzer 21 ;;; (which may also be part of a speech encoder and may be implemented in particular by analyzer 35) generating a LPC coefficients Mj corresponding to the input signal Dec 12 115328 kon, and an autocorrelator 21a (which may be implemented by autocorrelator 3a) which derives autocorrelation coefficients B If the analyzer 21 is implemented by the analyzer 3, then Mj = Ls and Bj = Aj. These autocorrelation coefficients are then applied to the weighting and summing means 22, 23 (corresponding to the elements 5, 6), which also receive an input signal from the autocorrelator 4 of the autocorrelation vector R1, thereby calculating the spectral similarity between the incoming speech frame and the previous speech frame. This dimension may be the Itakura-Saito-distortion measure between the current frame coefficients Rj and the previous frame's coefficients Bj, as above, or it may instead be calculated as a 10-bar Itakura-Saito distortion measure for the current frame coefficients Rj and Bj and subtracting (in subtracting means 25) a corresponding prior measure stored in buffer 24 to generate a spectral signal (in each case, the energy of the measure is normalized by dividing by R0). The buffer 24 is then naturally updated. This spectral signal, after the threshold comparison performed in the threshold circuit 26 described above, 15, indicates the presence or absence of speech, but we have found that although this measure is excellent for distinguishing noise from silent speech to distinguish noise from pronunciation.

Accordingly, the circuit 20 preferably further utilizes a spoken voice detector circuit having a pitch analyzer 27 (which may, in practice, function as part of a speech encoder and, in particular, may 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 '' '; 25 this signal is coupled with a threshold value derived from threshold circuit 26, a measured dimension (which is generally "true" in the presence of silent speech) OR to NO inputs 28 to generate a signal that is "false" in the presence of speech and "true" noise. . This signal is supplied to buffer 8 ·. (or inverse filter analyzer 3) so that the inverse filter coefficients Li are updated only during the noise periods.

1 I

The threshold adaptation circuit 29 is also coupled to receive a control signal output from the control signal generator circuit 20; The output of the threshold adapters ,, ,, ,, is supplied to the threshold circuit 7. The threshold adaptation circuit 115 115328 13 relaxes or reduces the threshold in steps proportional to the current threshold value until the threshold approximates the noise power level (which can be practically derived) .

5 When the input signal is very low, it may be advantageous to set the threshold automatically 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.

Further, "overrun" generating means 30 can be used to measure the duration of speech detection after threshold circuit 7, and when speech occurrence is detected over a predetermined time constant, the output is maintained in a higher state for a short "overrun" period. In this way, low-level centering of speech bursts is avoided and correct selection of the time constant prevents triggering of the crossing 15 generator 30 by short, incorrectly pronounced noise peaks.

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

I · · I · * t · I. 'As explained above, the voice detector can be practically implemented I *' '. 25 As part of the LPC codec. Alternatively, when the signal autocorrelation coefficients t> * or related dimensions (partial correlation, or "parcor" coefficients) are transmitted to a remote station t · I ·, voice detection can occur far from the codec.

* · ». »<·» 1 f I •: I · I · «• t * ·

Claims (7)

1. Voice activity detector apparatus comprising: (i) a first voice activity detector (3-6,14), which operates by creating a matte of the spectral similarity between a portion of the input signal and the portion of the input signal considered free from tai for forming such an output signal showing the presence or absence of speech in the input signal; (ii) data stored in memory (15) obtained from said voiceless portion, and (iii) an additional detector (20) for voice activity; ::: characterized in that the voice detector auxiliary detector (20) alone controls the updating of the memory (15), the voice activity auxiliary detector (20) acting by forming a *: measured by the spectral similarity between the current portion of the input signal and the previous portion of the signal. 2. Voice activity detector apparatus comprising: / (i) means (1) for receiving an input signal; (ii) memory (15) for storing the signal representing noise, which signal. * represents an estimate of the noise component of said input signal; (iii) means (3-6, 14) for periodically forming a measure of the spectral similarity between the input signal and said estimated noise component of the input signal and said signal representing ' (Iv) means (7) for comparing the measurement with a threshold value T for generating an output signal to indicate the presence or absence of tai; (v) an auxiliary detector (20) for voice activity; (iv) updating means for updating the memory from the input signal; Characterized in that the auxiliary detector for voice activity operates depending on the output generated from the measure of spectral similarity between the current portion of the input signal and the preceding portion of the input signal, which output indicates the presence or absence of tai, and that the memory update device is only operable for the memory when said control signal indicates absence of tai. 10 Device according to claim 2, characterized in that it further comprises means for controlling said threshold value during such periods when said control signal indicates the absence of tai. Device according to claim 2 or 3, characterized in that said additional detector for voice activity further comprises vocalized tai detection devices (27), which include speech height analyzers for generating a signal indicating the presence of vocalized tai, where it is controlled the additional detector (20) for voice activity also depends on said signal. 20 »»: 5. Device for encoding the voice signal, characterized in that it comprises a device according to any of the preceding claims. I # »• · • · · • · f Mobile phone device, characterized in that it comprises a device according to any preceding claim. A method of detecting voice activity in the input signal, which comprises the following steps: receiving said input signal; ** estimating the noise signal component of said input signal; / storing data representing said noise signal component; »: Creating the measurement M of the spectral similarity between a portion of the input signal and said noise signal component; and comparing a parameter derived from the measured M with a first threshold T to form an indication of the activity of the primary voice to show the presence or absence of tai, depending on whether or not this value is exceeded, wherein said estimation phase includes a additional display of voice activity and which method is characterized in that said additional display of voice activity includes; forming a measured spectral distortion of the similarity between the current portion of the input signal and the previous portions of the input signal; comparing the spectral distortion with one another to create an indication of the presence or absence of tai, depending on whether this value is exceeded or not; and updating said stored data of the input signal, only from such periods where said supplemental indication of voice activity indicates absence of tai. 15 «·» • · · • * • · I t <I • · • * (<> • «« • · »• * * • · •> •« * i i »*» »»