JP6783001B2 - Speech feature extraction algorithm based on dynamic division of cepstrum coefficients of inverse discrete cosine transform - Google Patents

Description

The present invention relates to a voice feature extraction technology field in which an unsupervised learning cluster analysis method is applied in a voice feature extraction direction, and particularly to a voice feature extraction algorithm based on dynamic division of the cepstrum coefficient of the inverse discrete cosine transform.

Speaker recognition technology includes both feature extraction and modeling recognition. Feature extraction is a key step in speaker recognition technology and directly affects the overall performance of the speech recognition system. Normally, after frame processing and windowing preprocessing for the audio signal, a high-dimensional amount of data is generated, and when extracting the characteristics of the speaker, the data is removed by removing the redundant information in the existing audio. It is necessary to lower the dimension of. In conventional methods, a triangular filter can be used to convert the speech signal into a speech feature vector that meets the requirements of the feature parameters, as well as to match perceptual traits that resemble the hearing of the human ear. It is possible to enhance the audio signal and suppress the non-audio signal to some extent.

The common feature parameters are as follows. The linear predictive analysis coefficient is a characteristic parameter obtained by simulating the principle of human pronunciation and analyzing the cascade model of the short tube, and the perceptual linear predictive analysis coefficient is calculated for spectral analysis based on the auditory model. It is a feature parameter that corresponds to the omnipolar model prediction polymorphism of LPC instead of the time region signal used for linear predictive coding LPC by applying and processing the input voice signal by the human ear-hearing model. The Boatleneck feature is a type 2 feature extracted by the neural network, and the Fbank feature based on the filter bank corresponds to the discrete cosine transform with the last step removed from the MFCC, and holds more initial voice data than the MFCC feature. did. The linearly predicted cepstrum coefficient is an important feature parameter that expresses formant features using dozens of cepstrum coefficients, while discarding the voice excitation information during the signal generation process based on the voice model, and the voice feature parameter MFCC. Is the most common cepstrum parameter, and as its extraction process, the cepstrum is first preprocessed by pre-emphasis, frame processing, windowing processing, fast Fourier transform, etc., and then the energy spectrum is extracted. After filtering with a set of Mel-scale triangular filter banks, calculating the logarithmic energy output by each filter bank, obtaining the MFCC coefficient by the discrete cosine transform (DCT), and obtaining the Mel-scale Cepstrum parameter. Extract the dynamic differential parameter, the mel cepstrum coefficient. Japanese Patent No. 3654831 discloses a method for extracting speech features based on an optimized wavelet of the basis. The method used unsupervised learning rather than cepstrum coefficients to analyze unlabeled signal data. In 2012, S.M. Such as Al-Rawahya (Alrawahy S; Hossen A; Heute U.Text-independent speaker identification system based on the histogram of DCT-cepstrum coefficients.International Journal of Knowledge-based and Intelligent Engineering Systems, 2012,16,141-161) is With reference to the MFCC feature extraction method, we submitted the Histogram DCT cepstrum coefficient method, which divides the DCT cepstrum coefficient obtained by preprocessing the voice into equal frequency domains. We have noticed that the dynamic characteristics of speech data are overlooked when the cepstrum coefficient is divided into equal frequency domains. Therefore, the present invention relates to S.A. Based on the method submitted by Al-Rawaya et al., A new speech feature extraction algorithm, that is, a method based on the dynamic division of the cepstrum coefficient of the inverse discrete cosine transform is provided. The present invention uses a hierarchical cluster method of unsupervised learning to perform cluster analysis on speech data based on the similarity of the dynamic features, thereby further explaining the speech characteristics. Extract the vector.

In conventional research, speech is matched in speaker mode by using MFCC as a speech feature vector and combining it with machine learning methods such as Gaussian mixed model (GMM), hidden Markov model (HMM), and support vector machine (SVM). Recognition technology is widely applied. The extraction process of MFCC is as follows. First, the audio signal is preprocessed by pre-emfasis, frame processing, windowing processing, and fast Fourier transform, and then the energy spectrum is filtered by a set of Mel-scale triangular filter banks, and each filter bank is used. The output logarithmic energy is calculated, the MFCC coefficient is obtained by the discrete cosine transform (DCT), the obtained logarithmic energy is introduced into the discrete cosine transform, the Mel-scale Cepstrum parameter is obtained, and then the dynamic differential. Extract the parameters, namely the mel cepstrum coefficient.

S. Al-Rawahya such as the 2012 study (Alrawahy S; Hossen A; Heute U.Text-independent speaker identification system based on the histogram of DCT-cepstrum coefficients.International Journal of Knowledge-based and Intelligent Engineering Systems, 2012,16, In 141-161), we discovered a new feature called DCT Cepstrum and submitted an algorithm for extracting voice features based on the constant frequency domain DCT Cepstrum coefficient. The preprocessed audio signal is converted to the frequency domain, that is, the preprocessed audio signal is converted from the time domain convolution to the form of multiplication of the spectrum of the frequency domain, the logarithm is taken, and the obtained components are added. The discrete cosine transform cepstrum coefficient (DCT Cepstrum coefficient) was obtained. The DCT cepstrum coefficient records the periodicity of the frequency range with a non-linear increment. That is, the characteristic section of the frequency domain is divided every 50 Hz between the frequency regions of 0 Hz to 600 Hz, and the characteristic section of the frequency domain is divided every 100 Hz in the frequency region of 600 Hz to 1000 Hz. The process can be regarded as a count of the number of cycles in the frequency range in a predetermined audio signal. It is simpler and faster than the MFCC feature extraction method.

Chinese Patent Application Publication No. 103354091 discloses a method and apparatus for extracting audio features based on frequency domain conversion. The method described in this patent document generates at least two segmented frequency domain signals by performing fractionation processing on the audio signal, and further, for the audio characteristics of each segmented frequency domain signal. The frequency domain conversion is performed to generate the conversion characteristics of each of the segmented frequency domain signals, and the conversion of each of the segmented frequency domain signals is based on the conversion characteristics of each of the segmented frequency domain signals. The high frequency components of the feature were acquired, thereby generating dynamic features to explain the melody characteristics of the audio signal, based on the high frequency components of the conversion feature of the at least two segmented frequency domain signals. By performing frequency domain conversion on an audio feature, a high frequency component of the modified feature after frequency domain conversion can be obtained, thereby extracting a dynamic feature for explaining the melody characteristic of the audio signal. This can be achieved, and the distinctiveness of high-frequency components of audio features can be improved. However, for audio signals, the present invention uses time to perform segmentation processing on audio, so that processing can be performed only on audio for 1s to 4s, and shorter audio is analyzed. In addition to being unable to do so, the invention removed elements with a low number of dimensions when extracting the feature vector, thus destroying the completeness of the audio feature. Therefore, the method described in the patent document is disadvantageous for extracting speech features.

An object of the present invention is mainly in view of the inaccuracy of the division frequency in the voice feature extraction algorithm based on the equal frequency domain division of the cepstrum coefficient of the inverse discrete cosine transform, and the dynamics of the cepstrum coefficient of the inverse discrete cosine transform. It is to provide an algorithm for extracting voice features based on target division.

The technical means of the present invention are as follows.

An algorithm for extracting speech features based on the dynamic division of the cepstrum coefficients of the inverse discrete cosine transform involves the following steps.

Step 1: Preprocess the audio signal.
Pre-emphasis, frame processing, and windowing processing are sequentially performed on the audio signal,
The pre-processing removes the effects of factors such as aliasing, higher harmonic distortion, and high frequencies from the human speech organ itself and the equipment that collects the audio signal on the quality of the audio signal, resulting in the signal obtained in the subsequent processing. It is guaranteed to be more uniform and smooth, can provide excellent parameters for extracting audio features, and improves the quality of subsequent processing.

Step 2: The preprocessed audio signal is subjected to format conversion processing from the time domain to the frequency domain.
The preprocessed audio signal is converted to the frequency domain, that is, the preprocessed audio signal is converted from the time domain convolution to the frequency domain spectral multiplication format, the logarithm is taken, and the obtained component is obtained. Shown in the addition format, the inverse discrete cosine transform cepstrum coefficient (IDCT Cepstrum coefficient) is obtained, and the specific process is performed according to the following equation (1).
C (q) = IDCT log | DCT {x (k)} | Equation (1)
In equation (1), DCT is the discrete cosine transform, IDCT is the inverse discrete cosine transform, x (k) is the input audio signal, i.e. the preprocessed audio signal, and C (q) is the output. The audio signal, the cepstrum coefficient of the inverse discrete cosine transform,
The cepstrum coefficient of the inverse discrete cosine transform is one data matrix, and when performing a hierarchical cluster with the existing frequency attributes of the voice, all the column attributes are the same, so the similarity of the adjacent column attributes is calculated. By doing so, clusters are performed in sequence.

Step 3: Using the cluster analysis method, calculate the similarity between the cepstrum coefficients of the inverse discrete cosine transform obtained in step 2, and sequentially merge the two adjacent classes with the highest similarity. , The cepstrum coefficient (DD-IDCT Cepstrum coefficient) of the inverse discrete cosine transform of the dynamic division obtained by repeating the above process so as to be clustered up to class 24 becomes a voice feature.

The pre-emphasis is realized by a digital filter, and the specific process is performed according to the following equation (2).
Y (n) = X (n) -aX (n-1) Equation (2)
In equation (2), Y (n) is a pre-emphasis output signal, X (n) is an input audio signal, a is a pre-emphasis coefficient, and n is a time.

The average power spectrum of the audio signal is affected by glottic excitation and mouth and nasal radiation, and the high frequency end is reduced by 6 dB / oct (octave) above about 800 Hz, and the higher the frequency, the more the corresponding component. As it becomes smaller, it is necessary to increase its high frequency portion before analyzing the audio signal.

"Short-time analysis technology" is applied throughout the entire process of voice analysis. The audio signal has a time-varying characteristic, but within a short time range (usually within a short time of 10 to 30 ms), the time-varying characteristic is basically unchanged, that is, relatively stable, and therefore. It may be considered a quasi-steady state process, i.e., the audio signal has short-term steady state. Therefore, the analysis and processing of any audio signal must always be based on a "short time", that is, a "short time analysis" must be performed and its characteristic parameters analyzed by the audio signal segment. Among them, each segment is called one "frame", and the length of one frame generally takes 10 to 30 ms. In this way, for the entire audio signal, what is analyzed is the time series of the feature parameters consisting of the feature parameters of each frame.

In the frame processing, the output signal after the pre-emphasis is divided so as to be 20 ms per frame.

After the frame is processed, it is windowed. It is recognized that the purpose of the windowing process is to make the entire audio signal more continuous, avoid the occurrence of the Gibbs phenomenon, and express the characteristics of a part of the periodic function in the originally non-periodic audio signal. The windowing process is a humming windowing process.

The conversion format is a cepstrum conversion.

The cluster analysis method is a hierarchical structure analysis method.

The calculation of the similarity is the calculation of the Euclidean distance.

The present invention has the following advantages over the prior art.

First, the present invention deeply analyzes the characteristics of the voice feature extraction algorithm by dividing the DCT Cepstrum coefficient into equal frequency domains, thereby converting the frequency domain without fully utilizing the voice dynamic features in the prior art. This can improve the drawbacks of doing so, the present invention has broader adaptability, and more accurate identification can be obtained in the identification of speakers.

Second, the present invention has the advantages of a simple process, fast speed, and low computational resources by applying unsupervised learning cluster analysis to the extraction of speech features.

In order to more clearly explain the technical means of the embodiment or the prior art of the present invention, the drawings in the description of the embodiment or the prior art will be briefly described below. The drawings below relate to embodiments of the present invention, and it will be apparent to those skilled in the art that other drawings can be obtained based on these drawings without the need for creative labor.

It is a flowchart of the speech feature extraction algorithm based on the dynamic division of the cepstrum coefficient of the inverse discrete cosine transform which concerns on embodiment of this invention. It is a tree graph of the cluster analysis which concerns on embodiment of this invention.

In order to more clearly explain the purpose, technical means and advantages of the examples of the present invention, the drawings in the examples of the present invention will be combined to clearly and completely explain the technical means in the examples of the present invention. It is clear that the examples described are not all examples, but only some examples according to the examples of the present invention. Based on the examples of the present invention, all other examples obtained by those skilled in the art without creative labor fall within the scope of the present invention.

As shown in FIG. 1, the speech feature extraction algorithm based on the dynamic division of the cepstrum coefficient of the inverse discrete cosine transform includes the following steps.

Step 1: Preprocess the audio signal.
Pre-emphasis, frame processing, and windowing processing are sequentially performed on the audio signal,
The pre-emphasis is realized by a digital filter, and the specific process is performed according to the following equation (2).
Y (n) = X (n) -aX (n-1) Equation (2)
In equation (2), Y (n) is the output signal after pre-emphasis, X (n) is the input audio signal, a is the pre-emphasis coefficient, n is the time, and the text. The value of a is 0.97.

The output signal after the pre-emphasis by the frame processing is divided so as to be 20 ms per frame.

The windowing process is a humming windowing process.

Step 2: The format of the preprocessed audio signal is converted from the time domain to the frequency domain.
The preprocessed audio signal was converted to the frequency domain, that is, the preprocessed audio signal was converted from the time domain convolution to the form of multiplication of the spectrum of the frequency domain, and the logarithm was obtained. The components are shown in an additive format, the cepstrum coefficient (IDCT Cepstrum coefficient) of the inverse discrete cosine transform is obtained, and the specific process is performed according to the following equation (1).
C (q) = IDCT log | DCT {x (k)} | Equation (1)
In equation (1), DCT is the discrete cosine transform, IDCT is the inverse discrete cosine transform, x (k) is the input audio signal, i.e. the preprocessed audio signal, and C (q) is the output. It is an audio signal, that is, the cepstrum coefficient of the inverse discrete cosine transform, and the conversion form is the cepstrum transform.

Step 3: Using the cluster analysis method, calculate the similarity between the cepstrum coefficients of the inverse discrete cosine transform obtained in step 2, and sequentially merge the two adjacent classes with the highest similarity into class 24. The above process is repeated so as to be clustered up to, and the cepstrum coefficient of the inverse discrete cosine transform of the dynamic division obtained becomes a voice feature, and the specific steps are as follows.

The matrix A represents the cepstrum coefficient of the n-dimensional inverse discrete cosine transform of m people obtained in step 2, and as shown in FIG. 2, the vectors V ₁ , V ₂ , of each dimension of the cepstrum coefficient of the inverse discrete cosine transform. ..., by regarding the _{V n} and n such, the Euclidean distance of the resulting _{V i} and _{V j} are
Will be.
The following are specific steps for cluster analysis.

1st cluster:
l ₁ = Dis (V ₁ , V ₂ )
l ₂ = Dis (V ₂ , V ₃ )
・ ・ ・
l _n-1 = Dis (V _n-1 , V _n )
If i = arg min (l ₁ , l ₂ , l ₃ , ..., l _n-1 ), the cluster result is as follows.
_{(V 1), (V 2} ), ..., (V i + V i + 1), ..., (V n), i.e.,

When updated, it is as follows
_{l i-1 = Dis (V} i-1, (V i + V i + 1))
_{l i = Dis ((V i} + V i + 1), V i + 2)
l _{i + 1} = l _{i + 2}
・ ・ ・
l _n-1 = l _n-2
Delete l _n-1

2nd cluster:
If j = arg min (l ₁ , l ₂ , l ₃ , ..., l _n-2 ), the cluster result is as follows.
_{(V 1), (V 2} ), ..., (V i + V i + 1), ..., (V j + V j + 1), ..., (V n), i.e.,

If you update again
l _j-1 = Dis (V _j-1 , (V _j + V _{j + 1} ))
l _j = Dis ((V _j + V _{j + 1} ), V _{j + 2} )
l _{j + 1} = l _{j + 2}
・ ・ ・
l _n-3 = l _n-2
Delete l _n-2

As described above, hierarchical clustering is performed until the result of the cluster is 24th class, and the obtained cepstrum coefficient of the inverse discrete cosine transform of the dynamic division becomes the voice feature, and the voice feature is introduced into the GMM model for identification. By doing so, the feasibility of the algorithm is determined.

The cluster analysis method is a hierarchical structure analysis method.

The calculation of the similarity is the calculation of the Euclidean distance.

Finally, it should be explained as follows. Each of the above examples is merely an explanation for the technical means of the present invention, and the scope of protection thereof is not limited to this scope. Although the present invention has been described in detail with reference to the above-described Examples, the technical means described in each of the above-described Examples may be modified, or the parts or all of the technical features may be switched equivalently. It will be understood by those skilled in the art that the essence of the corresponding technical means does not deviate from the scope of the technical means in the embodiment of the present invention even if the switching is performed.

(Additional note)
(Appendix 1)
Step 1 in which pre-emphasis, frame processing, and windowing processing are sequentially performed on the audio signal, and
The preprocessed audio signal is transformed into the frequency domain, i.e. the preprocessed audio signal is transformed from the time domain convolution into the form of multiplication of the frequency domain spectrum and its logarithm is obtained. The components are shown in an additive format, and the cepstrum coefficient of the inverse discrete cosine transform is obtained. The specific process is from the time domain to the frequency domain for the audio signal after preprocessing according to the following equation (1). Step 2 to perform the format conversion process to
C (q) = IDCT log | DCT {x (k)} | Equation (1)
In equation (1), DCT is the discrete cosine transform, IDCT is the inverse discrete cosine transform, x (k) is the input audio signal, i.e. the preprocessed audio signal, and C (q) is the output. The audio signal, the cepstrum coefficient of the inverse discrete cosine transform,
Using the cluster analysis method, the similarity between the cepstrum coefficients of the inverse discrete cosine transform obtained in step 2 is calculated, and the two adjacent classes with the highest similarity are sequentially merged into class 24. Step 3 where the cepstrum coefficient of the inverse discrete cosine transform of the dynamic division obtained by repeating the above process so as to be clustered up to is a voice feature,
A speech feature extraction algorithm based on the dynamic division of the cepstrum coefficients of the inverse discrete cosine transform, including.

(Appendix 2)
The pre-emphasis is realized by a digital filter, and the specific process is performed according to the following equation (2).
Y (n) = X (n) -aX (n-l) equation (2)
In equation (2), Y (n) is the output signal after pre-emphasis, X (n) is the input audio signal, a is the pre-emphasis coefficient, and n is the time. The extraction algorithm according to Appendix 1, which comprises the above.

(Appendix 3)
The extraction algorithm according to Appendix 1, wherein the frame processing is to divide the output signal after the pre-emphasis into 20 ms per frame.

(Appendix 4)
The extraction algorithm according to Appendix 1, wherein the windowing process is a humming windowing process.

(Appendix 5)
The extraction algorithm according to Appendix 1, wherein the format conversion is a cepstrum conversion.

(Appendix 6)
The extraction algorithm according to Appendix 1, wherein the cluster analysis method is a hierarchical structure analysis method.

(Appendix 7)
The extraction algorithm according to Appendix 1, wherein the calculation of the similarity is a calculation of the Euclidean distance.

Claims (4)

Step 1 in which pre-emphasis, frame processing, and windowing processing are sequentially performed on the audio signals of m speakers, and
The voice signal of the m speakers after preprocessing is converted into the frequency domain, that is, the voice signal of the m speakers after preprocessing is multiplied by the spectrum of the frequency domain from the time domain convolution. The step is to convert to the form of, take the logarithm, show the obtained components in the addition form, and obtain the cepstrum coefficient of the inverse discrete cosine transform. The specific process is preprocessing according to the following equation (1). In step 2, the format conversion process from the time domain to the frequency domain is performed on the voice signals of the m speakers after the process is performed.
C (q) = IDCT log | DCT {x (k)} | Equation (1)
In equation (1), DCT is the discrete cosine transform, IDCT is the inverse discrete cosine transform, and x (k) is the input audio signal, that is , the audio signal of the m speakers after preprocessing. , C (q) is the output audio signal, that is, the cepstrum coefficient of the inverse discrete cosine transform of the m speakers .
Using the cluster analysis method, the similarity between the two columns between the cepstrum coefficients of the inverse discrete cosine transform of the m speakers obtained in step 2 is calculated, and the adjacent two columns with the highest similarity are calculated. And repeat the above process so that they are clustered up to 24 columns , and the cepstrum coefficient of the inverse discrete cosine transform of the dynamic division obtained is the voice feature of the m speakers. ,
Including
The frame processing is to divide the output signal after the pre-emphasis into 20 ms per frame.
In step 3, the calculation of the similarity degree calculation der Euclidean distance is, is calculated by the following method,
A shown in Eq. (2) represents the cepstrum coefficient matrix of the n-dimensional inverse discrete cosine transform of m speakers obtained in step 2, and the column vector V ₁ of each dimension of the cepstrum coefficient of the inverse discrete cosine transform . V _2, ..., considers _{V n} and n such, Euclidean distance _{V i} and _{V j} of the resulting two rows,
Next,
In step 3, the process of clustering up to the 24 columns is performed by the following hierarchical clustering.
1st Column Cluster:
Equation (2)
l ₁ = Dis (V ₁ , V ₂ )
l ₂ = Dis (V ₂ , V ₃ )
・ ・ ・
l _n-1 = Dis (V _n-1 , V _n )
If i = arg min (l ₁ , l ₂ , l ₃ , ..., l _n-1 ), then the column cluster results are as follows:
_{(V 1), (V 2} ), ..., (V i + V i + 1), ..., (V n), i.e.,
Updating the Euclidean distances of two adjacent rows is as follows:
_{l i-1 = Dis (V} i-1, (V i + V i + 1))
_{l i = Dis ((V i} + V i + 1), V i + 2)
l _{i + 1} = l _{i + 2}
・ ・ ・
l _n-1 = l _n-2
Delete l _n-1
Part 2 Column Cluster:
If j = arg min (l ₁ , l ₂ , l ₃ , ..., l _n-2 ), the column cluster result is as follows.
_{(V 1), (V 2} ), ..., (V i + V i + 1), ..., (V j + V j + 1), ..., (V n), i.e.,
If you update the Euclidean distance of two adjacent columns again,
l _j-1 = Dis (V _j-1 , (V _j + V _{j + 1} ))
l _j = Dis ((V _j + V _{j + 1} ), V _{j + 2} )
l _{j + 1} = l _{j + 2}
・ ・ ・
l _n-3 = l _n-2
Delete l _n-2
Perform a hierarchical cluster until the result of the cluster is 24 columns,
A speech feature extraction algorithm based on the dynamic division of the cepstrum coefficient of the inverse discrete cosine transform, characterized in that the m people are a large number of people . The pre-emphasis is realized by a digital filter, and the specific process is performed according to the following equation ( 3 ).
Y (n) = X (n) -aX (n-l) equation ( 3 )
In equation ( 3 ), Y (n) is the output signal after pre-emphasis, X (n) is the input audio signal, a is the pre-emphasis coefficient, and n is the time. The extraction algorithm according to claim 1. The extraction algorithm according to claim 1, wherein the windowing process is a humming windowing process. The extraction algorithm according to claim 1, wherein the format conversion is a cepstrum conversion.