PT89978B - DEVECTOR OF THE VOCAL ACTIVITY AND MOBILE TELEPHONE SYSTEM THAT CONTAINS IT - Google Patents

Abstract

Voice activity detector (VAD) for use in an LPC coder in a mobile radio system, uses autocorrelation coefficients R0, R1..... of the input signal, weighted and combined, to provide a measure M which depends on the power within that part of the spectrum containing no noise, which is thresholded against a variable threshold to provide a speech/no speech logic output. The measure is <MATH> where Hi are the autocorrelation coefficients of the impulse response of an Nth order FIR inverse noise filter derived from LPC analysis of previous non-speech signal frames. Threshold adaption and coefficient update are controlled by a second VAD responsive to rate of spectral change between frames.

Description

DESCRIPTION

GIVES

INVENTION PATENT

No. 89 978

APPLICANT: BRITISH TELECOMMUNICATIONS public limited company, British, industrial and commercial, based at 81 Newgate Street, London EC1A 7AJ, England.

TITLE:

VOCAL ACTIVITY DETECTING DEVICE AND MOBILE TELEPHONE DEVICE CONTAINING 0

INVENTORS: Daniel Kennetw Freeman and Ivan Boyd.

Claim of the right of priority under Article 4 of the Paris Convention of 20 March 1883.

United Kingdom on 11 March 1988, 2006

June 1988 and on August 24, 1988, under the n ⁹ s ......

8805795, 8813346.7 and 88201105.8, respectively.

INPI. MOD. 113 R F 18732

Description referring to the BRITISH TELECOMMUNICATIONS public limited company, British, industrial and commercial patent, based at 81 Newgate Street, London EClA 7AJ, England, (inventors: Daniel Kennetw Freeman and Ivan Boyd, residing in England), for DETECTOR APPLIANCE VOCAL ACTIVITY AND MOBILE TELEPHONE APPLIANCE THAT CONTAINS IT.

description

A vocal activity detector is a device to which a signal is provided for the purpose of detecting periods of conversation or periods containing only noise. Although the present invention is not limited to this, an application of particular interest to such detectors is in mobile radiotelephone systems, where knowledge of the presence or absence of conversation signals can be exploited by a speech encoder to improve efficient use of the frequency spectrum, adioelectric and where it is also very likely that the noise level (from a unit mounted on the vehicle) is high.

The essence of detecting vocal activity is to determine a measure that differs appreciably between speech and non-speech periods. In devices that include a speech encoder, a number of parameters are available from either stage of the encoder and it is therefore desirable to save on the processing required by using some of these parameters. In many environments, the main noise sources occur in defined areas known from the frequency spectrum. For example, in a car, a large part of the noise (eg engine noise) is concentrated in the low frequency areas of the spectrum. When such knowledge of the spectral position of the noise is available, it is desirable to base the decision on whether speech is present or absent on measurements made from the portion of the spectrum that contains relatively little noise. It would of course be possible in practice to pre-filter the signal prior to its analysis to detect the vocal activation, but when the vocal activity detector follows the output of a speech encoder, pre-filtering distorts the vocal signal to be encoded.

According to a first aspect of the present invention, a device for detecting vocal activity is provided which comprises means for receiving an input signal, means for estimating the noise signal component in the input signal, means for continuously forming a measurement M of the spectral similarity between a portion of the input signal and the noise signal and means to compare a parameter derived from measure M with a threshold value T to produce an output to indicate the presence or absence of speech, depending on that value is exceeded or not.

According to a second aspect of the present invention there is provided an apparatus for detecting vocal activity which comprises: means for continuously forming a measure of the spectral distortion of similarity between a portion of the input signal and more primitive portions of the input signal and means to compare the degree of veriation between successive values of the measure with a threshold value, to produce an output that indicates the presence or absence of speech depending on whether this value is exceeded or not.

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

Other aspects of the present invention are defined in the claims.

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

Fig. 1, a block diagram of a first form

ORIGINAL BAD

of carrying out the present invention;

Fig. 2, a second embodiment of the present invention;

Fig. 3, a third preferred embodiment of the present invention.

general principle underlying a first vocal activity detector according to a first embodiment of the present invention is as follows.

From a plot of n samples of the signal the ' ^S l' ^S 2 ' ^S 3'., ^S n_jJ, when passing through a digital filter with finite response to the ideal fourth-order unitary impulse (FIR), with the response to the impulse unitary (1, h ^, h ^, h ₂ , h ^), a filtered signal will result (ignoring previous frame samples) '= < ^s 0>

(s. (s. (s + h _o s _o ), + ' _Vl ^h i ^s o' + h ₀ s ₂ + h _lSl + h ₂ s _Q ), + h ₀ s ₃ + 1> _{Λ +} + · _νο >, ^(S 5 ^{+ h} 0 ^S S ^{+ h} l ^S 3 ^{+ h} ? ^S 7 ^{+ h} ri '· ^(s l (s

0''5 + + h ₂ s ₃

1 '^ 3 ^S 2)' zero order autocorrelation coefficient is the sum of each term squared, which can be normalized, that is, divided by the total number of terms (for constant lengths of the frames it is easier to omit the division ); the value of the filtered signal will thus be

(so this value is therefore a measure of the power of the ideal filtered signal s ¹ - in other words, of the part of the signal s_ that falls in the pass band of the ideal filter.

Developing and neglecting the first 4 terms

R ¹ = (s + hs + hs „+ h., S. H _o s' 0 4 03 1 2 2130

+ is ,. + h _Q s ₄ + ^h i ^s 3 + ^h 2 ^s 2 ^{+ h} 3 ^S l?

+ ...

+ +

+ +

2 ^S 4 + ^h 0 ^S 4 ^S 3 + ^h l ^s 4 ^s 2 + ^h 2 ^S 4 ^S l + ^h 3 ^S 4 ^S 0 ^h 0 ^s 4 ^S 3 4- ^h 0 ^s 0 + Vl ^S 3 ^s 2 + ^h 0 ^h 2 ^s 3 ^s l + ^h 0 ^h 3 ^s 3 ^S C ^h l ^s 4 ^s 2 + ^h 0 ^h l ^S 3 ^S 2 + u2 2 ^h l ^s 2 + ^h l ^h 2 ^s 2 ^s l + ^h l ^h 3 ^s 2 ^s C ^h 2 ^s 4 ^s l + ^h 0 ^h L ^S 3 ^S l + ^h l ^h 2 ^S 2 ^s l + .2 2 ^h 2 ^s l + ^h 2 ^h 3 ^S l ^S 0 ^h 3 ⁼ 4 ^S , l + ^h 0 ^h 3 ^S 3 ^S 0 + ^h l ^h 3 ^s 2 ^s 0 + ^h 2 ^h 3 ^s l ^s 0 2 2 ^h 3 ^s 0

^R 0 ( ^{1 + h} 0 ^{+ h} l ^{+ h} 2 ^{+ h} 3 ^ + R (- ^h 0 + ^2h Q ^h i + ^2h i ^h 2 ^{+ 2h} 2 ^h 3 ^

4- R ₂ (2h ^ + 2h ₃ h ₃ + 2h _Q h ₂ ) + ( ^2h 2 ^{+ 2h} 9 ^h 3 ^

R ₄ '2h ₃ )

Thus, R'q can be obtained from a combination of the autocorrelation coefficients R ^, weighted by means of the constants in parentheses that determine the frequency band to which the value of R'q responds - In fact, the terms in parentheses are the autocorrelation coefficients of the ideal filter response to the unit impulse, so that the previous expression can be simplified to

N_ ^R 'o = ^R o ^H o ^{+ 2} T ~ ^r í ^h í- ⁽¹⁾ i = 1 where N is the order of the filter and H. are self-correcting coefficients (non-normalized) of the filter response to the unit impulse .

In other words, the effect on the signal autocorrelation coefficients of the filtering of a signal can be simulated by producing a weighted sum of the auto coefficients4

correlation of the signal (unfiltered), using the impulse response that the required filter would have.

Thus, a relatively simple algorithm, involving a small number of multiplication operations, can simulate the effect of a digital filter that typically requires a hundred times this number of multiplication operations.

This filtering operation can alternatively be seen as a way of comparing spectra, with the signal spectrum being compared with a reference spectrum (the inverse of the ideal filter response). As the ideal filter in this application is chosen in order to approximate the inverse of the noise spectrum, this operation can be considered as a spectral comparison between the speech and noise spectra and the zero order autocorrelation coefficient thus generated (this ie, the filtered reverse signal energy) as a measure of the dissimilarity between the spectra. The Itakura-Saito distortion measure in the LPC is used to establish the adaptation between the predictor filter and the input spectrum, in a form expressed by

Μ = R „A„ + 2 i = l where Αθ ... are the autocorrelation coefficients of the set of LPC parameters. It will be seen that this is very similar to the relation deduced above and, if we remember that the LPC coefficients are the derivations of a FIR filter that has the inverse spectral response of the input signal so that the set of the LPC coefficients is the impulse response of the inverse LPC filter, it will be evident that the measure of the Itakura-Saito distortion is in fact simply a form of equation (1), with the H response of the filter being the inverse of the spectral shape of the model with all poles of the input signal.

In fact, it is also possible to transpose the spectrum using the LPC coefficients of the test spectrum and the autocorrelation coefficients of the reference spectrum to obtain a different measure of spectral similarity.

The measurement of I-S distortion is discussed in detail

in Speech Coding based upon Vector Quantisation by A. Buzo, A.H. Gray, R. M. Gray and J.D. Markel, IEEE Trans. on ASSP, Vol. ASSP-28, N ° 5, October 1980.

As the signal frames are only finite in length and a number of terms are neglected (N, where N is the order of the filter), the previous result is only an approximation; it does however provide a surprisingly good indicator of the presence or absence of speech and can therefore be used as measure M in speech detection. In an environment where the care spectrum is well known and stationary, it is very possible to simply use fixed coefficients h ^, h ^, ... for the reverse noise filter model.

However, a device that can adapt to different noise environments is much more useful.

With reference to fig. 1, in the first embodiment, a signal from a microphone (not shown) is received at an input (1) and converted into digital samples _s with an appropriate sampling frequency via an analog-to-digital converter (2) (ADC ). An LPC encoding analysis unit (3) (in an LPC encoder of known type) then deducts, for successive frames of n (eg 160) samples, a set of N (eg 8 or 12) LPC filter coefficients, which are transmitted to represent speech at the entrance. The speech signal _s also enters a correlating unit (ACF) (4) (normally part of the LPC encoder (3), since the speech autocorrelation vector R ^ is also produced as a phase of the LPC analysis, although understand that a separate correlator can be provided). The correlator (4) produces the autocorrelation vector R ^, including the zero order correlation coefficient Ρθ and at least two more autocorrelation coefficients R ^, R ₂ , R ^. These are then supplied to a multiplier unit (5).

A second input (11) is connected to a second microphone located far from the speaker in order to receive only background noise. The input of this microphone is converted into a series of digital input samples, by the analog-to-digital converter ADC (12), and analyzed for the LPC by a second analysis.

_ LPC user (13). The noise LPC coefficients produced by the LPC analyzer (13) pass to the correlator (14) and the autocorrelation vector thus produced is multiplied term-by-term by the autocorrelation coefficients R of the input signal from the speech microphone, in the multiplier. (5), and the weighted coefficients thus produced are combined in the adder (6) according to equation 1 in order to apply a filter with the inverse form of the noise spectrum coming from the microphone that captures only the noise (which in practice it is the same shape as the noise spectrum in the microphone that captures the signal plus the noise), thus filtering out most of the noise. The resulting measure M is subjected to a threshold in the threshold circuit (7) to produce a logic output (8) that indicates the presence or absence of speech; if M has a high level, it is assumed that speech is present.

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

Alternatively, another embodiment uses a corresponding measure formed using the autocorrelations of the microphone (11) that captures only noise and the LPC coefficients coming from the main microphone (L), so that an extra autocorrelator is needed, instead of a another LPC analyzer.

These embodiments can therefore operate in different environments having noises with different frequencies or in a noise spectrum that varies in a given environment.

Referring now to fig. 2, in the preferred embodiment of the present invention, a buffer (15) is provided which stores a set of LPC coefficients (or the set's autocorrelation vector) deduced from the microphone input (1) in a period identified as being a period with no speech (that is, only with noise). These coefficients are then used to deduce a measure using equation 1, which of course also corresponds to the Itakura-Saito distortion measure, except that it is only used. a single frame is recorded in the memory of LPC coefficients cor7 corresponding to an approximation of the inverse noise spectrum, instead of the piresent frame of LPC coefficients.

The L vector of the LPC coefficients left from analyzer 1 (3) is also routed to a correlator (ACF) (14), which produces the autocorrelation vector of the LPC coefficient vector. The buffer memory (15) is controlled by the speech presence / absence output of the threshold circuit (7), so that during speech frames the buffer memory retains the noise autocorrelation coefficients, but during noise frames a new set of LPC coefficients can be used to update the buffer, for example by a multiple switch (16), through which the outputs of the correlator (14), which transmit each of the autocorrelation coefficients, are connected to the memory buffer (15). It will be understood that the correlator (14) can be placed after the buffer (15). In addition, the decision of the presence / absence of speech to update the coefficients does not need to be derived from the output (8), but can (and is, preferably) derived in another way.

As frequent speechless periods occur, the LPC coefficients stored in the buffer are updated from time to time, so that the device is thus able to follow changes in the noise spectrum. It will be understood that such an update of the buffer may be necessary only occasionally or may occur only once at the beginning of the detector's operation, if (as often happens) the noise spectrum is relatively stationary in time, but in the case of a mobile radiotelephone environment, frequent updating is preferred.

In a modification of this embodiment, the system initially uses equation 1 with terms of the coefficients corresponding to a simple fixed high-pass filter and then begins to adapt, by switching, to the use of the LPC coefficients of the noise periods. If, for any reason, speech detection fails, the system can return to using the simple high-pass filter.

It is possible to normalize the previous measure by dividing by Rq, so that the expression to be submitted to the threshold circuit has the form

N = ^A 0 + ² T ~ i = i Rq

This measure is independent of the total signal energy in a frame and is thus compensated for wide variations in level, but it gives a somewhat less marked contrast between noise and speech, and is therefore not preferably used in a high level environment.

Instead of using the LPC analysis to deduce the coefficients of the inverted filter of the noise signal (from periods of the noise microphone or from periods with only noise, as in the various previous embodiments), it is possible to model the inverse spectrum of the noise using an adaptive filter of known type; as the noise spectrum changes only slowly (as discussed below), a relatively slow adaptation rate of the coefficients common for these filters is acceptable. In an embodiment, which corresponds to fig. 1, the LPC analysis unit (13) is simply replaced by an adaptive filter (for example a transverse FIR filter or a mesh filter) connected in order to make the noise white at the input by modeling the reverse filter, and the its coefficients provided as before to the autocorrelator (14).

In a second embodiment, corresponding to fig. 2, the LPC analysis device (3) is replaced by such an adaptive filter, omitting the buffer memory (15), but the switch (16) works to prevent the adaptive filter from adapting its coefficients during speech periods .

A second speech activity detector will now be described in accordance with another aspect of the present invention.

From the above, it will be evident that the vector of the LPC coefficients is simply the response to the impulse of a 9-

another FIR that has a response that approximates the inverse spectral shape of the input signal. When the Itakura-Saito distortion measure is formed between adjacent frames, it is in fact equal to the signal strength, as it is filtered by the LPC filter of the previous frame. Thus, if the spectra of the adjacent frames differ little, a correspondingly small portion of the spectral power of a frame will be escorted to the filtering and the measurement will be low. Correspondingly, a large inter-frame spectral difference produces a high measure of the Itakura-Saito distortion, so that the measurement reflects the spectral similarity of the adjacent frames. In a speech encoder, it is desirable to minimize the data transmission rate, so that the duration of the frame is made as long as possible; in other words, if the frame duration is long enough, then a speech signal presents a significant spectral variation from frame to frame (if not, the coding is redundant).

On the other hand, the noise has a slow spectral variation from frame to frame and thus, in a period when speech is absent from the signal, the measure of the Itakura-Saito distortion will be correspondingly low - since the application of the inverse LPC filter from of the previous plot separates by filtration most of the power of the noise.

Typically, the measure of Itakura-Saito distortion between adjacent frames of a noise signal containing intermittent speech is higher during speech periods than during noise periods; the degree of variation (as illustrated by the standard deviation) is higher and less variable intermittently.

Note that the standard deviation of M is also a reliable measure; the effect of taking each standard deviation is essentially to make the measurement more uniform.

In this second form of the vocal activity detector, the measured parameter used to decide whether speech is present is preferably the standard deviation of the Itakura-Saito distortion measure, although other measures of variance and other measures of spectral distortion can be used ( based, for example, on the FFT (Fast Fourier Transform)

Fast Fourier)). It has been found to be advantageous to use an adaptive threshold in the detection of vocal activity. Such thresholds do not need to be adjusted during speech periods or the speech signal will be eliminated by a threshold. It is therefore necessary to control the threshold adapter using a signal to control the presence / absence of speech, it being preferable that this control signal is independent of the output of the threshold adapter.

threshold T is adjusted adaptively so as to maintain the threshold level precisely above the level of measure M when only noise is present. As the measurement will generally vary randomly when noise is present, the threshold variation is made by determining an average level extended to a certain number of blocks and adjusting the threshold to a level proportional to this average. In a noisy environment, however, this is not enough, so a determination of the degree of variation of the parameter in several blocks is also taken into account.

The threshold value T is therefore preferably calculated according to

T = M '+ K.d where Μ' is the average value of the measure over a number of consecutive frames, gives the standard deviation of the measure in these frames and K is a constant (which can typically be 2).

In practice, it is preferred not to resume adaptation immediately after the absence of speech has been indicated, but rather to wait to ensure that the fall is stable (to prevent repeated rapid switching between the adaptation and non-adaptation states).

Referring now to fig. 3, in a preferred embodiment of the present invention, which incorporates the above aspects, an input (1) receives a signal that is scanned and digitized by the analog-to-digital converter (ADC) (2), and supplied to the input of an analyzer reverse filter (3), which in practice is part of a speech encoder that the vocal activity detector has to work with and that generates coefficients (typically 8) of a filter corresponding to the inverse spectrum of the input signal. The digitized signal is also supplied to an autocorrelator (4) ACF (which is part of the analyzer (3)) that generates the vector R ^ of autocorrelation of the input signal (or at least as many lower order terms as the LPC coefficients). The operation of these parts of the apparatus is as described for figs. 1 and 2. Preferably, the autocorrelation coefficients are then transformed into an average spanning several successive speech frames (typically 5 to 20 ms in duration) to improve their reliability. This can be achieved by storing each set of autocorrelation coefficients by the autocorrelator (4) in a buffer (4a) and using an averaging calculator (4b) (AV) to produce a weighted sum of the current autocorrelation coefficients R. and the of the previous frames stored in the buffer memory (4a) and supplied by it. The mean Ra ^ of the autocorrelation coefficients thus obtained is provided to a weighting and addition medium (5,6) which also receives the autocolor vector A of the inverse filter coefficients in the memorized noise period, from the autocorrelator ( 14) through the buffer memory (15), forming from Ra ^ and A ^ a measure M preferably defined by

N

M = A _n + 2> Ra.A.

/ li

This measurement is then subjected to a threshold circuit (7) for comparison with a certain threshold level, the logical result providing an indication of the presence or absence of speech at the output (8).

In order for the coefficients of the inverse filter to correspond to a good estimate of the inverse of the noise spectrum, it is desirable to update these coefficients during periods of noise. do (and, of course, do not update them during speech periods) ,. However, it is preferable that the decision presence / absence of speech on which the update is based does not depend on the result of the update, or else a single signal frame that is wrongly identified may result in the vocal activity detector being detached and mistakenly identifying the

following plots. Therefore, preferably a control signal generator circuit (20) is provided, effectively a separate vocal activity detector, which forms an independent control signal that indicates the presence or absence of speech to control the reverse filter analyzer ( 3) (or the buffer (8)), so that the correlation coefficients of the inverse filter used to form the measure M are only updated during the noise periods. The control signal generator circuit (20) includes the LPC analyzer (21) (which again is part of a speech encoder and, specifically, it can be the analyzer (3)) which produces a set of LL coefficients of the LPC corresponding to the input signal and an autocorrelator ACF (21a) (which can be the autocorrelator (3a)) that deducts the autocorrelation coefficients of PL.

If the analyzer (21) is the analyzer (3), then M. = L. and B. = A ..

1111

These autocorrelation coefficients are then supplied to a weighting and addition device (22,23) (equivalent to (5,6)) which also receives the autocorrelation vector FL from the input signal from the autocorrelator (4). Thus, a measure of the spectral similarity between the incoming speech frame and the previous speech frame is calculated; this measure can be the measure of the Itakura-Saito distortion between R ^ of the present frame and the previous frame, as explained above, or instead, it can be derived by calculating the Itakura-Saito distortion measure for and from the present frame, and subtracting (in the subtractor (25)) the measurement corresponds to the previous frame stored in the buffer (24) to generate a signal of spectral difference (in any case, the measurement is preferably normalized with respect to energy, dividing by R ^ ). Of course, the buffer (24) is then updated. This signal of spectral difference, when subjected to the threshold circuit (26), as described above, is an indicator of the presence or absence of speech. However, it has been found that although this measure is excellent for distinguishing noise from non-vocal speech (a problem that systems in the prior art are generally unable to resolve) it is in general less able to distinguish noise from vocal speech. Therefore, in addition, in the circuit (20) it is preferably provided with a detection circuit

of vocal speech comprising a pitch pitch analyzer (27) (which in practice can be part of a speech encoder and in particular can measure the value of the long-term predictor delay produced in a multi-pulse LPC encoder). The analyzer (27) produces a logical signal that is true when vocal speech is detected, this signal, together with the measurement subject to the threshold circuit derived from the threshold circuit (26) (which will generally be true when the non-vocal speech), taken to the inputs of a NOR gate (28) to generate a signal that is false when speech is present and true when noise is present. This signal is supplied to the buffer memory (8) (or to the inverse filter analyzer (3)) so that the inverse filter coefficients L ^ are updated only during the noise periods.

The threshold adapter (29) is also connected to receive the control output of the speechless signal of the circuit (20). The output of the threshold adapter (29) is supplied to the threshold circuit (7) ,. The threshold adapter operates to increase or decrease or decrease the threshold, by steps that are proportional to the instantaneous value of the threshold, until the threshold approaches the level of the noise power (which can be conveniently derived, for example, from the weighting and addition (22,23)). When the input signal is very low, it may be desirable for the threshold to be automatically adjusted to a fixed, low level, since at low signal levels the effect of quantifying the signal produced by the analog-to-digital converter (2) can produce results unreliable.

A retention device (30) can also be provided, which works to measure the duration of the speech indications after the threshold circuit (7) and, when the presence of speech has been indicated for a predetermined period of time , the output is retained for a short retention time. This prevents the suppression of low-level speech data streams in the middle, preventing an appropriate selection of the time constant from triggering the retention generator (30) by short-term peaks that are falsely indicated as speech. Of course, it is understood that all of the above functions can be performed by a single digital processing device programmed in an appropriate manner, for example a digital signal processor formed on a DSP (Digital Signal Processing) plate, forming part of an LPC encoder-decoder. thus carried out (this being the preferred embodiment), or as a microcomputer or a microcontroller on a nameplate with the associated memory device.

Conveniently, as described above, the voice detection apparatus can be carried out in practice as part of an LPC codec. Alternatively, when signal autocorrelation coefficients or related measures (partial correlation or plot coefficients) are transmitted to a distant station, voice detection can be done at a distance from the codec.

Claims (1)

- Apparatus for detecting vocal activity, characterized by comprising means for receiving an input signal, means for adaptively estimating the noise signal component in the input signal, means for periodically forming a measure M of the spectral similarity between a portion of the input signal and the noise signal component, and means for comparing a parameter derived from measure M with a threshold value T, and means for producing an output to indicate the presence or absence of speech as a result of that value is exceeded or not. - Apparatus according to claim 1, characterized in that the adaptive means for estimating noise comprise means for calculating the correlation coefficients of the response to the unit impulse of a FIR filter that has a response that approximates the inverse of the spectrum. short-term component of the noise signal, and the means for forming the measurement comprise means for calculating the autocorrelation coefficients R ^ of the signal, connected means for receiving F ε A., and for calculating M from them, the parameter the value of M. - Apparatus according to claim 2, characterized in that it is M = R „A ₀ + R _i A _i Apparatus according to claim 2, characterized by being M = A + 25 R. THE . ⁰ Z, 1 1 ^R 0 Apparatus according to any of claims 2 to 4, characterized in that it further comprises an input arranged to receive: a second signal, similarly subject to noise, from which the vocal signal is absent, comprising the means of calculating the means. LPC analysis to derive A ^ values from the second signal. Apparatus according to any of claims 2 to 4, characterized in that it further comprises a connected buffer for storing data from which the autocorrelation coefficients A ^ of said filter response can be derived, said filter response being calculated periodically16 from the signal by the means of analysis LPC, the apparatus being connected and controlled in such a way that measure M is calculated using said stored data and said stored data being updated only from periods when it is found to be absent speech. Apparatus according to any of claims 1 to 4, characterized in that the means for estimating noise include an adaptive filter. Apparatus according to any of claims 2 to 6, characterized in that the means for calculating the autocorrelation coefficients of the signal are arranged to make this calculation according to the autocorrelation coefficients of several successive portions of the signal. Apparatus according to claim 1, characterized in that the measure M is a measure of spectral distortion. - Apparatus according to claim 9, characterized in that measure M is the measure of distortion Itakura-Saito. Apparatus according to any of the preceding claims, characterized in that it further comprises means for adjusting said predetermined threshold T during periods when speech is found to be absent. cr ' 123 Apparatus according to claim 11, further characterized in that it comprises second means of detecting vocal activity arranged to prevent the adjustment of the threshold value when speech is present. Apparatus according to claim 11 or 12, characterized in that the threshold value T, when adjusted, is adjusted to be equal to the measurement mean plus a term that is a function of the measurement standard deviation. Apparatus according to claim 6, characterized in that it further comprises means for indicating the absence of speech to control the updating of said stored data, the means for indicating the absence of speech consisting of a second device for detecting activity vocal. Apparatus according to claims 13 and 14, characterized in that the second vocal activity detection device controls both the adaptation of the threshold and the updating of the data. Apparatus according to claims 13 to 15 or 17, characterized in that said second vocal activity detection device comprises means for generating a measure of the spectral similarity between a part of the input signal and previous portions of the input signal. 17th Apparatus according to claim 16, characterized in that the means for generating the measure of similarity comprises means for providing a present measure of distortion from the data of the predictive linear encoding (LPC) filter and the autocorrelation data referring to a present part of the input signal, means for providing a measure of the equivalent distortion of a passed frame corresponding to a previous portion of the input signal and means for generating a signal indicating the degree of similarity between them as an indicator of the presence or absence of signals vocals. Apparatus according to claim 16 or 17, characterized in that said means of detecting vocal activity comprise means of detecting vocal signals which comprise means of analyzing the pitch of the sounds to generate a signal indicative of the presence of vocal signals of which also depends on the output of said means of detecting vocal activity. - 19§ Process for the detection of vocal activity in a signal, characterized by comprising the phases of comparing the signal spectrum with an estimated noise spectrum, forming a variable measure of the spectral similarity between them and comparing that measure with a threshold. - Device for encoding speech signals, characterized in that it includes devices according to any of the preceding claims. 21st Mobile telephone apparatus, characterized in that it includes apparatus according to any of the preceding claims. The applicant declares that the first applications for this patent were filed in the United Kingdom on 11 March 1988, 6 June 1988 and 24 August 1988 under the NQs. 8805795, 8813346.7 and 8820105.8, respectively. Lisbon, March 10, 1989.