Patent Grant
8306249
This application claims the priority, under 35 U.S.C. §119, of European application EP 09005597, filed Apr. 21, 2009; the prior application is herewith incorporated by reference in its entirety.
The present invention relates to a method, an acoustic signal processing device and a use of an acoustic processing device for estimating linear predictive coding coefficients.
In signal enhancement tasks, adaptive Wiener filtering is often used to suppress background noise and interfering sources. For constructing a Wiener filter it is necessary to have at least an estimate of the noise power spectral density (PSD). Conventional speech enhancement systems typically rely on the assumption that the noise is rather stationary, i.e., its characteristics change very slowly over time. Therefore, noise characteristics can be estimated during speech pauses but requiring a robust speech activity detection (VAD). More sophisticated methods are able to update the noise estimate even during speech activity and thus do not require a VAD. This is performed by decomposing the noisy speech into sub-bands and tracking minima in these sub-bands over a certain time interval. Because of the higher dynamics of the speech signal the minima should correspond to the noise PSD if the noise is sufficiently stationary. However, this method fails if the noise characteristics exceed a certain degree of non-stationarity and thus the performance in highly non-stationary environments (e.g., babble noise in a cafeteria) breaks down severely.
More recently, model-based speech enhancement methods have emerged that utilize a priori knowledge about speech and noise. In the reference by S. Srinivasan, titled “Codebook Driven Short-Term Predictor Parameter Estimation for Speech Enhancement”, IEEE Trans. Audio, Speech, and Language Process., vol. 14, no. 1, January 2006, pp. 163-176 one of these methods is described in detail. The main idea disclosed is to estimate linear predictive coding (LPC) coefficients, i.e., prediction coefficients and excitation variances (gains) of speech and noise from the noisy signal. The LPC coefficients directly correspond to spectral envelopes of the speech and noise signal parts. For distinguishing between speech and noise, trained codebooks are used that contain typical sets of prediction coefficients (i.e., typical spectral envelopes) of speech and noise.
The estimation method involves building every possible pair of speech and noise parameter sets taken from the respective codebooks and computing the optimum gains so that the sum of the LPC spectra of speech and noise fits best to the observed noisy spectrum. The proposed criterion is the Itakura-Saito distance between the sum of the LPC spectra and the observed noisy spectrum. The Itakura-Saito distance has shown a good correlation with human perception. The codebook combination with the respective gains that globally minimizes the Itakura-Saito distance is considered as the best estimate. With the corresponding LPC spectra a Wiener filter for noise reduction is constructed. It is disclosed that minimizing the Itakura-Saito distance results in the maximum likelihood (ML) estimate of the speech and noise parameters. The disclosed method has the advantage of enhancing every signal frame independently and thus it is able to react instantaneously to noise fluctuations. Therefore it can deal with highly non-stationary noise.
Besides the ML method, a minimum mean-square error (MMSE) approach has been disclosed in the reference by S. Srinivasan, titled “Codebook-Based Bayesian Speech Enhancement for Nonstationary Environments”, IEEE Trans. Audio, Speech, and Language Process., vol. 15, no. 2, February 2007, pp. 441-452. The parameter estimates are not single codebook entries anymore but a weighted sum of all possible combinations of codebook entries with the weights being proportional to the probability that the codebook entry combination corresponds to the observed noisy signal. This probability is called the likelihood and is denoted as p(x|θ), where x denotes a frame of noisy speech samples and θ is a vector containing the speech and noise LPC parameters. It is further disclosed that incorporating memory improves the estimation accuracy.
Memory is incorporated in the form of conditional probabilities and the weights are proportional to
p(x|θ)p({circumflex over (θ)}s,k-1|θs)p({circumflex over (θ)})n,k-1|θn). (1)
θs and θn denote the LPC parameters (without the gains) of speech and noise of the current frame. {circumflex over (θ)}s,k-1 and {circumflex over (θ)}n,k-1 are the estimates of the respective parameters from the preceding frame. By applying suitable models for the conditional probabilities p({circumflex over (θ)}s,k-1|θs) and p({circumflex over (θ)}n,k-1|θn) the estimation accuracy can be improved considerably because ambiguities arising from the Itakura-Saito-distance using as the only optimization criterion can be reduced.
The conditional probabilities p({circumflex over (θ)}s,k-1|θs) and p({circumflex over (θ)}n,k-1|θn) are modeled as multivariate Gaussian Random Walks N:
p({circumflex over (θ)}s,k-1|θs)˜N({circumflex over (θ)}s,k-1,Λs)
p({circumflex over (θ)}n,k-1|θn)˜N({circumflex over (θ)}n,k-1,Λn), (2)
where Λs and Λn are diagonal matrices with variances on their diagonals that are estimated from training data. It is reported that using this model the estimation accuracy of the speech parameters is not or at least only very little affected.
It is accordingly an object of the invention to provide a method and an acoustic signal processing device for estimating linear predictive coding coefficients which overcome the above-mentioned disadvantages of the prior art methods and devices of this general type, which improves noise and speech estimations.
The invention claims a method for estimating a set of linear predictive coding coefficients of a microphone signal using minimum mean-square error estimation with a codebook containing several predetermined sets of linear predictive coding coefficients. The method includes determining sums of weighted backward transition probabilities describing the transition probabilities between the predetermined sets of linear predictive coding coefficients. The backward transition probabilities are obtained from signal training data by mapping the signal training data to one set of the codebook and by determining relative frequencies of transitions between two sets of the codebook. Modelling the “memory” of the system according to the invention has the advantage that the estimation accuracy is increased considerably also for speech components.
In a preferred embodiment the method can include weighting every backward transition probability with a first weight of the corresponding predetermined set of linear predictive coding coefficients determined at a preceding time instant.
In a further embodiment the method can include weighting the predetermined sets of linear predictive coding coefficients with the corresponding weighted sum of backward transition probabilities.
In a preferred embodiment the first weights can be a measure for the probability that the combination of predetermined sets of linear predictive coding coefficients may have produced the microphone signal.
In a further embodiment the method can include determining second weights for all predetermined sets of linear predictive coding coefficients for a current time frame. The second weights denote a measure for the probability that the combination of predetermined sets of linear predictive coding coefficients may have produced the microphone signal at the current time frame. The method can further include summing all predetermined sets of linear predictive coding coefficients weighted with the determined weighted transition probabilities and the determined second weights yielding the estimated set of linear predictive coding coefficients at the current time frame.
Furthermore the method can be carried out with a speech codebook and a noise codebook.
The invention also claims an acoustic signal processing device for estimating a set of linear predictive coding coefficients of a microphone signal using minimum mean-square error estimation with a codebook containing several predetermined sets of linear predictive coding coefficients. The device includes a signal processing unit which determines sums of weighted backward transition probabilities describing the transition probabilities between the predetermined sets of linear predictive coding coefficients. The backward transition probabilities are obtained from signal training data by mapping the signal training data to one set of the codebook and by determining relative frequencies of transitions between two sets of the codebook.
In a preferred embodiment every backward transition can be weighted with a first weight of the corresponding predetermined set of linear predictive coding coefficients determined at a preceding time instant.
Furthermore the predetermined sets of linear predictive coding coefficients can be weighted with the corresponding weighted sum of backward transition probabilities.
In a further embodiment the first weight can be a measure for the probability that the combination of the predetermined sets of linear predictive coding coefficients may have produced the microphone signal.
In a preferred embodiment second weights can be determined for all predetermined sets of linear predictive coding coefficients for a current time frame. The second weights denote a measure for the probability that the combination of the predetermined sets of linear predictive coding coefficients may have produced the microphone signal at the current time frame. All predetermined sets of linear predictive coding coefficients can be weighted with the determined weighted transition probabilities and the determined second weights and can be summed yielding the estimated set of linear predictive coding coefficients at the current time frame.
Finally, estimating a set of linear predictive coding coefficients can be carried out with a speech codebook and a noise codebook.
The invention also claims a use of an acoustic signal processing device according to the invention in a hearing aid. The invention provides the advantage of an improved noise reduction.
Other features which are considered as characteristic for the invention are set forth in the appended claims.
Although the invention is illustrated and described herein as embodied in a method and an acoustic signal processing device for estimating linear predictive coding coefficients, it is nevertheless not intended to be limited to the details shown, since various modifications and structural changes may be made therein without departing from the spirit of the invention and within the scope and range of equivalents of the claims.
The construction and method of operation of the invention, however, together with additional objects and advantages thereof will be best understood from the following description of specific embodiments when read in connection with the accompanying drawings.
Since the present application is preferably applicable to hearing aids, such devices shall be briefly introduced in the next two paragraphs together with
Hearing aids are wearable hearing devices used for supplying hearing impaired persons. In order to comply with the numerous individual needs, different types of hearing aids, like behind-the-ear hearing aids and in-the-ear hearing aids, e.g. concha hearing aids or hearing aids completely in the canal, are provided. The hearing aids listed above as examples are worn at or behind the external ear or within the auditory canal. Furthermore, the market also provides bone conduction hearing aids, implantable or vibrotactile hearing aids. In these cases the affected hearing is stimulated either mechanically or electrically.
In principle, hearing aids have one or more input transducers, an amplifier and an output transducer as essential components. An input transducer usually is an acoustic receiver, e.g. a microphone, and/or an electromagnetic receiver, e.g. an induction coil. The output transducer normally is an electro-acoustic transducer like a miniature speaker or an electro-mechanical transducer like a bone conduction transducer. The amplifier usually is integrated into a signal processing unit. Such principle structure is shown in
The invention utilizes the MMSE estimation scheme described in the reference by S. Srinivasan, entitled “Codebook-Based Bayesian Speech Enhancement for Nonstationary Environments”, IEEE Trans. Audio, Speech, and Language Process., vol. 15, no. 2, February 2007, pp. 441-452. However, a completely different model is used for the conditional probabilities p({circumflex over (θ)}s,k-1|θs) and p({circumflex over (θ)}n,k-1|θn). The invention is based on the fact that the temporal evolution of the prediction parameters can be modeled as a Markov chain. A Markov chain consists of a finite set of states, which are equal to codebook entries θs, θn according to the invention, and transition probabilities between the states. Every codebook entry contains a set of LPC coefficients. The transition probabilities are obtained from training data by first mapping each frame of training data to one codebook entry and secondly computing the relative frequencies of transitions between two codebook entries (Markov states).
aij=p(Skj|Sk-1i) (3)
can be converted to the backward transition probabilities
bij=p(Sk-1j|Ski) (4)
via Bayes' rule. The backward transition probabilities bij directly correspond to the conditional probabilities p({circumflex over (θ)}s,k-1=θsj) modeling the memory. Given that the state estimate, i.e., the estimate of the spectral envelope, at the preceding time instant was
{circumflex over (θ)}s,k-1=θsj, (5)
we get
bij=p({circumflex over (θ)}s,k-1|θsi) (6)
and likewise for the noise. However, this only holds if the state estimate were uniquely defined by only one codebook entry.
In the MMSE estimation scheme, the state estimate is a weighted sum of all possible states, so the transition probabilities are a weighted sum of the backward transition probabilities bij, as well. In this case, the transition probabilities are computed as
where the ws,k-1j denote the weights of the states (i.e., the weights of the codebook entries) at the preceding time frame and Ns denotes the number of (speech) codebook entries. Similar holds also for the noise.
In the first step 100 Ns first weights ws,k-1j for all codebook sets for the time frame k−1 which is the preceding time frame to time frame k are determined. The first weights ws,k-1j denote a measure for the probability that a codebook set may may have produced the actual microphone signal at the preceding time frame k−1.
In step 101 the backward transition probabilities bij between every pair of codebook sets θsi, θsj, are used to weight the Ns weights ws,k-1j determined in step 100. The backward transition probabilities bij are obtained from signal training data by mapping the signal training data to one set of the codebook and by determining relative frequencies of transitions between two sets of the codebook.
In step 102 all Ns weighted backward transition probabilities bij are summed up for every Ns codebook set θsj resulting in Ns transition probabilities p({circumflex over (θ)}s,k-1|θsi).
In step 103 Ns second weights ws,kj for all codebook sets θsj for the current time frame k are determined. The second weights ws,kj denote a measure for the probability that a codebook set θsj may have produced the microphone signal at the current time frame k.
In the final step 104 sum of all Ns codebook set θsj weighted with the determined transition probabilities p({circumflex over (θ)}s,k-1|θsi) and the determined weights ws,kj is calculated which yields the estimated set {circumflex over (θ)}s,k of linear predictive coding coefficients for speech at the time frame k.
x(k)=s(k)+n(k). (7)
Speech and noise are assumed to be uncorrelated. With a filter h(k) an estimate ŝ(k) of the possibly time delayed clean speech signal can be obtained according to
ŝ(k)=h(k)*x(k), (8)
where “*” denotes linear convolution. The equivalent formulation in the frequency-domain reads
Ŝ(Ω)=H(Ω)×X(Ω). (9)
The optimal solution to this problem in the minimum mean-squared error (MMSE) sense is the well known Wiener filter 6
where Sss(Ω) and Sxx(Ω) denote the auto power spectral densities (PSD) of the clean speech signal s(k) and the noisy microphone signal x(k), respectively.
In a real noise reduction scheme, Sss(Ω) has to be estimated since only the noisy speech PSD Sxx(Ω) is accessible. However, in nearly all applications it is much easier to get an estimate of the noise PSD Snn(Ω). Given the fact that speech and noise are assumed to be uncorrelated the speech PSD Sss(Ω) can be expressed as the difference between Sxx(Ω) and Snn(Ω)
Sss(Ω)=Sxx(Ω)−Snn(Ω) (11)
that yields an alternative formulation of the Wiener filter 6
Equation 12 shows that for building a Wiener filter 6 it is also sufficient to have an estimate of the noise PSD Snn(Ω). So the noise reduction task can be reduced to the task of estimating the noise PSD Snn(Ω).
In accordance with the invention the noise PSD Snn(Ω) and/or the speech PSD Sss(Ω) can be calculated by using estimated linear predictive coding coefficients {circumflex over (θ)}s,k, {circumflex over (θ)}n,k. Therefore, the Wiener filter 6 can be built by estimating the linear predictive coding coefficients {circumflex over (θ)}s,k, {circumflex over (θ)}n,k according to the method described above. The estimation is performed in a signal processing unit 3.
Preferably, the acoustic processing device according to the invention is used in a hearing aid for reducing background noise and interfering sources.
| Number | Date | Country | Kind |
|---|---|---|---|
| 09005597 | Apr 2009 | EP | regional |
| Number | Name | Date | Kind |
|---|---|---|---|
| 5953697 | Lin et al. | Sep 1999 | A |
| 5966689 | McCree | Oct 1999 | A |
| 6009388 | Ozawa | Dec 1999 | A |
| 6182030 | Hagen et al. | Jan 2001 | B1 |
| 6226607 | Chang et al. | May 2001 | B1 |
| 6385578 | Lee et al. | May 2002 | B1 |
| 6732070 | Rotola-Pukkila et al. | May 2004 | B1 |
| 6807527 | Rozhdestvenskij et al. | Oct 2004 | B1 |
| 7587316 | Morii | Sep 2009 | B2 |
| 8116490 | Puder et al. | Feb 2012 | B2 |
| 20010053229 | Azizi | Dec 2001 | A1 |
| 20020176594 | Hohmann et al. | Nov 2002 | A1 |
| 20040023677 | Mano et al. | Feb 2004 | A1 |
| 20070124140 | Iser et al. | May 2007 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20100266152 A1 | Oct 2010 | US |
