The present invention generally relates to the multi-channel separation of a desired signal and of an interfering signal, by using a reference signal. More particularly, the present invention relates to the separation of a speech signal (the desired signal) and of an interfering signal, in the context of speech recognition applications, in the case where a mixture of the speech and of the interfering signal is recorded in one channel, and where the interfering signal is recorded in a second channel (the reference signal).
Robustness in the presence of noise and, more generally, of interfering signals is a crucial issue normally addressed in connection with speech recognition, especially when performance in a real-world environment is concerned. In cases where the signal interfering with the speech is stationary and where its characteristics are known in advance, robustness issues can, to a certain extent, be addressed during the training of the system. Particularly, the acoustic model of the speech recognition system can be trained on a representative collection of noisy data; this approach is known as “multi-style training” and has been shown to reduce the degradation of the recognition accuracy in the presence of noise.
However, in most applications, the signal corrupting the speech is neither known in advance nor stationary (for example, music or speech from competing speakers). Such cases typically cannot be handled by devising special training schemes, and they tend to require the use of on-line adaptive algorithms.
Particular needs have been recognized in connection with addressing the problem of separating a speech signal and an interfering signal (e.g. non stationary noise, music, competing speech) in the case where a recording of the interfering signal is available in a second channel. The signal contained in this second channel is called the reference signal. This occurs in a variety of contexts, such as:
To date, various efforts have been made in the contexts just described, yet various shortcomings and disadvantages have been observed.
Conventionally, the problem of separating a desired signal and an interfering signal with a known reference signal is often addressed by using decorrelation filtering techniques (see Ehud Weinstein, Meir Feder and Alan V. Oppenheim, “Multi-channel signal separation by decorrelation”, IEEE transactions on Speech and Audio Processing, volume 1, number 4, October 1993). The model underlying the conventional decorrelation filtering approach is illustrated in
The transfer function within each channel (from s1 to o1, and from s2 to o2) is assumed to be an identity system. Besides, it is assumed that there is no leakage of the desired signal s1 into the reference sensor, i.e., the cross-coupling function from the input channel of s1 to the output channel of o2 is zero. Under theses assumptions, the linear system reduces to the cross-coupling between the input channel of s2 and the output channel of o1. In decorrelation filtering techniques, the linear system is estimated with an iterative algorithm so that, by performing inverse filtering, the reconstructed signals s1 and s2 in the input channels are statistically uncorrelated. It can be shown that under the above assumptions, the linear system can be identified unambiguously. Once the linear system is identified, it is used to cancel the interfering signal component in the observed mixture.
The decorrelation filtering approach does suffer from some limitations in the context of a speech recognition application, such as:
Another conventional approach, the Codeword-Dependent Cepstral Normalization (CDCN) approach, is a mono-channel technique which is used during speech recognition to compensate for the combined effect of stationary noise and channel mismatch. (See Alejandro Acero, “Acoustical and Environmental Robustness in Automatic Speech Recognition”, PhD thesis, Department of Electrical and Computer Engineering, Carnegie Mellon University, Pittsburgh, Pa. 15213, September 1990.) CDCN does not operate in the waveform domain but, instead, in the cepstral domain, which is the domain where speech recognition is usually performed. A cepstra (see chapter 3 in L. Rabiner and B. H. Juang, Fundamentals of Speech Recognition, Prentice Hall Signal Processing Series, 1993) is a vector that is computed by the front end of a speech recognition system from the log-spectrum of a segment of speech waveform samples (usually this segment is about 100 ms long). The stream of cepstra corresponding to a speech utterance is typically computed from successive overlapping segments of the speech waveform samples (usually the shift between two adjacent segments is about 10 ms). In the CDCN framework, the cepstra of the noise is estimated by minimizing the difference between the cepstral space of the current utterance and the cepstral space of the clean speech (“clean speech” meaning non-noisy speech) characterized by a codebook of cepstral vectors. As the sources of mismatch are assumed to be stationary, the estimation is performed by averaging over the whole utterance.
Among the limitations of the mono-channel CDCN approach, though, is that non-stationary noise is not taken into account as accurately and effectively as may be possible. Particularly, a fundamental assumption of the mono-channel CDCN approach is that the noise is relatively stationary over periods of at least one or even a few seconds. The shorter the period during which the noise can be considered stationary, the more poorly conventional CDCN will perform. In the case of highly non-stationary noises, such as music, the mono-channel CDCN framework may even degrade the speech recognition accuracy instead of improving it.
Also included among conventional techniques are two-channel compensation techniques that operate in the cepstral domain (see Acero, supra). Such techniques can be characterized as follows:
Accordingly, similar disadvantages are encountered as in the case of the other conventional techniques described.
Consequently, and in brief recapitulation, various needs have been recognized in connection with overcoming the shortcomings and disadvantages observed in connection with conventional techniques.
In accordance with at least one presently preferred embodiment of the present invention, it is proposed that the effect of the interfering signal in the cepstral domain be compensated for.
Assuming that y1 is the cepstra of the desirable signal (i.e. the clean speech) and that x2 is the cepstra of the interfering signal, it is well known in the field of speech processing that the cepstra x1 of the observed mixed signal can be computed as x1=y1+f(y1, x2), where “f” is a known non linear function of both the cepstra of the clean speech and of the interfering signal. In at least one embodiment of the present invention, the cepstra x2 of the interfering signal is computed from the reference signal. For lack of knowing the cepstra y1 of the clean speech, the compensation term f(y1, x2) is preferably estimated by following the Codeword Dependent Cepstral Normalization (CDCN) approach described heretofore, such that f(y1, x2) is approximated with its expected value over y1, where y1 takes its values in a predefined codebook of cepstral vectors characterizing the cepstral space of the clean speech.
In accordance with at least one presently preferred embodiment of the present invention, the following advantages are presented in comparison with the conventional decorrelation filtering approach described further above:
Furthermore, since operation in the cepstral domain is contemplated, at least one embodiment of the present invention makes it especially appropriate for speech recognition applications that use this front end.
In comparison with CDCN as described heretofore, at least one presently preferred embodiment of the present invention involves a two-channel approach, where the cepstra of the noise is computed on a frame basis from the reference signal in the second channel. Therefore, unlike the CDCN approach, at least one embodiment of the present invention aims at compensating for the effect of both stationary and non stationary noise.
In comparison with conventional two-channel compensation techniques operating in the cepstral domain as described heretofore, a two-channel compensation technique in accordance with at least one presently preferred embodiment of the present invention has the following characteristics:
In one aspect, the present invention provides an apparatus for compensating for interference in speech recognition, the apparatus comprising: a first input medium which obtains an initial speech signal; a second input medium which obtains at least one interfering signal; a normalizing arrangement which reconciles the initial speech signal and at least one interfering signal with one another to produce a final speech signal; and the normalizing arrangement being adapted to account for non-stationary noise in the at least one interfering signal.
In another aspect, the present invention provides a method of compensating for interference in speech recognition, the method comprising the steps of: obtaining an initial speech signal; obtaining at least one interfering signal; and reconciling the initial speech signal and at least one interfering signal with one another to produce a final speech signal; the reconciling step comprising the step of accounting for non-stationary noise in the at least one interfering signal.
Furthermore, in an additional aspect, the present invention provides a program storage device readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for compensating for interference in speech recognition, the method comprising the steps of: obtaining an initial speech signal; obtaining at least one interfering signal; and reconciling the initial speech signal and at least one interfering signal with one another to produce a final speech signal; the reconciling step comprising the step of accounting for non-stationary noise in the at least one interfering signal.
For a better understanding of the present invention, together with other and further features and advantages thereof, reference is made to the following description, taken in conjunction with the accompanying drawings, and the scope of the invention will be pointed out in the appended claims.
y1(t)=x1(t)−f(y1(t),x2(t)),
For lack of knowing the cepstra y1 of the clean speech, the principle of the CDCN approach is to approximate the compensation term f(y1, x2) with its expected value over y1, given x1 and x2. The calculation of the expectation is carried out by assuming that y1 takes its values in a predefined codebook C consisting of cepstral codeword vectors {ci} characterizing the acoustic space of the clean speech:
Ey1{f(y1,x2)|x1,x2,C}=Σip(ci|x1,x2)f(ci,x2)
And the cepstra of the desired speech signal is estimated as:
y1(t)=x1(t)−Ey1{f(y1,x2)|x1,x2,C}
Note that in the conventional mono-channel CDCN scheme, the interfering signal is assumed to be stationary and its cepstra x2 is estimated either by averaging non-speech frames, or by applying an EM algorithm over all the frames of each utterance. On the other hand, in the two-channel CDCN scheme according to at least one embodiment of the present invention, the assumption of stationarity is removed and the cepstra x2 of the interfering signal is estimated as the cepstra of the reference signal in the second channel.
The disclosure now turns to a general overview of a procedure that may be employed in accordance with at least one embodiment of the present invention.
As far as off line requirements go, one may preferably build a codebook C of cepstral codeword vectors characterizing the cepstral space of the desired speech signal; this can be done by vector-quantizing a set of training utterances that have been pronounced in a non-noisy environment.
For pre-processing of the waveforms, one may preferably:
For feature extraction, one may preferably compute standard cepstral vectors x1 and x2 from the time-aligned and scaled input waveforms s1 and s2.
For two-channel CDCN compensation, for each input pair of cepstral vectors x1 and x2, one may preferably:
Preferably, one may then decode the speech from the estimated cepstra y1. Elaborating now on a preferred method of employing at least one embodiment of the invention, as pointed out heretofore, the two-channel CDCN scheme requires that the two input waveforms be previously time-aligned and scaled. In experiments that have been performed, time-alignment was performed by detecting the maximum of the cross-correlation function between the two waveforms. The scaling factor was estimated as the ratio between the means computed in each channel over non speech segments of about 450 ms (10,000 samples at 22 kHz).
As will be seen herebelow, experiments were conducted with codebooks of different sizes. As far as the test data are concerned, a codebook of 64 codewords appears to correspond to an optimal trade-off between recognition accuracy and complexity. Smaller codebooks (down to 8 codewords), leading to further reduced computation requirements, still provide significant improvement of the recognition accuracy over the baseline recognition scores.
The disclosure now turns to an evaluation of an embodiment of the invention as employed in the context of speech recognition in a car, where the signal interfering with the speech is the music played by the CD player of the car. The evaluation protocol is first explained, and then the recognition scores, obtained before and after the two-channel CDCN scheme according to an embodiment of the invention is applied, are compared.
The test data in the experiments included stereo recordings in a non-moving car, where a male subject is asked to utter digit strings while the CD player of the car is playing rock'n'roll music. The first channel comes from an AKG Q400 microphone mounted on the visor of the car, and recording the voice of the subject and the music from the car speakers. The second channel contains the music recorded directly from the output of the CD player. All the data are recorded at 22 kHz and downsampled to 11 kHz.
The test data include four subsets, each of which corresponds proportionally to an increased volume of music: 3, 5, 7 and 10 (10 corresponds to the loudest volume of the CD player of the car used in these experiments). The speech in all the subsets is uttered by the same male speaker, and each subset contains 24 sentences of either seven or eleven digits, so that the total number of sentences in the test data is 96.
The speech recognition system used in the experiments is particularly configured to be used in portable devices, or in automotive applications. It includes a set of speaker-independent acoustic models (156 subphones covering the phonetics of English) with about 9,000 context-dependent gaussians (triphone contexts tied by using a decision tree) (see L. R. Bahl et al., Performance of the IBM Large Vocabulary Continuous Speech Recognition System on the ARPA Wall Street Journal Task, Proceedings of ICASSP 1995, volume 1, pp. 41-44, 1995), trained on a few hundred hours of general English speech (about half of these training data has either digitally added car noise, or was recorded in a moving car at 30 and 60 mph). The front end of the system computes 12 cepstra+the energy+delta and delta-delta coefficients from 15 ms frames using 24 melfilter banks; see chapter 3 in Rabiner et al., supra).
The codebook of cepstral vectors characterizing the cepstral space of the clean speech was obtained (prior to experimentation) by quantizing a collection of a few thousand sentences uttered by both males and females and recorded with an AKG Q400 microphone in a non-moving car and in a non-noisy environment (especially, with no music being played), using the same setup as for the test data. The quantization was performed by using a splitting algorithm, where the distance between a cepstral vector and a cepstral codeword is computed with a Malahanobis distance (a weighted Euclidean distance) with a diagonal covariance matrix tied over all codewords (see chapter 5, section 5.2.2 in Rabiner et al., supra). The dimension of the cepstra was 24, as determined by the number of mel-filters in the front end of the speech recognition system. Codebooks of different sizes were experimented with, particularly, codebooks containing either 2, 4, 8, 16, 32, 64, 128, 256 codewords.
In accordance with an embodiment of the present invention, the two input waveforms are preferably time-aligned and scaled by following the procedure described heretofore with relation to
The scaling factor between the amplitudes in the two channels is preferably estimated by computing the mean values of non speech segments about 450 ms long, in each channel, and by computing the ratio between the mean values.
The two-channel CDCN scheme is preferably applied as described heretofore in connection with
In Table 1 below, shown are the Word Error Rates obtained after decoding the test data. Each column correspond to given volume of music. In the first row, shown is the WER obtained when no compensation at all is applied. In the other rows, shown is the WER obtained by using the two-channel CDCN scheme with codebooks of size 2, 4, 8, 16, 32, 64 and 128.
The two channel CDCN scheme allows to significantly reduce the WER at all music volumes, and for each size of codebook. The codebook with 64 codewords seems to correspond to an optimal trade-off between the recognition accuracy and computational requirements. Smaller codebooks still yield significant improvements over the baseline scores, with a reduced computational load.
It is to be understood that the present invention, in accordance with at least one presently preferred embodiment, includes a first input medium which obtains an initial speech signal, a second input medium which obtains at least one interfering signal, and a normalizing arrangement which reconciles the initial speech signal and at least one interfering signal with one another to produce a final speech signal. Together, the first input medium, second input medium and normalizing arrangement may be implemented on at least one general-purpose computer running suitable software programs. These may also be implemented on at least one Integrated Circuit or part of at least one Integrated Circuit. Thus, it is to be understood that the invention may be implemented in hardware, software, or a combination of both.
If not otherwise stated herein, it is to be assumed that all patents, patent applications, patent publications and other publications (including web-based publications) mentioned and cited herein are hereby fully incorporated by reference herein as if set forth in their entirety herein.
Although illustrative embodiments of the present invention have been described herein with reference to the accompanying drawings, it is to be understood that the invention is not limited to those precise embodiments, and that various other changes and modifications may be affected therein by one skilled in the art without departing from the scope or spirit of the invention.
Number | Name | Date | Kind |
---|---|---|---|
4537200 | Widrow | Aug 1985 | A |
5309378 | Beierle | May 1994 | A |
5487129 | Paiss et al. | Jan 1996 | A |
5539832 | Weinstein et al. | Jul 1996 | A |
5590241 | Park et al. | Dec 1996 | A |
5590242 | Juang et al. | Dec 1996 | A |
5598505 | Austin et al. | Jan 1997 | A |
5737485 | Flanagan et al. | Apr 1998 | A |
5745872 | Sonmez et al. | Apr 1998 | A |
5781883 | Wynn | Jul 1998 | A |
5924065 | Eberman et al. | Jul 1999 | A |
5956679 | Komori et al. | Sep 1999 | A |
6157909 | Mauuary et al. | Dec 2000 | A |
6487295 | Lofgren et al. | Nov 2002 | B1 |
Number | Date | Country | |
---|---|---|---|
20020143528 A1 | Oct 2002 | US |