JP2013037174A - Noise/reverberation removal device, method thereof, and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a noise/reverberation removal device which uses only exemplar models of clean voice to perform voice emphasis.SOLUTION: An emphasizing processing result reliability calculation unit outputs a value indicative of uncertainty of a primary voice emphasis signal in accordance with a feature quantity of an input signal and the primacy voice emphasis signal. A matching unit receives the primary voice emphasis signal, the value indicative of uncertainty of the primary voice emphasis signal, and exemplar models of learning data and outputs a learning data segment which gives a clean voice sequence closest to clean voice included in the input signal with respect to each time frame. A voice emphasis filtering unit receives the input signal and the learning data segment, reads out amplitude spectrum data pairing the learning data segment from an exemplar model storage unit to generate a Wiener filter, multiplies a power spectrum of the input signal by the Wiener filter to perform filtering, and outputs a voice emphasis signal.

Description

  The present invention relates to a noise / dereverberation apparatus that extracts an acoustic signal from which noise and reverberation have been removed from an acoustic signal accompanied by noise and reverberation, and a method and program thereof.

  When an acoustic signal is collected in an environment with noise or reverberation, it is observed as a signal in which acoustic distortion (noise or reverberation) is superimposed on the original signal. When the acoustic signal is speech, the clarity of speech is greatly reduced due to the effect of superimposed acoustic distortion. As a result, it becomes difficult to extract the nature of the original speech signal, and for example, the recognition rate of the speech recognition system decreases. In order to prevent this reduction in the recognition rate, it is necessary to devise a method (method) for removing the superimposed acoustic distortion.

  In addition to voice recognition, this noise / reverberation removal method can be used for, for example, a hearing aid, a TV conference system, a machine control interface, a music information processing system for searching for music, and recording music.

  FIG. 7 shows an example of the functional configuration of a conventional noise / dereverberation apparatus 700, and its operation will be briefly described. The noise / dereverberation apparatus 700 includes a matching unit 703, a speech enhancement filtering unit 704, and a case model 705. The matching unit 703 performs matching between the input signal feature quantity and the feature quantity cases included in the case model 705, and searches for a case closest to the input signal.

  The case model 705 is a model composed of clean speech data corresponding to a case and noise / reverberation speech feature quantities paired therewith. This example model 705 simulates observation signals in various environments using a large amount of clean speech obtained from a speech corpus and noise / reverberation data (noise signal waveform and room impulse response) obtained in any environment. It is generated in advance by a case model learning device using a signal obtained by converting the simulated observation signal into a feature amount region.

  The speech enhancement filtering unit 704 creates a filter for speech enhancement using the clean spectrum amplitude spectrum case data used when searching for the case closest to the input signal, and filters the input signal. According to this method, it has been reported that it is possible to remove noise that has been difficult in the past and has a very large time variation. The noise having a very large time change is a noise such as an alarm sound of an alarm clock with respect to the background noise.

J. Ming and R. Srinivasan, and D. Crooke, “A C0rpus-Based Approach to Speech Enhancement From Nonstationary Noise,” IEEE Trans. On Acoustics, Speech and Signal Processing, 19 (4), pp. 822-836, 2011 .

  However, in the conventional method, noise / reverberation data for simulating noise / reverberation environment of any environment is necessary at the time of learning, and the amount of data is not sufficient, and the condition is sufficiently close to the noise / reverberation data at the time of speech enhancement Is not prepared as an example, it was difficult to perform accurate speech enhancement. In addition, it is possible to simulate the noise / reverberation environment of any environment, and even when sufficiently close examples are included in the case model at the time of speech enhancement, the number of cases becomes enormous, and the number of cases is the largest for the input signal. There is a problem that the amount of calculation for searching for a nearby case becomes very large.

  The present invention has been made in view of such a problem, and even if all noise / reverberation data is not prepared at the time of learning, a clean voice that seems to be closest to the clean voice included in the input signal is used as an example. It is an object to provide a noise / dereverberation apparatus that can be found using a model and perform accurate speech enhancement, and a method and program thereof.

  The noise / dereverberation apparatus of the present invention includes a speech enhancement processing unit, an enhancement processing result reliability calculation unit, a case model storage unit, a matching unit, and a speech enhancement filtering unit. The speech enhancement processing unit receives a speech digital signal on which noise and reverberation are superimposed as an input signal, and outputs a primary speech enhancement signal in a feature amount region obtained by performing primary speech enhancement processing on the input signal. The enhancement processing result reliability calculation unit outputs a value indicating the uncertainty of the primary speech enhancement signal from the feature amount of the input signal and the primary speech enhancement signal. The case model storage unit stores a case model of learning data and amplitude spectrum data thereof. The matching unit receives a primary speech enhancement signal, a value indicating the uncertainty of the primary speech enhancement signal, and a case model of learning data as input, and is closest to the clean speech included in the input signal for each time frame. A learning data segment giving a clean speech sequence is output. The speech enhancement filtering unit receives the power spectrum of the input signal and the learning data segment, reads out the amplitude spectrum data paired with the learning data segment from the case model storage unit, generates a Wiener filter, and generates the power spectrum of the input signal. A voice emphasis signal is output after filtering by the winner filter.

  According to the noise / dereverberation apparatus of the present invention, since a case model generated only from clean speech is used, the amount of calculation for case search can be reduced. At the same time, primary speech enhancement processing is performed on the input signal, and matching is performed in consideration of the uncertainty (reliability) of the speech enhancement processing, thereby making it possible to search for an example of an appropriate clean speech. Although specific effects will be described later, according to the present invention, the SN ratio for noise / reverberation removal can be improved as compared with the prior art while reducing the amount of calculation.

The figure which shows the function structural example of the noise / dereverberation apparatus 100 of this invention. The figure which shows the operation | movement flow of the noise / dereverberation apparatus 100. The figure which shows the function structural example of the example model production | generation apparatus 200. FIG. The figure which shows the operation | movement flow of the example model production | generation apparatus 200. FIG. It is a figure which shows the spectrogram of an evaluation experiment result, (a) is a clean voice, (b) is a reverberation voice, (c) is a conventional method, (d) is an output signal that has undergone matching processing without considering uncertainty, (E) is an output signal of the noise / dereverberation apparatus 100 of the present invention. It is a figure which shows an evaluation-experiment result by segmental SNR and logarithmic spectral distance, (a) is segmental SNR, (b) is logarithmic spectral distance. The figure which shows the function structural example of the conventional noise / dereverberation apparatus 700.

    Embodiments of the present invention will be described below with reference to the drawings. The same reference numerals are given to the same components in a plurality of drawings, and the description will not be repeated.

  FIG. 1 shows a functional configuration example of a noise / dereverberation apparatus 100 of the present invention. The operation flow is shown in FIG. The noise / dereverberation removing apparatus 100 includes a speech enhancement processing unit 102, an enhancement processing result reliability calculation unit 103, a case model storage unit 104, a matching unit 105, a speech enhancement filtering unit 106, and a control unit 107. It has. The function of each part of the noise / dereverberation apparatus 100 is realized by a predetermined program being read into a computer constituted by, for example, a ROM, a RAM, and a CPU, and the CPU executing the program.

  The output signal region of the noise / dereverberation apparatus 100 can be output in various signal regions such as a time region, a power spectrum region, an amplitude spectrum region, and a feature amount region, and is selected according to the use of the output signal. In the description of this embodiment, the input signal will be described as the power spectrum region and the output signal will be described as the time domain signal.

  Since the input signal is given in the power spectrum region, the feature amount generation unit 101 is provided in this embodiment. The feature value generation unit 101 generates a feature value (for example, mel frequency cepstrum coefficient) for each frame from the input power spectrum (step S101). If the input signal is given in the feature quantity region, the feature quantity generation unit 101 is not necessary. Therefore, the feature quantity generation unit 101 is indicated by a broken line.

The input signal y _t of the feature region is modeled as shown in equation (1).

y _t is an input signal of time frame t, _st is clean speech, and b _t is an acoustic distortion component (noise or rear reverberation component). Modeling noise as an additive term in this manner is widely performed, and posterior reverberation is often modeled as an additive term (Reference 1: K. Kinoshita, M.). Delcroix, T. Nakatani, and M. Miyoshi, “Suppression of late reverberation effect on speech signal using long-term multiple-step linear prediction,” IEEE TASLP, 17 (4), pp. 534-545, 2009.). In the following description, the signals in the power spectrum region are denoted as Y _t ² , S _t ² , and B _t ² , respectively.

Speech enhancement processing unit 102, the superimposed audio digital signal of the noise-reverberation as an input signal, and outputs the primary audio enhancement signals ^~ s _t of the feature region subjected to first-order speech enhancement process on the input signal (Step S102). ^The correct notation is that the position of ^~ is located immediately above the variable as in the notation of (Expression (2)). Enhancement processing result reliability calculation unit 103, an input signal y _t, and a primary audio enhancement signals ^~ s _t outputted by the sound enhancement processor 102, a value indicating the uncertainty of the primary audio enhancement signals ^~ s _t sigma _bt is output (step S103).

The case model storage unit 104 stores a case model of learning data and amplitude spectrum data thereof. Matching unit 105 includes a primary audio enhancement signals ^~ s _t outputted by the sound enhancement processing unit 102, enhancement processing result value indicating the uncertainty of the primary audio reliability calculation unit 103 outputs enhancement signal ^~ s _t ^Σ _bt When, and outputs the learning data segments to provide a closest clean speech sequence to clean speech included in the input signal y _t and case model M of the learning data stored in the case model storage unit 104, as inputs (step S105 ).

The speech enhancement filtering unit 106 receives the power spectrum Y _t ² of the input signal and the learning data segment output from the matching unit 105 as input, and reads out the amplitude spectrum data paired with the learning data segment from the case model storage unit 104. A Wiener filter is generated and filtered by multiplying the power spectrum Y _t ² of the input signal by the Wiener filter to output a speech enhancement signal (step S106). The control unit 107 controls time-series operations between the above-described units.

  By operating as described above, the noise / dereverberation apparatus 100 uses a case model generated only from clean speech, uses a small amount of calculation for case search, and has a good SN ratio. Allows removal.

  In the following, the function of each part of the noise / dereverberation apparatus 100 will be described in more detail.

[Speech enhancement processor]
Speech enhancement processing unit 102 of this embodiment, the input signal because the feature quantity region, directly to the input signal y _t, subjected to first-order speech enhancement. The process for obtaining the primary speech enhancement signals ^~ s _t, applicable any conventional speech enhancement method, a method of applying the like should be suitably chosen depending on the type of acoustic distortion contained in the input signal . For example, a method of linearly predicting a reverberation component from a past signal and removing it in the power spectrum region (Reference Document 2: Table 2007/100137) can be used.

[Enhancement processing result reliability calculation section]
Enhancement processing result reliability calculation unit 103, a primary audio enhancement signals ^~ s _t, using the feature quantity y _t of the input signal, a value indicating the uncertainty of the enhanced speech (primary speech enhancement signals ^{~ s} _t) ^Σ _bt is calculated and output. Although the total covariance matrix may be used as the value Σ _bt indicating the uncertainty, in this embodiment, Σ _bt is a diagonal covariance matrix that is a covariance matrix having a diagonal component of zero, The k-th diagonal element σ _k is calculated as shown in Equation (2).

  k is an index representing the order of the feature vector.

That is, enhancement processing result reliability calculation unit 103, the value sigma _bt indicating the uncertainty of the primary audio enhancement signals ^~ s _t, a difference between the feature quantity y _t and the primary audio enhancement signals ^~ s _t of the input signal A covariance matrix is used as a component.

[Case model generator]
Here, a case model generation apparatus 200 that generates a case model stored in the case model storage unit 104 will be described. FIG. 3 shows a functional configuration example of the case model generation apparatus 200. The operation flow is shown in FIG. The case model generation apparatus 200 includes a Fourier transform unit 201, a feature value generation unit 202, a Gaussian mixture model learning unit 203, a maximum likelihood Gaussian distribution calculation unit 204, and a control unit 205. The function of each part of the case model generation apparatus 200 is realized by, for example, reading a predetermined program into a computer constituted by a ROM, a RAM, a CPU, and the like, and executing the program by the CPU.

  The Fourier transform unit 201 receives clean speech of a speech digital signal as an input signal, and the input signal is windowed with a short Hamming window of about 30 ms, for example, and each windowed input signal is subjected to discrete Fourier transform to an amplitude spectrum. Conversion is performed (step S201). An amplitude spectrum is amplitude data of a frequency spectrum.

The feature value generation unit 202 converts all of the amplitude spectrum output from the Fourier transform unit 201 into a mel cepstrum feature value s _i . In general, the mel cepstrum widely used is about 10 to 20th order, but in order to accurately represent the case data, a mel cepstrum having a high order (for example, about 30 to 100th order) is used. Note that feature quantities other than the mel cepstrum may be used.

The Gaussian mixture model learning unit 203 uses the feature quantity s _i in each short time frame i obtained by the feature quantity generation unit 202 as learning data, and uses a Gaussian mixture model g (formula (3)) by a normal maximum likelihood estimation method. Get.

g (s | q) represents the q-th Gaussian distribution with mean mu _q, the dispersion sigma _q, w (q) represents a mixture weight for it. Q represents the number of mixtures.

The maximum likelihood Gaussian distribution calculation unit 204 obtains an index q _i of the Gaussian distribution in the Gaussian mixture distribution g giving the maximum likelihood for each time frame i, and uses the time series of the index q _i as a case model M. Obtained (step S204). The case model M is expressed as shown in Expression (4) using a set of Gaussian distribution indices q _i and a Gaussian mixture model g.

Here, q _i is the index of the Gaussian distribution that gives the maximum likelihood for the i-th frame of the feature s _i, the distribution g in the Gaussian mixture distribution q | represents (s q _i) Yes. The model M is referred to as a case model M that captures detailed time-frequency characteristics of the learning data s. The case model M is stored in the case model storage unit 204 (FIG. 1), for example, together with the amplitude spectrum data A of clean speech for learning that is paired with the learning data s.

[Matching part]
Matching unit 105 includes a feature amount y _t of the input signal, the segment closest training data to the feature quantity y _t of the input signal, and probed with case model M, the clean speech s included in the input signal y _t learning seems to give the closest clean speech series in _t data segment ^M _{t u:} to output the _{u + τmax.} Matching unit 105, in consideration of the value sigma _bt indicating the uncertainty of the primary audio enhancement signals ^~ s _t, but is intended to explore the nearest clean speech sequence to clean speech, the conventional method without consideration of sigma _bt For the purpose of clarifying the difference from the above, a matching method that does not take into account the value Σ _bt indicating uncertainty is described first.

It is assumed that the input signal is composed of T time frames, and the input signal is y = {y _t : t = 1, 2,..., T}. Also, let _{yt: t + τ be} a sequence from the time frame t to t + τ of the input signal. Then, M _{u: u + τ} = {g, q _i : i = u, u + 1,..., U + τ} is a Gaussian distribution sequence corresponding to continuous time frames from u-th to u + τ-th in the learning data s. .

Definition and of the distance between a segment in the input signal y _t and the training data s, as a search method of the input signal y _t and the closest training data, be considered a Euclidean distance, etc., several other methods I can do it. Here, it is assumed that the learning data segment closest to the time frame t of the input signal y has the longest length among learning data segments that closely match the input signal. In other words, the closest training data segments M ^{t u} to the input _{signal: u + tau} can be determined by maximizing a posterior probability shown in the following equation.

Here, p (M _{u: u + τ} | y _{t: t + τ} ) represents the posterior probability, and when y _{t: t + τ} and M _{u: u + τ} are relatively well matched, τ is The longer it is, the higher the posterior probability is. By taking a measure of searching for a longer segment, it becomes difficult to be affected by noise that exists locally at a certain time, and it can be expected that relatively robust matching is performed with respect to noise. In equation (6), for simplicity, p (M _{u: u + τ} ) can assume equal probabilities for all learning data segments. This corresponds to the assumption that the sequence patterns observed in the training data can all occur with the same probability when noise / dereverberation is removed.

The numerator term p (y _{t: t + τ} | M _{u: u + τ} ) in equation (6) is the likelihood of the speech enhancement signal y _{t: t + τ} for the learning data segment corresponding to M _{u: u + τ.} Degree. The likelihood is calculated by the following equation.

For simplicity, it is assumed that adjacent frames are independent. The denominator of Expression (6) is a value obtained by taking the sum of p (y _{t: t + τ} | M _{u: u + τ} ) for all patterns included in the case model M.

Here, if the input signal y _t is close enough to clean speech, i.e. the closer to zero sufficiently audio distortion component b _t, mismatch between the clean speech data is reduced using at the time of learning, close to the clean speech s _t Patterns can be searched from learning data. In general, however, the input signal y _t and the clean speech s _t there is a difference due to noise / reverberation, the difference directly affects the matching process. Therefore, it is not easy to discover from the learning patterns a pattern close to the clean speech s _t as it is. It is necessary to devise a technique for reducing the influence of the difference due to the noise / reverberation.

Therefore, the noise / dereverberation apparatus 100 of the present invention takes into account uncertainty (reliability) for the purpose of reducing the influence of the difference due to noise / reverberation. That is, the noise / dereverberation apparatus 100 of the present invention, while considering the reliability by matching the input signal y _t and learning data, most close to the input signal training data segments M ^{t _u:} searching the _{u + .tau.max.}

Therefore, in order to take into account that there is a difference between the primary audio enhancement signals ^~ s _t a clean speech s _t explicitly considers the reliability / uncertainty of the primary audio enhancement signals ^~ s _t. Specifically, probabilistically formulate the input signal y _t.

First, it is assumed that the noise / reverberation component b _t follows the following Gaussian process.

Here, ^ _{b t} is the estimated value of the difference of the primary audio enhancement signals ^~ s _t and the input signal _{y t, ^ b t = y} t - calculated as ~ _{s t,} the primary speech enhancement signal value sigma _bt indicating the uncertainty likeness of ^~ s _t is a variable of the covariance matrix when _{b t.} By using this formulation, the likelihood of the input signal y _t is by marginalizing the clean speech signal joint probability can be determined as follows.

In the derivation, the probabilistic multiplication theorem was used. From the equation (9), when the change of the covariance matrix Σ _bt can be thought of as a measure of the uncertainty likeness of the primary speech enhancement signal ^~ s _t. For example, for an uncertain feature amount with low reliability, the covariance matrix _Σbt corresponding to the feature amount increases, and as a result, the influence of the feature amount on the result is reduced.

Thus the task of varying correction when the dispersion section of the Gaussian distribution, by inserting the equation (6), primarily first-order speech enhancement signals ^~ s _t reliability is the result of the speech enhancement / taking into account the uncertainty likeness, learning closer to the clean speech signal s _t data segment M ^{t _u:} it is possible to search the _{u + .tau.max.}

[Speech enhancement filtering part]
Speech enhancement filtering unit 106, the learning data segment M ^{t u} matching unit 105 _outputs: the _{u + .tau.max,} performs filtering using the case of the amplitude spectrum of the clean speech corresponding thereto.

First, the matching result M ^{t _u:} the amplitude spectrum of the clean speech corresponding to _{u + .tau.max,} read from case model storage unit 104 attempts to restore the amplitude spectrum of the clean speech component s contained in the input signal yt. Assuming that ε (ε = 1, 2,..., T) is a time frame index of a target for which the amplitude spectrum of the clean speech is to be restored, the clean speech amplitude spectrum ^ S _ε is estimated and restored as follows.

Here, A (u ^t _ε ) is an example of the amplitude spectrum of the clean speech paired with the learning data segment M ^t _{u: u + τmax,} and u ^t _ε is the likely learning data segment u obtained at each frame t. = Index corresponding to ε of {u, u + 1,..., U + τmax}. A set [A] of clean speech amplitude spectrum data is {A (i): i = 1, 2,..., I _s }.

Next, to construct a Wiener filter H _epsilon using amplitude spectrum ^ S _epsilon that the estimated (equation (11)).

The estimated value of noise / reverberation component B ² _ε is obtained as shown in equation (12).

Here, α is a smoothing coefficient, and max [k, k ′] is a function that selects and outputs the larger of k and k ′. The Wiener filter H _epsilon as H _t, is multiplied to the H _t to the power spectrum Y _t ² of the input signal, it is possible to obtain a final output signal.

An output signal obtained by multiplying the power spectrum Y _t ² of the input signal by the Wiener filter H _t is subjected to inverse Fourier transform, converted into a time domain signal, and output.

[Evaluation experiment]
An evaluation experiment was conducted for the purpose of evaluating the performance of the noise / dereverberation apparatus 100 of the present invention. The experimental conditions were as follows.

  For training of the Gaussian mixture model g, TIMIT core training-set consisting of 1088 sentences and 136 speakers was used. A sampling frequency is 8 kHz, and a vector obtained by connecting a 40th-order mel cepstrum coefficient and a logarithmic energy term is used as a feature vector used for learning a Gaussian mixture model. As the mixture number Q of the Gaussian mixture model, 4096, which is a sufficiently large value, is used in order to accurately model various temporal frequency patterns included in the learning data. The frame length used for Fourier transform was 20 ms, and the shift width of the short time window was 10 ms.

In the experiment, assuming a room with a size of 5m x 5m x 5m and a reverberation time of 0.5 seconds, the room impulse response that would be measured in a situation where the speaker was 2.5m away from the microphone in the room. Simulated above. Input signals y _t to noise / dereverberation apparatus 100, produced by convoluting a and sound 64 sentences included in the room impulse response and TIMIT core training-set. The speech enhancement processing to obtain a 1 a-order speech enhancement signal primary audio enhancement signals ^~ s _t, using the method of Reference 2 described above.

  FIG. 5 shows the experimental results in a spectrogram. The horizontal axis represents time, and the vertical axis represents frequency. The intensity of the frequency is represented by black and white shading. (A) is an input signal, (b) is a reverberant voice, (c) is an output signal according to a conventional method, (d) is an output signal that has been subjected to matching processing without taking uncertainty into consideration, and (e) is the noise of the present invention. This is an output signal of the dereverberation apparatus 100.

  Looking at the output signal (c) according to the conventional method, although a certain degree of dereverberation effect can be confirmed, the energy of the portion where the original speech energy exists is excessively suppressed, and the inaccuracy of the processing is confirmed. I can do it. On the other hand, the output signal (d) of the processing that has been matched without considering the uncertainty outputs the sound with slightly less distortion than the conventional method (c) by connecting the processing based on the case. .

  The output signal (e) of the noise / dereverberation apparatus 100 of the present invention is about 0.54 seconds, 0.81 seconds, indicated by an arrow ↓, indicating that dereverberation is more effective than the above two processed sounds. It can be seen from the recovery of the harmonic structure around 0.96 seconds.

  Next, in order to more objectively evaluate the effect of the noise / reverberation removal method of the present invention, a sectional SNR and a logarithmic spectral distance were calculated. The higher the segmental SNR, the more accurately the acoustic distortion is removed. Conversely, the smaller the logarithmic spectral distance, the closer the sound is to clean sound. The average value of the results obtained from all the evaluation voices is shown in FIG. The horizontal direction in FIG. 6 is a processing method, which is the input signal (□) from the left, the conventional method, matching processing without considering uncertainty, and the present invention (■). In the vertical axis direction, (a) is the segmental SNR (dB), and (b) is the logarithmic spectral distance (dB).

  As described above, according to the noise / reverberation removal method of the present invention, the best numerical value can be obtained for both the sectional SNR and the logarithmic spectral distance by using only the case model generated only from clean speech. That is, according to the noise / reverberation removal method of the present invention, noise / reverberation data at the time of learning is no longer required, so that the amount of calculation can be reduced and the SN ratio of noise / dereverberation removal can be improved over the prior art. It becomes possible.

  When the processing means in the noise / dereverberation apparatus 100 and the example model generation apparatus 200 described above is realized by a computer, the processing contents of functions that each apparatus should have are described by a program. Then, by executing this program on the computer, the processing means in each apparatus is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, as a magnetic recording device, a hard disk device, a flexible disk, a magnetic tape or the like, and as an optical disk, a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), a CD-ROM (Compact Disc Read Only). Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording medium, MO (Magneto Optical disc), etc., semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Further, the program may be distributed by storing the program in a recording device of a server computer and transferring the program from the server computer to another computer via a network.

  Each means may be configured by executing a predetermined program on a computer, or at least a part of these processing contents may be realized by hardware.

Claims (7)

A speech enhancement processing unit which outputs a primary speech enhancement signal in a feature amount region obtained by performing primary speech enhancement processing on the input signal, using a speech digital signal on which noise and reverberation are superimposed;
An enhancement processing result reliability calculation unit that outputs a value indicating the uncertainty of the primary speech enhancement signal from the feature amount of the input signal and the primary speech enhancement signal;
A case model of learning data, a case model storage unit for storing the amplitude spectrum data,
Using the primary speech enhancement signal, a value indicating the uncertainty of the primary speech enhancement signal, and the example model of the learning data as inputs, the cleanest closest to the clean speech included in the input signal for each time frame A matching unit that outputs a training data segment that gives a speech sequence;
Using the power spectrum of the input signal and the learning data segment as input, amplitude spectrum data paired with the learning data segment is read from the case model storage unit to generate a Wiener filter, and the winner spectrum is added to the power spectrum of the input signal. A voice enhancement filtering unit that filters and filters to output a voice enhancement signal;
A noise / dereverberation apparatus comprising: The noise / dereverberation apparatus according to claim 1,
The emphasis processing result reliability calculation unit is
A noise / dereverberation apparatus characterized in that a value indicating the uncertainty of the primary speech enhancement signal is a covariance matrix whose component is a difference between the feature amount of the input signal and the primary speech enhancement signal. The noise / dereverberation apparatus according to claim 1 or 2,
The learning data segment that gives the clean speech sequence closest to the clean speech included in the input signal for each time frame output by the matching unit is the learning data segment that closely matches the feature quantity of the input signal. Noise / dereverberation device characterized by being the longest. A speech enhancement process for outputting a primary speech enhancement signal in a feature amount region obtained by performing a primary speech enhancement process on the input signal using a speech digital signal on which noise and reverberation are superimposed;
An enhancement processing result reliability calculation process for outputting a value indicating the uncertainty of the primary speech enhancement signal from the feature amount of the input signal and the primary speech enhancement signal;
A case model of learning data, a case model storage unit for storing the amplitude spectrum data,
The primary speech enhancement signal, a value indicating the uncertainty of the primary speech enhancement signal, and the case model of the learning data stored in the case model storage unit are input and included in the input signal for each time frame. A matching process that outputs a training data segment giving the clean speech sequence closest to the clean speech;
Using the power spectrum of the input signal and the learning data segment as inputs, the amplitude spectrum data stored in pairs with the learning data segment is read from the case model storage unit to generate a Wiener filter, and the input signal A voice enhancement filtering process of outputting a voice enhancement signal by multiplying the power spectrum by the Wiener filter and filtering;
A noise / dereverberation method comprising: The noise / dereverberation method according to claim 4,
The emphasis processing result reliability calculation process is as follows:
A noise / reverberation removal method characterized in that a value indicating the uncertainty of the primary speech enhancement signal is a covariance matrix whose component is a difference between the feature amount of the input signal and the primary speech enhancement signal. The noise / dereverberation method according to claim 4 or 5,
The learning data segment that gives the clean speech sequence closest to the clean speech included in the input signal for each time frame output by the matching process is the learning data segment that closely matches the feature quantity of the input signal. Noise / dereverberation method characterized by being the longest.   A program for causing a computer to function as the noise / dereverberation apparatus according to any one of claims 1 to 3.

Images