JP2005031524A - Speech signal extracting method and speech recognition device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a method of extracting a speech signal from one input speech signal including a mixed signal such as noise. <P>SOLUTION: A signal storage part 11 stores finite past values of one input signal. A plural-filter determination part 12 determines a plurality of filters through independent component analysis based upon the finite past values and uses the finite past values as inputs to obtain a plurality of filter outputs. A filter output selection part 13 selects a filter output corresponding to a speech signal component from a plurality of filter outputs. A speech signal composition part 14 composes a speech signal based upon the selected filter output. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a speech signal extraction method for extracting only a speech signal from one signal including a speech signal component and a noise component in an environment where noise exists, and a speech recognition apparatus using the speech signal extraction method. About.

As a signal extraction method using a plurality of input signals from a plurality of microphones, for example, as described in Non-Patent Document 1, there is a signal extraction method using independent component analysis. In this signal extraction method, focusing on the fact that speech and noise are statistically independent, speech signals are extracted using independent component analysis.
TW. Lee, AJ Bell and R. Orglmeister. “Blind Source Separation of Real World Signals”, Proceedings of IEEE International Conference Neural Networks, June 97, Houston, pp 2129-2135.

  Speech signal extraction using independent component analysis is a technique that is expected to have high-precision extraction performance. However, the prior art has the following problems. First, a plurality of input signals are required. That is, an input signal “number of noise components + 1 (audio signal to be extracted)” is required. The number of noise components changes from moment to moment, which is unrealistic. Furthermore, the hardware becomes complicated as a plurality of input signals are processed.

  The present invention has been made in view of the above points, and uses an audio signal extraction method and an audio signal extraction method capable of extracting an audio signal from one input signal including an audio signal component and a noise component. An object of the present invention is to provide a voice recognition apparatus.

  In order to achieve the above object, in the audio signal extraction method according to claim 1, focusing on the fact that the audio signal and other noise signals are statistically independent, one input signal is converted into a plurality of filters. To decompose signal components (filter outputs) that are statistically independent of each other. Since speech and noise can be considered to be statistically independent, there is no signal component in which speech and noise are mixed among the decomposed signal components. Therefore, the extraction of the audio signal is realized by selecting the audio signal component and acquiring the audio signal from the selected signal component.

  When determining a plurality of filters, as described in claim 2, the plurality of filters can be determined using independent component analysis. Thereby, the output of the determined filter can be made statistically independent. As a specific example of this filter, a digital FIR (Finite Impulse Filter) filter is used as described in claim 3, or a digital IIR (Infinite Impulse Filter) filter is used as described in claim 4. can do.

  6. The audio signal extraction method according to claim 5, wherein N (N ≧ 1) filter outputs are selected as filter outputs for acquiring the audio signal in the order in which the filter outputs are separated from the Gaussian distribution. And Noise existing in the world generally has an amplitude distribution characteristic close to a Gaussian distribution. On the other hand, the audio signal has an amplitude distribution separated from the Gaussian distribution. Therefore, a filter output corresponding to the audio signal can be selected by selecting a specific number of signal components (filter outputs) in order from the signal component having the distribution farthest from the Gaussian distribution. Moreover, regardless of the Gaussian distribution, as described in claim 6, it is possible to select a filter output corresponding to an audio signal based on the audio feature amount of each filter output. And when there are a plurality of filter outputs selected in this way, as described in claim 7, the audio signal can be synthesized by taking the sum of the plurality of filter outputs.

  Claims 8 to 14 describe a speech recognition apparatus using the above-described speech signal extraction method. That is, a voice recognition unit that recognizes the acquired voice signal after the voice signal is acquired using the voice signal extraction method described above is provided. As described above, the accuracy of speech recognition can be improved by using the speech signal acquired using the speech signal extraction method described above for speech recognition.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a block diagram showing the configuration of the speech recognition apparatus 20 according to the present embodiment. The voice recognition device 20 extracts a voice signal using a voice signal extraction method described below, and recognizes the extracted voice signal.

  In FIG. 1, reference numeral 10 denotes a microphone, which generates a signal including ambient noise (noise) together with voice. Reference numeral 11 denotes a signal storage unit that stores an input signal from the current time point to a time point that is one second back from the signal generated by the microphone 10. The input signal is sampled at a sampling frequency of 10000 Hz, converted into a digital signal, and stored in the signal storage unit 11. Therefore, 10,000 digital signals are stored in the signal storage unit 11 as input signals from the current time point to a time point that is one second backward.

  The 10,000 digital signals are sent to the multiple filter determination unit 12 every second. The multiple filter determination unit 12 determines the coefficients of three FIR (Finite Impulse Response) filters having a length of 3, and inputs the 10,000 digital signals sent from the signal storage unit 11 using the coefficients. The filter output is calculated. The multiple filter determination unit 12 sends the calculated three FIR filter outputs and FIR filter coefficients to the filter output selection unit 13. The filter output selection unit 13 selects two filter outputs having a distribution away from the Gaussian distribution based on the non-approximation with the Gaussian distribution indicating a noise component.

  The audio signal synthesis unit 14 receives the two filter outputs selected by the filter output selection unit 13 and the coefficients specifying the two amplitudes, and synthesizes the audio signal. In this way, extracted speech signals having a length of 1 second are synthesized one after another.

  That is, an audio signal extraction method for extracting an audio signal from one input signal including an audio signal component and a noise component is performed by the entire processing performed on the input signal in the audio signal synthesis unit 14 from the signal storage unit 11. It is realized.

  The voice recognition unit 15 receives the voice signal synthesized by the voice signal synthesis unit 14 and performs voice recognition of the input voice signal. The speech recognition result in the speech recognition unit 15 is output to a processing unit that uses recognized speech.

  Hereinafter, the processing of each block will be described in detail. In the following formula and FIG. 2, the symbol N represents 10,000.

  The signal storage unit 11 creates an array mm (u) (u = 0, 1,..., 10000-1) from the input signal. The multiple filter determination unit 12 creates a signal vector x (u) represented by the following Equation 1 from the array mm (u) of the signal storage unit 11.

Next, a matrix W is created as shown in Equation 2 below, where the coefficients of the three FIR filters are W _ij (i: filter number, 0, 1, 2, j = 0, 1, 2).

Here, if y (u) is a vector whose elements are the three filter outputs y ₀ (u), y ₁ (u), and y ₂ (u) when mm (u) is input, y (u) Can be expressed by Equation 3 below.

Determining the filter coefficients so that the outputs are statistically independent of each other means that the elements of the vector y (u) (u = 2,3, ..., 10000-1) in Equation 3 are statistical In other words, “determining the matrix W so as to be independent”. W uses standard independent component analysis such as the Infomax algorithm (see Bell AJ and Sejnowski TJ 1995. “An information maximisation approach to blind separation and blind deconvolution”, Neural Computation, 7, 6, pp.1129-1159) To decide. FIG. 2 shows an example of a technique for determining the matrix W using the Infomax algorithm. Filter outputs y ₀ (u), y ₁ (u), y ₂ (u) are calculated from the matrix W thus determined.

The filter output selection unit 13 selects a filter output used for speech synthesis from the filter outputs y ₀ (u), y ₁ (u), and y ₂ (u) obtained by the multiple filter determination unit 12. First, after normalizing the average of the filter outputs y ₀ (u), y ₁ (u), y ₂ (u) to 0 and the variance to 1, the index g _i (i representing the distance from the Gaussian distribution of the filter output = 0,1,2) is calculated by the following Equation 4. (See A. Hyvarinen. “New Approximations of Differential Entropy for Independent Component Analysis and Projection Pursuit”, In Advances in Neural Information Processing Systems 10 (NIPS * 97), pp. 273-279, MIT Press, 1998.)

The index g _i takes a positive value, and the larger the value, the farther from the Gaussian distribution. The filter output selection unit 13 sends the two filter outputs having the largest and second largest values of the index g _i of the three filter outputs to the audio signal synthesis unit 14.

The audio signal synthesizer 14 synthesizes audio using the filter output selected by the filter output selector 13 and the matrix W obtained by the multiple filter determiner 12. Let the selected filter output be y ₀ (u), y ₁ (u). Further, assuming that the inverse matrix of the matrix W is A, the signal vector x (u) is expressed by the following Equation 5.

  When attention is paid to the first element of the signal vector x (u) in Expression 5, the following Expression 6 is established.

Since x ₀ (u) in Equation 6 is the original input signal itself, the input signal is the sum of three of A ₀₀ y ₀ (u), A ₀₁ y ₁ (u), and A ₀₂ y ₂ (u). It will be disassembled. Therefore, as shown in Formula 7, the audio signal is synthesized by taking the sum of A ₀₀ y ₀ (u) and A ₀₁ y ₁ (u).

  By the processing of each block described above, extracted speech signals having a length of 1 second are synthesized one after another, and the synthesized extracted speech signals are sent to the speech recognition unit 15. That is, according to the audio signal extraction method according to the present embodiment, an audio signal can be extracted from one input signal by independent component analysis using a plurality of filters.

  The present invention is not limited to the above-described embodiments, and various modifications can be made without departing from the spirit of the present invention.

  For example, in the above embodiment, the filter output selection unit 13 obtains a distance from each Gaussian distribution of each filter output, and selects a filter output having a large distance. However, it is also possible to calculate a feature value representing the likelihood of speech of each filter output y0 (u), y1 (u), y2 (u) and select a filter output that seems to be speech.

  In the above embodiment, the filter output selection unit 13 selects two filter outputs. However, the number of filters to be selected may be one, or when the number of filter outputs is four or more, Three or more filter outputs may be selected.

  Furthermore, in the above-described embodiment, the multiple filter determination unit 12 uses a digital FIR filter as a filter. However, for example, a digital IIR (Infinite Impulse Filter) filter may be used instead.

It is a block diagram showing the structure of the speech recognition apparatus by embodiment of this invention. It is explanatory drawing which shows an example of the method of determining the matrix W using an Infomax algorithm.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Microphone 11 Signal memory | storage part 12 Multiple filter determination part 13 Filter output selection part 14 Voice signal synthesis | combination part 15 Voice recognition part 20 Voice recognition apparatus

Claims (14)

Always storing one signal including an audio signal component and a noise component from a current time point to a finite past time point T;
Determining a plurality of filters so that the outputs are statistically independent of each other when the stored signal is input;
Selecting a filter output corresponding to the audio signal component from the outputs of the plurality of filters;
And obtaining an audio signal from the selected filter output.   2. The audio signal extraction method according to claim 1, wherein the step of determining the plurality of filters determines the plurality of filters using independent component analysis.   2. The audio signal extraction method according to claim 1, wherein the step of determining the plurality of filters determines a digital FIR (Finite Impulse Filter) filter as the plurality of filters.   2. The audio signal extraction method according to claim 1, wherein the step of determining the plurality of filters determines a digital IIR (Infinite Impulse Filter) filter as the plurality of filters.   2. The audio signal extraction method according to claim 1, wherein the filter output selecting step selects N (N ≧ 1) filter outputs in the order in which the filter outputs are separated from the Gaussian distribution. Method.   2. The audio signal extraction method according to claim 1, wherein the filter output selecting step selects N (N ≧ 1) filter outputs in accordance with the audio feature amount of each filter output. Method.   7. The audio signal extraction method according to claim 5 or 6, wherein the step of selecting the filter output selects a plurality of filter outputs, and the step of acquiring the audio signal takes a sum of the plurality of filter outputs. A voice signal extraction method characterized by synthesizing a voice signal. A signal input unit that receives one signal including an audio signal component and a noise component;
A signal storage unit that always stores one input signal from the signal input unit from a current time point to a finite past time point T;
When a signal stored in the signal storage unit is input, a plurality of filter determination units that determine a plurality of filters so that their outputs are statistically independent from each other;
A filter output selection unit that selects a filter output corresponding to an audio signal component from a plurality of filter outputs determined by the plurality of filter determination units;
An audio signal acquisition unit for acquiring an audio signal from the filter output obtained by the filter output selection unit;
A voice recognition apparatus comprising: a voice recognition unit that receives the voice signal acquired by the voice signal acquisition unit.   The speech recognition apparatus according to claim 8, wherein the plurality of filter determination units determine a plurality of filters using independent component analysis.   9. The speech recognition apparatus according to claim 8, wherein the filter determined by the plurality of filter determination units is a digital FIR (Finite Impulse Filter) filter.   9. The speech recognition apparatus according to claim 8, wherein the filter determined by the plurality of filter determination units is a digital IIR (Infinite Impulse Filter) filter.   9. The speech recognition apparatus according to claim 8, wherein the filter output selection unit selects N (N ≧ 1) filter outputs in order of separation of the filter output from the Gaussian distribution.   9. The speech recognition apparatus according to claim 8, wherein the filter output selection unit selects N (N ≧ 1) filter outputs according to a speech feature amount of the filter output.   14. The speech recognition apparatus according to claim 12, wherein the filter output selection unit selects a plurality of filter outputs, and the speech signal acquisition unit synthesizes speech signals from the plurality of filter outputs. Voice recognition device.

Images