KR100791349B1 - Method and Apparatus for coding speech signal in Distributed Speech Recognition system - Google Patents

Abstract

According to the present invention, in a method of encoding a speech signal in a distributed speech recognition system, a log spectrum is generated by applying a log function to a speech signal subjected to mel filtering, and a long-term gain and a short-term excitation signal are generated using the generated log spectrum. And quantize the generated long-term gain and short-term excitation signal using a double codebook, and then transmit the feature vectors that cannot be corrected in the error correction code by using the property that the log spectrum has strong band-by-band correction. In addition, it is possible to improve the performance of speech recognition in a distributed speech recognition system having poor network conditions.

Distributed Speech Recognition, Encoding, Decoding, Quantization, Capstrum, Log Spectrum

Description

Coding method of speech signal in distributed speech recognition system and apparatus therefor {Method and Apparatus for coding speech signal in Distributed Speech Recognition system}

1 is a block diagram schematically showing the configuration of a distributed speech recognition system according to the present invention.

FIG. 2 is a block diagram illustrating in detail the DSR signal generator shown in FIG. 1; FIG.

3 is a block diagram illustrating in detail the quantization unit illustrated in FIG. 2;

4 is a flowchart illustrating a method for encoding a speech signal by a client in a distributed speech recognition system according to the present invention.

5 is a flowchart illustrating a method of decoding a speech signal by a server in a distributed speech recognition system according to the present invention.

<Explanation of symbols for the main parts of the drawings>

100: client 102: voice input unit

104: DSR signal generator 106: voice vocoder analyzer

108: adder 110: transmitter

150: server 152: receiving unit

154: log spectrum loss compensation unit 156: discrete cosine conversion unit

158: speech recognition unit 200: frame generation unit

210: preprocessor 220: Fourier transform unit

230: mel-filtering unit 240: log spectrum generation unit

250: quantization unit 300: long-term gain prediction unit

310: short excitation signal generator 320, 340: coupling unit

330: synthesis filter 350: error minimizing unit

The present invention relates to a method and apparatus for coding a speech signal in a distributed speech recognition system capable of clearly explaining the relationship between noise and speech by performing noise preprocessing of speech recognition in a log spectrum, which is a pre-capstrum stage, in a distributed speech recognition system. will be.

Distributed speech recognition systems separate feature extraction from feature matching and have two methods implemented by two different processing units disposed at two different locations to compensate for system performance degradation caused by transmitted speech. Will try. An example Distributed Speech Recognition (DSR) mobile or wireless communication system or client comprising a first communication device and a second mobile communication device extracts a feature vector from an input speech signal and encodes the feature vector.

The client then sends the encoded feature to the server via an error proof data channel.

The server then receives the encoded recognition features and performs only feature matching, i.e., the server matches the encoded features with the features of the known data set.

In the conventional distributed speech recognition frame as described above, all processes up to extracting the feature vector by transmitting the feature vector are necessarily coupled to the client part.

In general, the most commonly used feature vectors used for speech recognition are capstram feature vectors. The capturm feature vector has a relatively strong characteristic against variations in speaker and pronunciation speed. However, in an environment where actual speech recognition is used, noise of the surrounding environment, rather than the above variations, degrades the performance of speech recognition.

However, as described above, the capturm feature vector has a disadvantage in that it is difficult to form a mathematically clear relationship between the environmental noise and the voice.

In addition, since the noise preprocessing methods performed before the capstrum must be combined with the client portion requiring less computation and less memory, there is a problem that it is difficult to actually use algorithms capable of accurate noise processing while requiring a lot of computation.

In addition, since the independent component transmits a strong capstrum, when a specific frequency range is lost, only the result value of the error correction code is used to recover, and thus, even if the correction code is not restored, recognition cannot be performed.

Accordingly, an object of the present invention is to provide a method and apparatus for coding a speech signal in a distributed speech recognition system capable of clearly explaining the relationship between noise and speech by performing noise preprocessing of speech recognition in a log spectrum, which is a pre-capstrum step. have.

Another object of the present invention is to provide a method and apparatus for coding a speech signal in a distributed speech recognition system that can be easily recovered by using a log spectrum property when a signal of a specific band is lost.

According to an aspect of the present invention to achieve the above object, in a method of encoding a speech signal in a distributed speech recognition system, a log spectrum is generated by applying a log function to a speech signal subjected to mel filtering, and the generated A long-term gain and short-term excitation signal are generated using a log spectrum, quantized using a double codebook of the generated long-term gain and short-term excitation signal, and then transmitted. This is provided.

According to another aspect of the present invention, in a method of decoding a speech signal in a distributed speech recognition system, a log spectrum is received from a client, a lost band is compensated for the received log spectrum, and the lost band is compensated for. The present invention provides a method for coding a speech signal in a distributed speech recognition system, characterized in that a discrete cosine transform is performed on a converted log spectrum to convert to a capstrum and speech recognition is performed using the converted capstrum.

According to another aspect of the present invention, in the apparatus for encoding a speech signal in a distributed speech recognition system, a log spectrum generator for generating a log spectrum by applying a log function to a speech signal subjected to mel filtering, the log spectrum generation A long term gain and a short term excitation signal are generated using the log spectrum generated by the unit, and a quantization unit quantizes using the double codebook of the generated long term gain and short excitation signal, and a quantized log spectrum is transmitted by the quantization unit. Provided is a coding apparatus for a speech signal in a distributed speech recognition system comprising a transmitter.

According to another aspect of the present invention, in the apparatus for decoding a speech signal in a distributed speech recognition system, a receiver for receiving a log spectrum from a client and applying a loss characteristic theory to a log spectrum outputted from the receiver may determine a lost band. Compensating log spectral loss compensation unit, a discrete cosine transform unit performing a discrete cosine transform on the log spectrum compensated by the log spectral loss compensation unit to convert into a capstrum region, and using the capstrum transformed by the discrete cosine transform unit. In the distributed speech recognition system, a speech signal coding apparatus is provided.

Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a block diagram schematically showing the configuration of a distributed speech recognition system according to the present invention, FIG. 2 is a detailed block diagram showing the DSR signal generating unit shown in FIG. 1, and FIG. 3 is a detailed block diagram showing the quantization unit shown in FIG. It is also.

Referring to FIG. 1, a distributed speech recognition system recovers a loss band through correction of a log spectrum transmitted from a client 100 and a client 100 that quantizes a log spectrum of an input speech signal using a double codebook. Server 150 to perform voice recognition.

The client 100 may include a voice input unit 102, a DSR (Distributed Speech Recognition) signal generator 104, a voice vocoder analyzer 106, an adder 108, and a transmitter 110 that receive a voice from a user. Include.

The DSR signal generator 104 extracts spectral data regarding the voice input received through the voice input unit 102 and generates a coded signal representing the extracted spectrum data.

The DSR signal generator 104 will be described in more detail with reference to FIGS. 2 and 3.

Referring to FIG. 2, the DSR signal generator 104 may include a frame generator 200, a preprocessor 210, a Fourier transform 220, a mel-filter 230, and a log spectrum generator 240. And a quantization unit 250.

The frame generator 200 separates the input voice signal into a frame having a length of about 20-30 msec every 10 msec in order to obtain continuous short time spectrum information of the voice signal.

The preprocessor 210 performs preemphasis processing and hamming window processing on the voice signal of each frame. The pre-emphasis processing is an operation of emphasizing the high frequency components of the input signal, and the hamming window processing is an operation of multiplying the pre-emphasis processed speech signal by a window function to minimize discontinuity at each frame boundary.

The Fourier transform unit 220 outputs spectrum information by performing Fourier transform on the frame-specific speech signal passing through the preprocessor 210.

The mel filtering unit 230 sets a plurality of sections in the spectrum obtained through the Fourier transform, obtains energy for each section, and generates filter bank energy.

The log spectrum generator 240 generates a log spectrum by applying a log function to the generated filter bank energy.

The quantization unit 250 is a long-term gain and short-term gain generated by the technique of analysis-by-synthesis for the log spectrum prediction for clear transmission and efficient processing on the log spectrum generated by the log spectrum generator 240 Accurate quantization of the log spectrum is performed using the double codebook of the excitation signal.

The quantization unit 250 performing the above role will be described in more detail with reference to FIG. 3.

Referring to FIG. 3, the quantizer 250 includes a long-term gain predictor 300, a short-term excitation signal predictor 310, a first combiner 320, a synthesis filter 330, and a second combiner 340. And an error minimization unit 350.

The long-term gain predictor 300 predicts the long-term gain by using an analysis-by-synthesis technique on the log spectrum output from the log spectrum generator 240.

Looking at how the long-term gain predictor 300 predicts the long-term gain in more detail, in general, the relationship between speech and additional noise on the log spectrum is expressed by Equation 1.

......시간 도메인

...... time domain

......주파수 도메인

Frequency domain

......로그 스펙트럼 도메인

Log Spectrum Domain

Here, x (t), X (w), x is the voice, n (t), N (W), n is noise.

In addition, it is generally known that the statistical distribution of noise follows a minimum value among window values of a specific section. Algorithms based on these minimum-value theories perform better to track the real gain.

In order to accurately estimate the gain, several long-term frame values are considered and predicted so that the temporal correlation in different frames can be considered.

There are several algorithms to consider temporal correlation.

One of the least statistical methods is used to predict long-term gains as follows.

In the spectral domain, the power spectral density has the following relationship.

는 noisy power spectral density,

는 현재 입력 신호,

는 과거 입력신호,

,

는 시간, k는 주파수(frequency),

는 잡음의 power를 말한다. here,

Is the noisy power spectral density,

Is the current input signal,

Is the past input signal,

,

Is the time, k is the frequency,

Is the power of noise.

By taking a logarithmic function on the noisy power spectral density, Eq. (2), we obtain the long-term gain (g) as

The short term excitation signal predictor 310 may obtain a short term excitation signal using the long term gain obtained by the long term gain predictor 300.

)를 구할 수 있다. That is, the short-term excitation signal predictor 310 normalizes the log spectrum in each band by using the long-term gain obtained in the spectral domain of the log spectrum domain. Then, the short-term excitation signal prediction unit 310 is a short-term excitation signal (Equation 4)

) Can be obtained.

=

=

The short-term excitation signal such as Equation 4 is expressed as Equation 5.

는 현재 입력되는 신호를 말한다. Where

Refers to the signal currently being input.

According to another embodiment of the present invention, when making a codebook for the long-term gain, each codebook may be obtained using an EM algorithm such as Equation 6 through the relationship between the long-term gain and the short-term excitation signal.

The first combiner 320 combines the long-term gain obtained by the long-term gain predictor 300 with the short-term excitation signal obtained by the short-term excitation signal predictor 310 and transmits the short-term excitation signal to the synthesis filter 330.

The synthesis filter 330 receives the long-term gain signal and the short-term excitation signal transmitted from the first combiner 320 to generate an estimate of the input signal based on the quantized spectral coefficients and the combined excitation signal. The synthesis filter 330 then transmits an estimate of the generated input signal to the second combiner 340.

The second combiner 340 subtracts the input signal estimate output from the synthesis filter 330 from the log spectrum output from the log spectrum generator 240. The difference between the input log spectrum and the input signal estimate is applied to the error minimizer 350.

In addition, the second combiner 340 combines the long-term gain and the short-term excitation signal output from the synthesis filter 330 with the log spectrum output from the log spectrum generator 240 to generate an index signal for each. do.

The error minimizer 350 generates a cognitively weighted error signal based on a difference and a weighting function between an input signal estimate output from the second combiner 340 and the input log spectrum.

The cognitively weighted error signal is then used to determine an optimal set of excitation vector-related parameters that yield an optimal estimate of the input signal. The optimal set of quantized spectra and parameters is then delivered to server 150 via the transmitter 110.

As described above, the quantization unit 250 stores the long-term gain obtained by the long-term gain predictor 300 and the short-term excitation signal obtained by the short-term excitation signal predictor 310, respectively, by reflecting long-term and short-term statistics. Think of it as a bond.

As a result, the quantization unit 250 performs quantization using a double codebook in the form of an adaptive codebook as a technique for speech coding.

The voice vocoder analyzer 106 extracts additional data regarding the voice input that can be used for speech reconstruction.

The adder 108 combines the signal coded from the DSR signal generator 104 and the additional data extracted by the voice vocoder analyzer 106 into an integrated signal and transmits the combined signal to the transmitter 110. .

The transmitter 110 may be a radio frequency transmitter or a transceiver.

The signal transmitted through the transmitter 110 is transmitted to the server 150, the server 150 reconstructs the input voice signal using the log spectrum.

That is, the server 150 compensates for the damage to the log spectrum transmitted from the client 100, and then extracts a capstrum to perform voice recognition.

The server 150 may include a receiver 152, a log spectrum loss compensator 154, a discrete cosine converter 156, and a voice recognizer 158 that receive the log spectrum transmitted from the client 100. Include.

The log spectrum loss compensator 154 determines whether a lost band exists using the log spectrum received through the receiver 152. If the lost band exists, the received log is determined. The spectrum is used to recover the lost bands.

The log spectrum x received by the receiver 152 is expressed by Equation 7 below.

.

.

은 손실된 대역의 로그 스펙트럼,

는 손실되지 않은 대역의 로그 스펙트럼을 말한다. Where

Is the log spectrum of the lost band,

Is the log spectrum of the unlost band.

Assuming that we already know the correlation in the log spectrum of two codewords, the lost log spectrum is compensated by equation (8).

은 손실된 로그 스펙트럼이 복구된 스펙트럼,

는 평균을 말한다. Where

Is the spectrum from which the lost log spectrum has been recovered,

Says average.

As described above, the log spectrum loss compensator 154 recovers the lost log spectrum by applying a missing feature theory among the enhancement methods in the log spectrum received through the receiver 152.

이라고 하고 손상되지 않은 부분을

라고 가정하고 난 이후, 손상되지 않은 dimension과 correlation matrix를 이용하여 손상된 부분의 log spectrum을 복구한다. The missing feature theory measures the dimension of the missing part of the 23rd log spectrum.

And the undamaged part

After that, we restore the log spectrum of the damaged part using the undamaged dimension and correlation matrix.

The discrete cosine transformer 156 extracts a feature vector from the log spectrum of which the log spectrum lost by the log spectrum loss compensator 154 is compensated.

In other words, the discrete cosine transform unit 156 converts 23 log spectral values into a capstem region. At this time, only the first 13 values are calculated, and the remaining 10 values are not calculated. The discrete cosine transformer 156 then transmits the extracted feature vector to the speech recognizer 158.

Then, the speech recognition unit 158 performs speech recognition using the extracted feature vector and outputs the speech recognition.

4 is a flowchart illustrating a method of encoding a voice signal by a client in a distributed speech recognition system according to the present invention.

Referring to FIG. 4, the client generates a log spectrum by applying a log function to the filter bank energy generated through mel filtering (S400).

After performing step 400, the client predicts a long term gain using the generated log spectrum (S402), and predicts a short term excitation signal using the predicted long term gain (S404).

That is, the client predicts a long-term gain by considering the temporal correlation in different frames with respect to the generated log spectrum. The client then uses the predicted long term gain to normalize the log spectrum in each band to predict short term excitation signals.

After performing step 404, the client performs quantization using the double codebook of the long-term gain and the short-term excitation signal (S406). That is, the client quantizes the long-term gain and the short-term excitation signal as a combination of independent parameters reflecting long-term and short-term statistics, respectively, and quantizes the combination between the long-term and short-term parameters using an adaptive codebook type dual codebook.

The quantized log spectral value is transmitted to a server.

A method of recognizing a voice signal using the log spectrum transmitted from a client by the server will be described with reference to FIG. 5.

5 is a flowchart illustrating a method of decoding a speech signal by a server in a distributed speech recognition system according to the present invention.

Referring to FIG. 5, when a log spectrum is received from a client (S500), the server determines whether there is a band lost in the log spectrum (S502).

If there is a lost band in the log spectrum as a result of the determination in step 502, the server recovers the lost band using the received log spectrum (S504).

That is, the server recovers the lost log spectrum by applying missing feature theory among the enhancement methods in the log spectrum.

After performing step 504, the server extracts the capstrum by applying the log spectrum recovered in step 504 to the received log spectrum (S506).

After performing step 506, the server performs speech recognition using the extracted capstrum (S508).

If a lost band does not exist in the received log spectrum as a result of the determination of step 502, the server extracts a capstrum from the received log spectrum (S510) and performs step 508.

The method of the present invention as described above may be implemented in a program and stored in a recording medium in a computer-readable form. Since this process can be easily carried out by those skilled in the art will not be described in more detail.

The present invention is not limited to the above embodiments, and many variations are possible by those skilled in the art within the spirit of the present invention.

As described above, according to the present invention, the performance of speech recognition in a distributed speech recognition system in which network conditions are poor by correcting feature vectors that cannot be corrected in an error correction code by using a property of strong band-by-band correction. The present invention provides a method and apparatus for coding a speech signal in a distributed speech recognition system.

In addition, according to the present invention, a method of coding a speech signal in a distributed speech recognition system capable of transmitting an efficient and accurate log spectrum instead of a capstrum in a distributed speech recognition system frame by modeling the log spectrum as a long-term gain and a short-term excitation signal. And the apparatus.

In addition, according to the present invention, there is provided a method and apparatus for coding a speech signal in a distributed speech recognition system capable of clearly explaining the relationship between noise and speech by performing noise preprocessing of speech recognition in a log spectrum that is a pre-capstrum step. Can be.

In addition, according to the present invention, it is possible to provide a method and apparatus for coding a speech signal in a distributed speech recognition system that can be easily recovered by using a log spectrum property when a signal of a specific band is lost.

Claims (11)

In a speech signal encoding method in a distributed speech recognition system, Generating a log spectrum by applying a log function to the voice signal on which the mel filtering is performed; Generating a long term gain and a short term excitation signal using the generated log spectrum; and Quantizing using the generated long-term gain and the short-term excitation signal using a double codebook, and then transmitting Method of coding a speech signal in a distributed speech recognition system comprising a. The method of claim 1, The long term gain (g) is

Saved by here,

Is the noisy power spectral density

Saved by

Is the current input signal,

Is the past input signal,

,

Is the time, k is the frequency,

The coding method of the speech signal in a distributed speech recognition system, characterized in that the power of the noise. The method of claim 1 The short excitation signal (

) Is formed by normalizing the log spectrum in each band using the long-term gain,

=

The method of coding a speech signal in a distributed speech recognition system, wherein g is a long term gain and z (w) is an input signal. The method of claim 1, The quantization of the generated long-term gain and the short-term excitation signal using a double codebook is performed by an analysis-by-synthesis technique. A method of decoding a speech signal in a distributed speech recognition system, Receiving a log spectrum from a client; Compensating for the lost band for the received log spectrum; Performing discrete cosine transform on the log spectrum compensated for the lost band to convert to a capstrum; and Performing speech recognition using the converted capstrum Method of coding a speech signal in a distributed speech recognition system comprising a. The method of claim 5, The method of coding a speech signal in a distributed speech recognition system, characterized in that the compensation for the lost band for the received log spectrum is compensated using a missing feature theory. The method of claim 5, Compensation of the lost band (

)silver

Made by Where

Is the log spectrum of the lost band,

Is the log spectrum of the unlost band,

The coding method of the speech signal in a distributed speech recognition system, characterized in that the average. An apparatus for encoding a speech signal in a distributed speech recognition system, A log spectrum generator for generating a log spectrum by applying a log function to the voice signal on which the mel filtering is performed; A quantization unit generating a long term gain and a short term excitation signal using the log spectrum generated by the log spectrum generator and quantizing the generated long term gain and the short term excitation signal using a double codebook; Transmitter for transmitting the quantized log spectrum in the quantization unit Apparatus for coding a speech signal in a distributed speech recognition system comprising a. The method of claim 8, The quantization unit A long term gain predicting unit predicting a long term gain using the log spectrum generated by the log spectrum generating unit; A short-term excitation signal predictor for generating a short-term excitation signal by normalizing a log spectrum in each band using the long-term gain predicted by the long-term gain predictor; A first combiner combining the long term gain generated by the long term gain predictor and the short term excitation signal generated by the short term excitation signal predictor; A second combiner configured to combine the long-term gain and the short-term excitation signal output from the first combiner with the log spectrum output from the log spectrum generator to generate an index signal for each; And an error minimizer for minimizing an error of a signal output from the second combiner and feeding back the long-term gain predictor and the short-term excitation signal predictor. The method of claim 8, And the transmitter is one of a radio frequency transmitter and a transceiver. An apparatus for decoding a speech signal in a distributed speech recognition system, Receiving unit for receiving the log spectrum from the client; A log spectrum loss compensator for compensating a lost band by applying a loss characteristic theory to the log spectrum output from the receiver; A discrete cosine transform unit performing a discrete cosine transform on the log spectrum compensated by the log spectrum loss compensator and converting the transformed signal into a capstrum region; and A speech recognition unit for performing speech recognition using the capstrum transformed by the discrete cosine transform unit Apparatus for coding a speech signal in a distributed speech recognition system comprising a.

Images