DE68929442T2 - Device for recording speech sounds - 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

A speech sound detector is a device to which a signal is applied with the aim of detecting periods of speech or periods containing only noise. Although the present invention is not so limited, an application of particular interest for such detectors is mobile radio telephone systems where knowledge of the presence or absence of speech sounds can be used by a speech encoder to improve the efficient use of the radio spectrum and where the noise level (from vehicle-mounted units) is also likely to be high.

The essence of speech sound detection is the location of a measurement which clearly distinguishes between periods in which speech sounds are present and periods in which speech sounds are not present. In devices incorporating a speech encoder, a number of parameters from one or other stage of the encoder are available at any time and it is therefore desirable to make the necessary processing economical by using some of these parameters. In many environments the main sources of noise occur in known, defined regions of the frequency spectrum. In a moving vehicle, for example, much of the noise (e.g. engine noise) is concentrated in the low frequency regions of the spectrum. If such knowledge of the spectral position of the noise is available, it is desirable to base the decision as to whether or not speech sounds are present on measurements taken from the part of the spectrum which contains relatively little noise. In practice it would of course be possible to pre-filter the signal before analysis for speech sound detection, but if the speech detector is to be connected to the output a speech coding device, pre-filtering would distort the speech signal to be coded.

US4358738 discloses a speech detector which compares the input signal with a given noise characteristic by filtering the input signal through two manually balanced bandpass filters (for which analog components are used) to produce two frequency dependent energy segments. This method is of limited use for many reasons; firstly, such a crude arrangement ignores the fact that many types of noise between two bands can have an energy balance similar to a speech signal, secondly, balancing the filters is labor intensive and requires manual detection of the noise periods for balancing, and thirdly, such a device cannot adapt to changing noise or spectrum changes in the environment (or the communication channels).

In "IEEE Transactions on Acoustics, Speech Signal Processing", vol. ASSP-25, no. 4, August 1977, pages 338-343, Rabiner et al., "Application of an LPC Distance Measure to the Voiced Unvoiced Silence Detection Problem", a classifier for distinguishing between silence, unvoiced speech and voiced speech transmitted over a telephone line is disclosed. The method involves initially using manually classified frames of "silence" and "voiced" and "unvoiced" speech signals to drive reference patterns and then comparing the input signal to each of these frames using a comparison measure and selecting the reference frame that is closest to the input signal. This method shares some of the disadvantages of US4358738 in that it requires extensive manual intervention in selecting the frames representing "silence" from learned data and generating the reference pattern from this; and that changes in the environment lead to erroneous identifications because the reference pattern is fixed. These problems are significantly more severe in high-noise environments (such as a moving vehicle) than in the low-noise environments (the silence of a telephone line) described by Rabiner.

European patent application published under 01277 15A and US patent 4672669 describe an apparatus for detecting speech sounds in which a first check is made on the signal amplitude and a second check is based on the analysis of changes in the short-term signal spectrum. More specifically, the spectral analysis is performed by comparing the autocorrelation of the signal with that of an earlier part of the signal which is considered to be speech-free.

According to one aspect of the present invention, a device for detecting speech sounds is provided which

(i) a first speech detector (3-6, 14) operating by producing a measure of spectral similarity between a portion of an input signal and a portion of the input signal assumed to be free of speech, to produce an output signal indicative of the presence or absence of speech sounds in the input signal;

(ii) a memory (15) for storing data extracted from the non-speech part and

(iii) an auxiliary speech detector (20);

characterized in that the auxiliary speech detector (20) controls the updating of the memory (15), the auxiliary speech detector (20) operating by generating a measure of the spectral similarity between a current part of the input signal and an earlier part of the input signal.

According to a further aspect, the invention provides a device for detecting speech sounds, which

(i) means (1) for receiving an input signal;

(ii) a memory (15) for storing a noise representative signal representing an estimated noise component of the input signal;

(iii) means (3-6, 14) for periodically generating a measure of the spectral similarity between a portion of the input signal and the estimated noise component from the input signal and the signal representing noise;

(iv) means (7) for comparing the measured value with a threshold value to produce an output indicative of the presence or absence of speech sounds;

(v) an auxiliary speech detector (20) and

(vi) memory updating means for updating the memory based on the input signal;

characterized in that the auxiliary speech detector is operable to generate a control signal indicative of the presence or absence of speech sounds in response to a measurement of the spectral similarity between a current portion of the input signal and a preceding portion of the input signal, and in that the memory updating means is operable to that it updates the memory based on the input signal only when the control signal indicates the absence of speech sounds.

Further aspects of the present invention are as defined in the claims.

With reference to the accompanying drawings, some embodiments of the invention are described below by way of example. In the drawings:

Fig. 1 is a block diagram of a speech detector disclosed in our European Patent No. 335,521;

Fig. 2 shows a second speech detector disclosed in our European Patent No. 335 521 and

Fig. 3 shows an embodiment of the present invention.

The general principle underlying the first speech detector disclosed in our European Patent No. 335 521 is as follows.

A frame of n signal samples (s₀, s₁, s₂, s₅, s₄, ..., Sn-1) yields (if the samples from previous frames are ignored) after passing through a fictitious, finite, fourth-order digital impulse response filter (digital FIR filter) with an impulse response (1, h₀, h₁, h₂, h₃) a filtered signal for which the following applies:

s' =

(s0),

(s₁ + hos₀),

(s 2 + h 0 s 1 + h 1 s 0 ),

(s 5 + h 0 s 2 + h 1 s 1 + h 2 s 0 ),

(s 4 + h 0 s 3 + h 1 s 2 + h 2 s 1 + h 1 s 0 ),

(s5 + h0 s4 + h1 s3 + h2 s2 + h3 s4 ,

(s 6 + h 0 s 5 + h 1 s 4 + h 2 s 3 + h 3 s 2 ),

(p.7...)

The zero-order autocorrelation coefficient is the sum of the square of each term, which can be normalized, i.e. divided by the total number of terms (with constant frame lengths it is easier to dispense with division); thus the sum of the filtered signal is

R'₀ = (s'i)²,

and this is therefore a measurement of the power of the fictitious filtered signal s' - in other words, the part of the signal s that falls within the passband of the fictitious filter,

The extension is considered to be neglecting the first four terms

R'0 = (s 4 + h 0 s 3 + h 1 s 2 + h 2 s 1 + h 3 s 0 )²

+ (s5 + h0 s4 +h1 s3 + h2 s2 + h3 s1 )²

+ ...

= s&sub4;² + h&sub0;s&sub4;s&sub3; + h&sub1;s&sub4;s&sub2; + h&sub2;s&sub4;s&sub1; + h&sub3;s&sub4;s&sub0;

+ h0 s4 sa + h02 s32 + h0 h1 s3 s2 + h&sub0;h&sub2;s&sub3;s&sub1; + h&sub0;h&sub3;s&sub3;s&sub0;

+ h&sub1;s&sub4;s&sub2; + h&sub0;h&sub1;s&sub3;s&sub2; + h&sub1;²s&sub2;² + h&sub1;h&sub2;s&sub2;s&sub1; + h&sub1;h&sub3;s&sub2;s&sub0;

+ h&sub2;s&sub4;s&sub1; + h&sub0;h&sub2;s&sub3;s&sub1; + h&sub1;h&sub2;s&sub1; + h&sub2;²s&sub1;² + h&sub2;h&sub3;s&sub1;s&sub0;

+ h&sub3;s&sub4;s&sub0; + h&sub0;h&sub3;s&sub3;s&sub0; + h&sub1;h&sub3;s&sub2;s&sub0; + h&sub2;h&sub3;s&sub1;s&sub0; + hs&sub2;s&sub0;&sub2;

+ ...

= R0 (1 + h0² + h12 + h22 + h31

+ R&sub1; (2h0 + 2h0 h1 + 2h1 h2 + 2h2 h3 )

+ R&sub2; (2h1 + 2h1 h3 + 2h0 h2 )

+ R&sub3; (2h2 + 2h0 h3 )

+ R4; (2h3)

Therefore, R'0 can be determined from a combination of the autocorrelation coefficients Ri weighted by the constants in parentheses that determine the frequency band to which the value of R'0 responds. In fact, the terms in parentheses are the autocorrelation coefficients of the impulse response of the fictitious filter, so that the above expression can be reduced to

R'0 = R0 H0 + 2 RiHi ... (1)

where N is the filter order and Hi is the (unnormalized) autocorrelation coefficients of the impulse response of the filter.

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

Therefore, 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.

The filtering operation can alternatively be viewed as a form of spectrum comparison, where the signal spectrum is compared with a reference spectrum (the inverse of the response of the fictitious filter). Since the fictitious filter in this application is chosen to approximate the inverse of the noise spectrum , this operation can be considered as a spectral comparison between speech and noise spectra, and the zeroth autocorrelation coefficient thus generated (i.e. the energy of the inverted, filtered signal) as a measure of the dissimilarity between the spectra. The Itakura-Saito distortion measure is used in an LPC to assess the agreement between the predictor filter and the input spectrum, and is given in a form by

M = R�0;A�0; + 2 RiAi

where Aa, etc. are the autocorrelation coefficients of the LPC parameter set. It can be seen that this is very similar to the relationship derived above, and if it is remembered that the LPC coefficients are the taps of an FIR filter with the inverted spectral response of the input signal, so that the LPC coefficient set is the impulse response of the inverse LPC filter, it is obvious that the Itakura-Saito distortion measure is actually just a form of equation 1, where the response H of the filter is the inverse of the spectrum shape of an all-pole model of the input signal.

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

The I-S distortion measure is discussed in more 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, No. 5, October 1980.

Since the frames of a signal have only a finite length and a number of terms (N, where N is the order of the filter) are neglected, the above result is only an approximation; however, it provides a surprisingly good indicator of the presence or absence of speech and can thus be used as a measurement M in speech detection. In an environment where the noise spectrum is well known and stationary, it is quite possible to simply use fixed coefficients h₀, h₁, etc. to form the inverse noise filter.

However, a device that can adapt to environments with different background noises has a wider range of uses.

Referring to Fig. 1, in the first speech detector described in our European Patent No. 335,521, a signal from a microphone (not shown) is received at an input 1 and converted into digital samples s by an analog-to-digital converter at a suitable sampling rate. An LPC analysis unit 3 (of a known type of LPC encoder) then derives, for successive frames of n (e.g. 160) samples, a set of N (e.g. 8 or 12) LPC filter coefficients Li which are transmitted to represent the input speech. The speech signal s also enters a correlator unit 4 (which is normally part of the LPC encoder 3, since the autocorrelation vector Ri of the speech is normally also generated as a step of the LPC analysis, although of course a separate correlator could be provided). The correlator 4 generates the autocorrelation vector Ri including the zeroth order correlation coefficient R0 and at least two further autocorrelation coefficients R₁, R₂, R₅. These are then fed to a multiplication unit 5.

A second input 11 is connected to a second microphone which is located away from the loudspeaker so that it receives only the background noise. The input from this microphone is converted into a chain of digital input samples by an AD converter 12 and subjected to LPC analysis by a second LPC analysis device 13. The LPC "noise" coefficients produced by the analysis means 13 are fed to the correlator unit 14, the autocorrelation vector thus produced is multiplied term by term by the autocorrelation coefficients Ri of the input signal from the speech microphone by the multiplication means 5, and the weighted coefficients thus produced are combined in an adder means 6 according to equation 1 to apply a filter having the inverse shape of the noise spectrum from the microphone transmitting only the noise (which in practice corresponds to the shape of the noise spectrum of the microphone transmitting both the signal and the noise) and thereby filter out most of the noise. The resulting measurement value M is compared with a threshold value by a threshold device to produce a logic output 8 indicating the presence or absence of speech; if M is high, speech is assumed to be present.

However, the speech detector requires two microphones and two LPC analyzers, which makes the required device more expensive and complex.

In an alternative implementation of the first speech detector, a corresponding one is generated using the autocorrelations from the noise microphone 11 and the LPC coefficients from the main microphone 1 generated measurement value is used, which means that instead of an LPC analysis device an additional autocorrelator would be required.

Both implementations of the first speech detector are therefore suitable for working in different environments with background noise at different frequencies or in a given environment with a changing noise spectrum.

Referring to Figure 2, in a second speech detector disclosed in our European Patent No. 335,521, a buffer 15 is provided in which a set of LPC coefficients (or the autocorrelation vector of the set) derived from the microphone input 1 in a "speech-free" (i.e., one containing only noise) period is stored. These coefficients are then used to derive a measurement using Equation 1 which, of course, also corresponds to the Itakura-Seito distortion measurement, with the exception that instead of the present frame of LPC coefficients, a single, stored frame of LPC coefficients corresponding to an approximation of the inverse of the noise spectrum is used.

The LPC coefficient vector Li output by the analysis device 3 is also passed to a correlator 14 which generates an autocorrelation vector of the LPC coefficient vector. The buffer memory 15 is controlled by the speech/no-speech indicating output of the threshold device 7, for example by means of a multiple switch 16 via which the outputs of the correlator 14 carrying each autocorrelation coefficient are connected to the buffer 15, so that the buffer stores the "noise" signal during the speech-containing frame. autocorrelation coefficients, but during the "noise" frame a new set of LPC coefficients can be used to update the buffer. It is obvious that the correlator 14 could be placed after the buffer 15. Furthermore, the decision as to whether or not speech is present for the coefficient update need not be made from the output 8, but could (and preferably is) derived from elsewhere.

Since periods of no speech often occur, the LPC coefficients stored in the buffer are updated from time to time, thereby making the device suitable for tracking changes in the noise spectrum. It is obvious that such an update of the buffer may only be necessary occasionally, or only once at the start of operation of the detector if the noise spectrum is relatively stationary over time (as is often the case), but in a cellular environment frequent updating is preferable.

In a modification of the second speech detector, the system initially uses Equation 1 with a simple, fixed high-pass filter corresponding coefficient terms and then begins to adapt by switching to the use of LPC "noise period" coefficients. If speech recognition fails for any reason, the system can revert to using the simple high-pass filter.

It is possible to normalize the measurements described above by dividing by 1%, so that the expression to be compared with the threshold takes the following form:

This measurement is independent of the total signal energy in a frame and is therefore compensated for significant signal level changes, but results in a rather less pronounced contrast between "noise" and "speech" levels and is therefore preferably not used in environments with strong background noise.

Instead of using LPC analysis to derive the inverse filter coefficients of the noise signal (either from the noise microphone or from periods containing only the noise, as in the first and second speech detectors described above), it is possible to model the inverse noise spectrum using an adaptive filter of a known type; since the noise spectrum changes only slowly (as discussed below), the relatively slow coefficient adaptation speed typical of such filters is acceptable. In an implementation corresponding to Fig. 1, the LPC analysis unit 13 is simply replaced by an adaptive filter (for example a transversal FIR or bridge filter) connected to lighten the input noise by modeling the inverse filter, and its coefficients are fed to the autocorrelator 14 as before.

In a second implementation, corresponding to that of Fig. 2, the LPC analysis device 3 is replaced by such an adaptive filter and a buffer device 15 is omitted, but the switch 16 functions to prevent adaptation of the coefficients by the adaptive filter during periods of speech.

A speech detector according to an embodiment of the present invention will be described below.

From the above, it is clear that the LPC coefficient vector is simply the impulse response of an FIR filter with a response approximating the inverted spectral shape of the input signal. When the Itakura-Saito distortion measurement is produced between successive frames, it actually corresponds to the power of the signal filtered by the LPC filter of the previous frame. Therefore, if the spectra of successive frames differ little, a correspondingly small amount of the spectral power of a frame escapes filtering and the measurement is low. Accordingly, a large difference between the spectra of the frames produces a high Itakura-Saito distortion measurement, so that the measurement reflects the spectral similarity between successive frames. In a speech encoder, it is desirable to minimize the data rate, so the frame length is set as long as possible; in other words, a speech signal should show a significant spectral change from frame to frame if the frame length is long enough (if this is not the case, the coding is redundant). The noise, on the other hand, has a spectral shape that changes only slowly from frame to frame, and so the Itakura-Saito distortion measure is correspondingly low during a period when speech is absent from the signal, since the application of the inverse LPC filter of the previous frame "filters out" most of the noise energy.

Typically, for a highly noisy signal containing intermittent speech, the Itakura-Saito distortion measure between consecutive frames is higher during periods of speech than during periods of noise only; the degree of change is (as indicated by the standard deviation) is also higher and changes with fewer interruptions.

It should be noted that the standard deviation of the standard deviation of M is also a reliable measure; the effect of using any standard deviation is essentially to smooth the measurement.

In this embodiment, the measured parameter used to determine whether speech is present is preferably the standard deviation of the Itakura-Saito distortion measure, but other measures of deviation and other spectral distortion measures (e.g. based on FFT analysis) may also be used.

It has been found advantageous to use an adaptive threshold when detecting speech sounds. Such thresholds must not be adjusted during periods containing speech or the speech signal will be removed by the threshold. Accordingly, it is necessary to control the threshold adjusting means using a speech/no speech indicating control signal; and preferably this control signal is independent of the output of the threshold adjusting means.

The threshold T is adjusted adaptively to keep the threshold level just above the level of the measured value M when noise is present. Since the measured value generally fluctuates randomly when background noise is present, the threshold is varied by determining an average level over a series of blocks and adjusting the threshold to a level proportional to the average. However, in a noisy environment this is usually not sufficient and Therefore, a determination of the degree of change of the parameter over several blocks is also taken into account.

The threshold value T is therefore preferably determined according to

T = M' + K d

where M' is the average of the measurements over several consecutive frames, d is the standard deviation of the measurements over these frames and K is a constant (which is typically 2).

In practice, adaptation is preferably not resumed immediately after the absence of speech is detected, but rather one waits to ensure that the decay is stable (to prevent repeated, rapid switching between the state in which adaptation occurs and the state in which no adaptation occurs).

Referring to Fig. 3, in a preferred embodiment of the invention incorporating the aspects discussed above, an input 1 receives a signal which is sampled and digitized by an analog-to-digital converter (ADC) 2 and fed to the input of an inverse filter device 3 which in practice is part of a speech coding device intended to process the speech sounds and produces (typically eight) coefficients Li of a filter corresponding to the inverse of the input signal spectrum. The digitized signal is also fed to an autocorrelator 4 (which is part of the analysis device 3) which produces the autocorrelation vector Ri of the input signal (or at least as many low order terms as there are LPC coefficients). The operation of these parts of the device is as described in Figs. 1 and 2. Preferably, an average value of the Autocorrelation coefficients Ri are formed to improve their reliability. This can be achieved by storing each set of autocorrelation coefficients output by the autocorrelator 4 in a buffer 4a and using an average means 4b to generate a weighted sum of the current autocorrelation coefficients R1 and the autocorrelation coefficients of the previous frames stored in and supplied from the buffer 4a. The average autocorrelation coefficients Rai thus derived are supplied to a weighting and adding means 5, 6 which also receives from an autocorrelator 14 via a buffer 15 the autocorrelation vector Ai of the stored inverse filter coefficients Li for a period in which noise is present and generates measured values from Rai and Ai which are preferably defined as follows:

M = A₀ + Σ

This measurement is then compared with a threshold level by the threshold device 7, and the logical result provides an indication of the presence or absence of speech at the output 8.

In order for the inverse filter coefficients Li to correspond to an approximate estimate of the noise spectrum, it is desirable to update these coefficients during the noise period (and of course not to update them during periods of speech). Preferably, the decision whether or not speech is present, on which the update is based, does not depend on the result of the update, otherwise a single wrongly identified frame of a signal could cause the speech detector to subsequently "lose sight"and incorrectly identifies the following frames. Therefore, a control signal generating circuit 20, effectively a separate speech detector, is preferably provided which generates an independent control signal indicating the presence or absence of speech to the inverse control filter analysis device 3 (or the buffer 8), so that the autocorrelation coefficients Ai of the inverse filter used to generate the measured value M are only updated during the periods in which only noise is contained. The control signal generating circuit 20 comprises an LPC analysis device 21 (which in turn may be part of the speech coding device and in particular may be implemented by the analysis device 3) which generates a set of LPC coefficients Mi corresponding to the input signal, and an autocorrelator 21a (which may be implemented by an autocorrelator 3a) which derives the autocorrelation coefficients Bi from M;. When the analysis device 21 is implemented by the analysis device 3, Mi = Li and Bi = Ai. These autocorrelation coefficients are then fed to the weighting and adding means 22, 23 (equivalent to 5, 6) which also receives the autocorrelation vector Ri of the input signal from the autocorrelator 4. Thus, a measure of the spectral similarity between the input speech frame and the previous speech frame is calculated; this may be the Itakura-Saito distortion measure between Ri of the current frame and Bi of the previous frame as disclosed above, or may instead be derived by calculating the Itakura-Saito distortion measure for Ri and Bi of the current frame and subtracting the corresponding measure for the previous frame stored in the buffer 24 (in the subtraction means 25) to produce a spectral difference signal (in both cases the measure is preferably energy normalized by division by R₀). The buffer 24 is then, of course, updated. This spectral difference signal, after being thresholded by a thresholding means 26, as discussed above, is an indicator of the presence or absence of speech. However, we have found that this measure, although excellent at distinguishing between noise and unvoiced speech (a task for which prior art systems are generally inadequate), is generally less suitable for distinguishing between background noise and voiced speech. Accordingly, the circuit 20 preferably further includes a voiced speech detection circuit comprising a pitch analyzer 27 (which in practice may function as part of the speech encoder and in particular may measure the long term prediction delay value produced by a multi-pulse LPC encoder). The pitch analyzer 27 produces a logic signal which is "true" when voiced speech is detected, which signal, together with the measurement determined by the thresholder 26 by comparison with the threshold (which is generally "true" when unvoiced speech is present) is fed to the inputs of a NOR gate 28 to produce a signal which is "false" when speech is present and "true" when noise is present. This signal is fed to the buffer 8 (or the inverse filter analyzer 3) so that the inverse filter coefficients Li are updated only during the periods when only noise is present.

The threshold adjustment device 29 is also connected to provide the control output of the control signal generating circuit 20 in the form of a signal indicating the absence of speech. signal. The output of the threshold adjuster 29 is fed to the threshold adjuster 7. The threshold adjuster operates by counting the threshold up or down in steps proportional to the current threshold until the threshold approaches the noise energy level (which may conveniently be derived from the weighting and summing circuits 22, 23, for example). If the input signal is very weak, it may be desirable to automatically set the threshold to a fixed, low level, since the signal quantization effect produced by the ADC 2 may give unreliable results at low signal levels.

A "hangover generator" 30 may also be provided which measures the duration of indications of speech after the threshold means 7 and if the presence of speech has been indicated for a period exceeding a predetermined time constant, the output is held high for a short "hangover" period. In this way, clipping of the center of low level speech sounds is prevented and appropriate selection of the time constant prevents activation of the hangover generator 30 by short noise peaks which are falsely indicated as speech.

It is of course obvious that all the functions described above can be performed by a single, suitably programmed digital processing device, for example a digital signal processing chip (DSP chip) as part of an LPC codec implemented in this way (this is the preferred implementation) or by a suitably programmed microcomputer or microcontroller chip with an associated storage device.

Conveniently, the speech detection device can be implemented as part of an LPC codec as described above. Alternatively, voice detection can be performed remotely from the codec when the signal's autocorrelation coefficients or associated measurements (partial correlation or "parcor" coefficients) are transmitted to a remote station.

Claims (4)

1. Device for recording speech sounds with (i) a first speech detector (3-6, 14) operating by producing a measure of the spectral similarity between a portion of an input signal and a portion of the input signal assumed to be free of speech to produce an output signal indicative of the presence or absence of speech sounds in the input signal; (ii) a memory (15) for storing data extracted from the non-speech part and (iii) an auxiliary speech detector (20); characterized in that exclusively the auxiliary speech detector (20) controls the updating of the memory (15), whereby the auxiliary speech detector (20) operates by generating a measurement value of the spectral similarity between a current part of the input signal and an earlier part of the input signal. 2. Device for recording speech sounds with (i) means (1) for receiving an input signal; (ii) a memory (15) for storing a noise representative signal representing an estimated noise component of the input signal; (iii) means (3-6, 14) for periodically generating a measurement value of the spectral similarity between a portion of the input signal and the estimated noise component based on the input signal and the signal representing noise; (iv) means (7) for comparing the measured value with a threshold value to produce an output indicating the presence or absence of speech sounds; (v) an auxiliary speech detector (20) and (vi) a memory updating device for updating the memory based on the input signal; characterized in that the auxiliary speech detector is operable to generate a control signal indicative of the presence or absence of speech sounds, depending on a measure of the spectral similarity between a current portion of the input signal and a previous portion of the input signal, and in that the memory updating means is operable to update the memory based on the input signal only when the control signal indicates the absence of speech sounds. 3. The apparatus of claim 2, further comprising means for adjusting the threshold during periods in which the control signal indicates the absence of speech sounds. 4. Device according to claim 2 or 3, wherein the auxiliary speech detector further comprises a device (27) for detecting articulated speech sounds with a pitch analysis device for generating a signal indicating the presence of articulated speech and on which the control signal generated by the auxiliary speech detector (20) also depends.