GB2212962A - Noise compensation in speech recognition - Google Patents

Description

NOISE COMPENSATION IN SPEECH RECOGNITION The present invention relates to apparatus and methods for alleviating the effect of noise in sound recognition, particularly whole-word pattern matching in speech recognition using stochastic models of words.

Improved pattern matching has been achieved by using stochastic models of words instead of simple templates, assuming that speech can be approximated by hidden Markov processes (see "An Introduction to the Application of the Theory of Probabilistic Functions of a Markov Process to Automatic Speech Recognition" by Levinson, Rabiner and Sondhi, Bell Systems Technical Journal, Vol. 62, No. 4, April 1983, pages 1035 to 1074).

Briefly, incoming sounds are frequency analysed by, for example, a bank of filters and the resulting signal levels in each filter are smoothed to provide estimates of the short term power spectrum (called frames) typically every 10 ms. These signals are used after further processing together with a number of probability density functions (p.d,f.s) to give the probabilities that the ,oncoming signal producing the channel outputs corresponds to a state in a Markov model which is a finite state machine representing a word to be recognized. Each Markov model comprises a number of states and there is in general one p.d.f. for each channel relating to each state* the p.d.f.s being obtained by previously training a recognizer using examples of words to be recognized.In operation, the recognizer employs the Markov models to calculate the word most likely to have been uttered by considering the likelihoods that a current sound arises from each of the states and the probabilities of transition from one state to another within a Markov model. The Viterbi algorithm may be used in finding the most likely word on this basis.

In practice the negative of the logarithm of a likelihood is used and in this specification is referred to for brevity as "distance", by analogy with Dynamic Time Warping (DTW) recognizers.

The state p.d.f.s are often assumed to be multivariate normal distributions with diagonal covariance matrices, and so are characterised by a mean, m, and variance, s, for each filter-bank channel. This is a crude approximation to the speech signal that is currently in widespread use. The theory presented in this specification is equally applicable to stochastic models with p.d.f.s that are not multivariate normal distributions with diagonal covariance matrices.

In this specification the word "input" means the input to a speech recognizer during operational use, and "cell" means the level in a particular filter-bank channel or equivalent in a particular frame, in either input or a training. Filter-bank analysis is usually preferable for the present invention because methods of acoustic analysis that do not keep the different parts of the spectrum separate (e.g. Linear Predictive Coding or cepstrum methods) are not so amenable. to noise compensation. These other methods of acoustic analysis mix together noisy parts of the signal spectrum with components caused mainly by speech, and it is not then possible to identify which parts of the spectrum are contaminated by noise.

The background noise signal needs to be estimated as it varies with time. This can be done by using the microphone signal when no speech is present. It can also be done using a separate microphone which only gives the noise signal.

It has to be accepted that in conditions of high noise it is not possible in principle to distinguish between words that differ only in low-level regions of the spectrum, where they are seriously contaminated by noise. A technique is required which makes full use of any speech information in the high-level parts of the spectrum that can act as true evidence for word identity, but ignores any information that is too corrupted by noise to be useful.

When the speech in the training phase is completely uncontaminated by noise and the input cell, f, is above the input noise level, then in the case of a multivariate normal distribution with diagonal covariance matrix, the p.d.f. for each channel has the form:

The distance is therefore: -ln(N(f,m,s)) = lln(2N) + ln(s) + (fm)2 2S2 However the situation is very different when the input cell is known to be noisy. its actual value is unlikely to be sensibly related to the underlying signal, and may, in fact even be quite low because of chance cancellation of signal by the noise. It is therefore necessary to use a different method to derive a distance measure for noisy input cells.

According to a first aspect of the present invention there is provided apparatus for use in sound recognition comprising means for deriving a plurality of input signals during recognition which are each representative of signal levels in respective regions in the frequency spectrum, means for storing a plurality of groups of p.d.f. values representing probability density functions indicating the likelihoods that input signals arise from states in finite state machine models of groups of sounds to be recognized, means for estimating the input noise level, and means for recognizing sounds from the input signals, the stored p.d.f. values and the models, employing respective distance measures, each derived from one input signal and one p.d.f. as represented by one group of the said values, each distance measure representing a likelihood of obtaining a region signal level from one p.d.f., when the input signal is above a predetermined level related to the noise level in the corresponding spectrum region, and representing the cumulative likelihood of obtaining from the said p.d.f. a region signal level below the said predetermined level, when the input signal is below, or equal to, the predetermined level.

The groups of sounds are usually words where the apparatus is for speech recognition and the spectral regions are usually channels.

An advantage of the first aspect of the invention is that the input signals are used in obtaining the distance measures when they are reliable; that is when they are above the predetermined level which is usually near to, or at, the level. The predetermined level is used, instead of the input signals, when the input signals are unreliable. This is because the input signals are near to or below the noise level so there is no reliable information about the level of the underlying speech signal, except that it was not above the noise level. Using the cumulative distribution of the p.d.f. over all levels up to the noise level therefore gives a more reliable comparison between states than using the probability derived from the p.d.f. at the noise level.

The means for recognizing sounds may comprise means for deriving masked input signals during recognition by representing any channel of an input signal below noise level with a masking level which is representative of the noise level in that channel.

The means for estimating the input noise level may comprise a separate microphone recording the noise signal alone or means for differentiating between noise only and noise plus speech on a single microphone.

Several different distributions may be found useful in calculating likelihoods and cumulative likelihoods but the normal distribution is usually used in speech recognition. Assuming the normal distributions each likelihood measure is preferably derived from -ln [N(f,m,s)] when the noise level is below the input signal and from -ln!erf((A-m)/s)] when the noise level is above the input signal; where A is the noise level in the spectrum region corresponding to the input signal, the known cumulative distance function

and N(x,O,l) corresponds to a normally distributed p.d.f. with independent variable x, mean equal to zero and variance equal to one.

The invention also includes methods corresponding to the first aspect of the invention.

Another problem arises in deriving groups of values representing p.d.f.s where in training the sample utterances are somewhat contaminated by noise. This is particularly important in environments where the voice quality changes because of the noise, or where noise and voice quality are inseparable consequences of the environment. Examples are stress-induced voice changes in aircraft, particularly in an emergency, and shouting in high noise levels.

Any solution to this problem should also give useful improvements in less severe noisy environments.

If a large proportion of the measurements used to derive any one p.d.f. are corrupted by noise, there is no prospect of making reliable estimates of the underlying speech distribution. It is, however, important for any channel that such evidence as there is to suggest that different states have different underlying distributions should be taken into account in estimating the state parameters.

Therefore, according to a second aspect of the present invention there is provided a method of training a sound recognition system comprising deriving a plurality of groups of input signals from repetitions of nominally the same sound, each group being representative of signal levels in respective regions in the frequency spectrum, and deriving a plurality of groups of p.d.f. values representing probability density functions indicating the likelihoods that input signals arise from states in finite state machine models for a vocabulary of groups of sounds to be recognized, the p.d.f. values being derived only from input signals above the noise levels in corresponding spectrum regions, and the derivation being so carried out that the groups of values represent substantially whole probability functions although obtained from input signals above noise levels only.

Preferably the noise level used in each region of the frequency spectrum is the highest found in deriving the input signals for that region for all training repetitions of all sounds in the vocabulary.

If a normal distribution is assumed for each p.d.f. and the p.d.f.s are assumed to be uncorrellated and each group of values 2 comprises the true mean m and the true variance s , then m and 5 may be estimated from m = M.erf-1 (F)-B.Q on equation 1 erf-(F)-Q(F) s = B ~1m equation 2 erf-(F) where B is noise level, M is the mean of samples above the noise level, F is the proportion of input signals below the noise level, erf(F) is as defined above, and Q(F) = N(erf ~1(F),0,1).

In practice Q(F) and erf (F) can be found by look-up in a table of pre-computed values.

if more than half of the training cell measurements used for a state are identified as noisy, it is implied that the underlying mean is in fact below the noise level. It is then unwise to use only the tail of a distribution in an attempt to estimate the true mean and variance.

According to another feature of the invention therefore a constant mean and a constant variance are substituted for the said values in any said group representing a p.d.f. derived in training in which the proportion of input signals which are noise exceeds a predetermined value, greater than 0.5, and typically equal to 0.8.

Where the proportion is below the predetermined value equations 1 and 2 may be used, but preferably for a range of proportions between for example 0.5 and 0.8, smooth transition for mean and variance values is arranged without discontinuities by replacing B in equations 1 and 2 with a function dependent on B and F derived from a look-up table, and by appropriately modifying the tables for erf (F) and Q(F) in this range of F values.

It is also preferable to add a standard minimum variance to all computed variances to overcome the danger with all stochastic models of limited training leading to the computed variances being too low by chance. If the standard minimum variance is chosen to be very large, the variances for all states are in effect made equal and the distance measure reduces to the squared Euclidean distance measure that is widely used in DTW matching. If it is made too small, there is the danger of attaching too much significance to unreliable statistics that arise through inadequate training. It is therefore desirable to set the standard minimum variance by experiment to optimise the performance under any given practical conditions.

An advantage can be obtained, if in deriving the said groups of values the standard minimum variance for a particular p.d.f. is scaled by a function of the number of input signal samples used to derive said group of values for that p.d.f. since a variance derived from a large number of samples is more likely to represent the true speech variability than an equal variance derived from only a few samples.

The invention also includes apparatus corresponding to the methods of the second aspect of the invention.

Certain embodiments of the invention will now be described by way of example with reference to the accompanying drawings, in which: Figure 1 is a block diagram of apparatus employing the invention, Figures 2(a) and 2(b) show probability density functions under different input noise conditions during recognition, and Figure 3 illustrates the derivation of the true mean from a probability density function partially masked by noise in training.

In Figure 1 a microphone 10 receives speech to be recognized and after passing through a non-linear amplifier 11 to allow for a wide dynamic range of input signals the resulting signals reach an analogue to digital converter (A/D) 12. A digital filter bank 13 provides output signals representative of the signal levels of signals in various portions of the frequency spectrum and although only four outputs are shown there are, in practice, usually between ten and twenty. Signals from each of the bandpass filters are passed to a smoother 14 providing output signals typically every 10 ms. Each of these longer periods is known as a frame and each smoother output (corresponding to a spectrum portion) is known as a channel level, and for one frame is known as a cell.

Noise masking is now carried out in a masker circuit 15 where the input noise spectrum is analysed by the components 10 to 14 and a noise estimator 20 in intervals in speech. The circuit 15 emits two signals for each channel, one indicating whether the signal in a channel is noise and one representative of the signal level in that channel. Two connections 16 and 17 for these signals for one channel only are shown in Figure 1 but similar connections for other channels are also provided.

The two signals are passed to a Markov Recognizer 18 which is usually a microprocessor system or a system employing a signal processor integrated circuit. The system is programmed to carry out conventional recognition using hidden Markov models except that likelihoods (that is the distance measures) are determined according to the table below rather than always according to the first mathematical expression in the second column of the table.

Input Likelihood Above Noise -ln[N (f,m, s)] At or below noise -ln[erf((A-m)/s)] Since the invention is concerned with inputs below noise level, consider the situation where the input signal level is at or below the input noise level. The only information reliably known about the underlying speech is that it was below the noise mask. Conditions when the means of the Markov state p.d.f. are below and above the masked input cell level are illustrated in Figures 2(a) and 2(b) respectively, where m is the mean and A is the input noise level.

In Figure 2(a), in relation to the model, which assumes a normal distribution of speech levels but constant (in the short term) noise, the mean of the p.d.f. is below the input noise level. This means that the input speech level did not occupy the right-hand tail of the distribution, shown hatched. If the speech really did arise from this distribution, there is only a small probability of its giving a level above the mask and thus the observed data gives very little evidence against it being generated by the p.d.f. of Figure 2(a). The required small distance measure is given in this case by the negative logarithm of the area of the unhatched region, which is nearly equal to the total area under the p.d.f., i.e. unity. In the case of Figure 2(b) the converse applies. As there is only a small likelihood that this p.d.f.

would give a level in the unhatched region, the distance measure in this case is high. The cumulative distribution function of the p.d.f. is therefore used for a noise-masked input cell to estimate distance, that is for both the situations shown in Figures 2a and 2b, rather than using the p.d.f. directly. The difference in treatment of noise-marked and unmarked cells shown by the above table does not cause serious problems, because it is only the relative contributions to the distance for the various states that affect the recognition decisions, and for any one input cell all distance contributions are calculated in the same way either directly from the p.d.f. or from the cumulative distribution function of the p.d.f.

In the case of a normal distribution the cumulative distribution function is not available as a simple mathematical expression, but its values depend on only one parameter, (A - m)/s, and are available in published tables. Where necessary linear interpolation can be used between tabulated values.

Recognition by the recognizer 18 depends on p.d.f.s relating to the states of models held by a store 19. Each p.d.f. is 2 represented by a mean m and a variance s , the values m and s being supplied to the recognizer 18 to allow the likelihoods to be calculated according to the above table. At every frame a likelihood is calculated for each combination of an input cell and a p.d.f., and these likelihoods, transition probabilities between states of the model, and cumulative probabilities from previous frames are used to determine which sound is most likely to have been uttered in that frame. The process is known for likelihoods calculated in the conventional way, see for example the above mentioned paper by Levinson et al.

In training if less than half the measurements are noisy a reasonable estimate of the mean and variance can be obtained from the truncated distribution shown in Figure 3. Consider a normal 2 distribution of training cell values with mean m and variance s2, where a fraction, F, (F ( 0.5) of the observations has been removed from the left-hand tail as a result of using a noise mask level, B.

Given M (the mean of the remaining samples) and F, it can be shown that good estimators of m and s are given by equations 1 and 2 above. The apparatus of Figure 1 is used except that the microprocessor or processor circuit used as the Markov recognizer is programmed to carry out the algorithm given below and the store 19 stores the parameters of the p.d.f.s and the Markov models.

During training each word in the vocabulary to be recognized is repeated a number of times, typically five times, and the following algorithm is carried out for each cell in each frame: (a) Store the signal levels obtained for that cell in each frame at each repetition; (b) Determine the appropriate masking level for each channel (for example by detecting the maximum noise level occurring during training of the whole vocabulary); (c) Determine the proportion of stored signal levels which are above the appropriate masking level; (d)(1) If the proportion is greater than, or equal to, 0.5 estimate the true mean and variance from the mean and variance of those stored levels which are above the maximum noise level using equations 1 and 2; or (2) If the proportion is below, or equal to, 0.2 allocate a predetermined mean and variance; or (3) If the proportion is between 0.2 and 0.5 use equations 1 and 2 to determine the mean and variance except replace B with a function dependent on the maximum noise level and the proportion F and use appropriately modified functions instead of erf 1 (F) and Q(F), for example weighted sums of these functions and constants where the weighting varies as F changes between 0.2 and 0.5; (e) Scale a predetermined minimum variance by a function of the number of frames occurring while the current word is repeated; (f) Add the variance determined in (d) to the scaled predeter mined minimum variance determined in (e); and (g) Store the means determined in (d) and the variances determined in (f) in the store 19 to represent the p.d.f.s used in recognition.

The invention may be put into practice in many other ways than those specifically described. For example many other techniques can be employed to obtain cell outputs and some or all of the circuits 12 to 15 may be replaced by a signal processor or microprocessor. Other techniques than Markov modelling may be used for recognition and p.d.f.s of other distributions than the normal distribution may be employed. Other functions may be employed to obtain likelihoods and cumulative likelihoods during recognition and during training the true means and variances may be estimated according to other functions. Some or all of steps d(2), d(3), (e) and (f) may be omitted in obtaining values to represent p.d.f.s.

Claims (12)

1. Apparatus for use in sound recognition comprising means for deriving a plurality of input signals during recognition which are each representative of signal levels in respective regions in the frequency spectrum, means for storing a plurality of groups of p.d.f. values representing probability density functions indicating the likelihoods that input signals arise from states in finite state machine models of groups of sounds to be recognized, means for estimating the input noise level, and means for recognizing sounds from the input signals, the stored p.d.f. values and the models, employing respective distance measures, each derived from one input signal and one p.d.f. as represented by one group of the said values, each distance measure representing a likelihood of obtaining a region signal level from one p.d.f. when the input signal is above a predetermined level related to the noise level in the corresponding spectrum region, and representing the cumulative likelihood of obtaining from the said p.d.f. a region signal level below the said predetermined level, when the input signal is below, or equal to, the predetermined level.

2. A method for use in sound recognition comprising the steps of deriving a plurality of input signals during recognition which are each representative of signal levels in respective regions in the frequency spectrum, storing a plurality of groups of p.d.f. values representing probability density functions indicating the likelihoods that input signals arise from states in finite state machine models of groups of sounds to be recognized, estimating the input noise level, and recognizing sounds from the input signals, the stored p.d.f.

values and the models, employing respective distance measures, each derived from one input signal and one p.d.f. as represented by one group of the said values, each distance measure representing a likelihood of obtaining a region signal level from one p.d.f. when the input signal is above a predetermined level related to the noise level in the corresponding spectrum region, and representing the cumulative likelihood of obtaining from the said p.d.f. a region signal level below the said predetermined level, when the input signal is below, or equal to, the predetermined level.

3. Apparatus or method according to Claim 1 or 2 wherein the groups of sounds are words and the regions are channels defined by a bank of filters.

4. Apparatus according to Claim 1 or 3 wherein the means for recognizing sounds comprises means for deriving masked input signals during recognition by representing any channel of an input signal below noise level with a masking level which is representative of the noise level in that channel.

5. Apparatus according to Claim 1, 3 or 4 wherein the means for estimating the noise level includes a first microphone for sensing noise and the means for deriving input signals includes a second separate microphone for sensing noise plus sounds to be recognised.

6. Apparatus according to Claim 1, 3 or 4 wherein the means for deriving input signals and the means for estimating noise level include a common microphone, and means for differentiating between noise only and noise plus sounds to be recognized.

7. Apparatus or method according to any preceding claim wherein 2 each said distance measure is derived from tNn(2n) + in(s) + (f-m) 2s2 when the noise level is below the input signal and from -ln[erf((A-m)/s)) when the noise level is above the input signal; where A is the noise level in the spectrum region corresponding to the input signal, f is the input level in a region due to sounds to be recognized, and m and s are a mean and a variance representing the p.d.f. from which that distance measure is derived.

8. A method of training a sound recognition system comprising deriving a plurality of groups of input signals from repetitions of nominally the same sound, each group being representative of signal levels in respective regions in the frequency spectrum, and deriving a plurality of groups of p.d.f. values representing probability density functions indicating the likelihoods that input signals arise from states in finite state machine models for a vocabulary of groups of sounds to be recognized, the p.d.f. values being derived only from input signals above the noise levels in corresponding spectrum regions, and the derivation being so carried out that the groups of values represent substantially whole probability functions although obtained from input signals above noise levels only.

9. A method according to Claim 8 wherein the noise level used in each region of the frequency spectrum is the highest found in deriving the input signals for that region for all training repetitions of all sounds in the vocabulary.

10. A method according to Claim B or 9 wherein each p.d.f. is assumed to have a normal distribution and each group of p.d.f.

values comprises the true mean m and the true variance s2 of the distribution.

11. A method according to Claim 10 wherein m and s are determined from M.erf1 (F)-B.O(F) m = -l erf (F)-Q(F) B-m erf (F) where B is noise level, M is the mean of samples above the noise level, F is the proportion of input signals below the noise level, erf(F) is as defined above, and -l Q(F) = N(erf (F),O,1).

12. A method according to Claim 2 wherein each probability density function is assumed to have a normal distribution and each group of said values comprises an estimated true mean and a modified variance which is the sum of an estimated true variance and a predetermined minimum variance scaled by a value dependent on the number of input signals used to calculate the estimated true variance.

12. A method according to Claim 10 or 11 wherein a constant mean and a constant variance are substituted for the said values in any said group representing a p.d.f. derived in training in which the proportion of input signals which are noise exceeds a predetermined value greater than 0.5.

13. A method according to Claim 12 wherein the predetermined value is 0.8.

14. A method according to Claim 12 wherein if the said proportion is between the predetermined value and a lower further predetermined value then a smooth transition of values is arranged over the range between the two predetermined values.

15. A method according to any of Claims 10 to 14 wherein a standard minimum variance is added to all calculated variances, the standard minimum variance having a value such that the variances for different states differ significantly.

16. A method according to any of Claims 10 to 16 wherein in deriving the said groups of values, the standard minimum variance for a particular p.d.f. is scaled by a function of the number of input signal samples used to derive the said group of values for that p.d.f.

17. Apparatus for use in sound recognition substantially as hereinbefore described with reference to Figure 1 of the accompanying drawings.

18. A method for use in sound recognition substantially as hereinbefore described.

19. A method of training sound recognition apparatus substantially as hereinbefore described.

Amendments to the claims have been filed as follows 1. A method of training a sound recognition system comprising: deriving a plurality of groups of input signals from repetitions of nominally a same sound, each said group being representative of a signal level in a corresponding region in a frequency spectrum in which frequency components appear when sounds are to be recognized; estimating noise levels in each of said regions of the frequency spectrum; and deriving a plurality of groups of values representing respective probability density functions indicating likelihoods that input signals arise from states in finite state machine models for a vocabulary of groups of sounds to be recognized; values in each said group of values being derived only from those of said input signals in one of said groups of input signals which represent signal levels above the estimated noise level in said corresponding region, and the derivation being so carried out that the groups of values represent substantially whole probability functions although obtained from input signals above noise levels only.

2. A method according to Claim 1 wherein the groups of said input signals are derived for each sound in a vocabulary of sounds, estimating noise level includes the step of finding a highest estimated noise level for each region for all repetitions of all sounds in said vocabulary, and deriving groups of input signals includes, for each region, choosing the estimated noise level used to determine which input signals are used in deriving the said values, as said highest estimated noise level, for that region.

3. A method according to Claim 1 or 2 wherein each said probability density function (p.d.f.) is assumed to have a normal distribution and each group of values representative of a p.d.f.

comprises an estimated true mean m and an estimated variance s2 of the distribution.

4. A method according to Claim 3 wherein for each region and each said p.d.f., m and s2 are determined from m = M.erf1(F)-B.O(F) erf1 (F)-Q(F) s=B - m erf1 (F) where B is noise level in said each region, M is the mean of samples derived from input signals in that region above the noise level B, F is the proportion of input signals below the noise level B, erf (F) is as defined above, and Q(F) = N(erf-1F),O,1), where

and N(x,0,1) corresponds to a normally distributed p.d.f. with independent variable x, mean equal to zero and variance equal to one.

5. A method according to Claim 3 or 4 wherein a derived mean and a derived variance are substituted for the estimated true mean and the estimated true variance in all said groups of said values derived in training from every said region in which the proportion of input signals which are below said estimated highest noise level for that region exceeds a predetermined value greater than 0.5.

6. A method according to Claim 5 wherein the predetermined value is 0.8.

7 A method according to Claim 5 wherein the derived mean and the derived variance have predetermined fixed values.

8. A method according to Claim 5 wherein for every said region in which said proportion is between the predetermined value and a lower further predetermined value, a substitute mean and a substitute variance are substituted for the estimate true mean and the estimated true variance, over the range between the two predetermined values, the substitute mean and the substitute variance being taken from a smooth transition from said derived mean to a predetermined fixed mean and from said derived variance to a predetermined fixed variance, respectively, according to position in said range.

9. A method according to any of Claims 5 to 8 wherein in order to derive the derived variance a standard minimum variance is added to all true variances, the standard minimum variance having a value which is small enough to ensure that the derived variances for different states differ significantly where true variances are significantly different.

10. A method according to any of Claims S to 8 wherein in deriving said derived variance, a scaled minimum variance is added to each true variance, the scaled minimum variance for a particular p.d.f.

being derived from the number of input signal samples used to derive the said group of values for that p.d.f.

11. A method according to Claim 2 wherein each probability density function is assumed to have a normal distribution and each group of said values comprises an estimated true mean and a modified variance which is the sum of an estimated true variance and a predetermined minimum variance scaled by a fixed value.