KR101236539B1 - Apparatus and Method For Feature Compensation Using Weighted Auto-Regressive Moving Average Filter and Global Cepstral Mean and Variance Normalization - Google Patents

Abstract

The present invention is to provide a feature compensation apparatus and method using a weighted autoregressive moving average filter, global spectral averaging and variance normalization, which can prevent performance degradation of speech recognition systems in long-distance environments where noise and echoes exist. Receiving input training speech and recognition speech into a plurality of frames, extracting a training speech cepstrum and a recognition speech cepstrum with respect to the speech signal of each frame, and extracting the extracted training speech cepstrum And normalizing the training speech column and the recognized speech column with the global mean and variance of the recognized speech column, and in normalizing the recognized speech column, the first sentence to the last (n). Up to the first sentence, respectively, the global mean and the variance of the currently input cepstrum based on the normalized value of the cepstrum of the previous input. Continuously weighting, performing weighted autoregressive moving average filtering on the normalized cepstrum time series by obtaining weights of the normalized training speech and the presence or absence of speech per frame of the recognized speech cepstrum, and performing the weighting The HMM acoustic model is updated to maximize the likelihood of the HMM corresponding to the training speech cepstrum filtered out through the autoregressive moving average filtering, and weighted among the updated HMM acoustic models through Viterbi decoding. And selecting a recognition sentence that maximizes the likelihood of the HMM acoustic model for the autoregressive moving average filtered recognition speech cepstrum.

Description

Apparatus and Method For Feature Compensation Using Weighted Auto-Regressive Moving Average Filter and Global Cepstral Mean and Variance Normalization}

TECHNICAL FIELD The present invention relates to the field of speech signal processing technology, and more particularly, to a feature compensation algorithm of a novel method that can be applied to speech recognition in a remote environment in which various noises and echoes exist. The new feature compensation algorithm applies weights for the presence of speech to the existing autoregressive moving average filter based feature compensation method, and uses the global spectral mean and variance of the training and recognition data to generate Normalize

Recently, voice recognition technology has been applied to the real world beyond the demonstration level of the laboratory. However, the current speech recognition system shows relatively good performance in limited environment, but if it is applied to the actual recognition environment, the performance can be drastically degraded. This is because the actual environment in which voice recognition is performed involves factors that degrade recognition performance such as ambient noise, speech distance, microphone characteristics, channel distortion, and speaker variation.

This additional noise contaminates the speech signal and changes the feature vectors representing the speech. Thus, it causes variation of the statistical properties of the feature vector. For example, white noise reduces the dynamic range (or variance) of a feature vector, such as a cepstrum, representing envelope information of a spectrum.

The trained system shows good performance under similar conditions as when the actual recognizer is used. Therefore, voice recognition technology developers have made efforts to minimize the degradation of recognition performance caused by inconsistency between learning and recognition environment.

According to this research and development, the speech recognition system is classified into a training mode using clean voice and a recognition mode using noise mixed voice. In other words, if a speech distorted by noise and echo is input to the speech recognition system, the feature compensation algorithm compensates and extracts the feature. An acoustic model is constructed by applying the same feature compensation algorithm to a clear speech in advance, and speech recognition is performed using the acoustic model.

On the other hand, when the distance between the sound source and the microphone when the voice recognition, there is a problem that the recognition performance is degraded, because of the discrepancy between the training environment and the recognition environment in the voice recognition. Especially at long distances, there are echo components as well as additional noise, making the problem of environmental mismatch larger.

Various algorithms have been studied to solve this inconsistency problem, and can be classified into model compensation, noise cancellation, echo cancellation, and feature compensation.

Among these, the model compensation can compensate the acoustic model from the clean voice by using information of the recognition environment in advance, such as Parallel Model Combination (PMC) and Vector Taylor Series (VTS). This method can be used effectively if you have enough information of the recognition environment, but there is a limit to the actual use of this model method because it is usually difficult to predict the recognition environment.

On the other hand, the improvement of sound quality improves sound quality by estimating noise without prior knowledge of recognition environment. The improvement of sound quality is focused on the reduction of distortion caused by added noise. However, this method of improving sound quality can efficiently remove added noise, but it is difficult to remove echoes as well. Therefore, research to reduce echo was conducted separately. Among the echo cancellation algorithms, MSLP (Multi-Step Linear Prediction) shows excellent performance. However, MSLP has a disadvantage in that it has a large amount of calculation because it uses correlation of speech waveforms.

On the other hand, the feature compensation method using the autoregressive moving average filter shows a good performance with a small amount of calculation. This method enhances the characteristics by passing a time series of cepstrum normalized by cepstrum mean and variance through an autoregressive moving average filter. However, this has a problem that the characteristics of the noise section affects the characteristics of the speech section in the section that changes from noise to speech, which may be a factor that degrades the performance of the autoregressive moving average filter.

In addition, most speech recognition systems currently operate in a short-range environment with less added noise and less reflection. However, in order for the voice recognition system to be applied to various fields, it must work well in a remote environment. In general, as the distance between the sound source and the microphone increases, the additional noise and echo components increase, which degrades the speech recognition performance. In practice, voice recognition systems often have a limited amount of computation or only one microphone available.

Therefore, in order to increase the performance of speech recognition by lowering the false recognition rate, these problems are acting as a big problem, and the demands are increasing due to the interest and expansion of the application fields according to the recent speech recognition.

Therefore, the present invention has been made to solve the above problems, weighted autoregressive moving average filter and global spectral mean and variance normalization that can prevent the performance degradation of speech recognition system in long-distance environment with noise and echo It is an object of the present invention to provide a feature compensation apparatus and method using the same.

Features of the feature compensation device using the weighted autoregressive moving average filter and the global spectral mean and variance normalization according to the present invention for achieving the above object are divided into a plurality of frames by receiving a training voice and a recognition voice, respectively A MFCC (Mel-Frequency Cepstral Coefficients) feature extractor which performs Fourier transform on the speech signal of each frame and extracts a training speech cepstrum and a recognition speech cepstrum, and the training speech output from the MFCC feature extractor. A spectral mean and variance normalizer that normalizes the training speech cognition and the recognition speech cepstrum by the global mean and variance of the cepstrum and recognition speech cepstrum, respectively, and normalizes the recognized speech cepstrum in the normalization In this case, the first sentence to the last (n) sentence are respectively inputted based on the normalized value of the cepstrum of the previous input. An updater for continuously updating the global mean and the variance of the cepstruum, and whether the normalized training speech cepstrum and the recognized speech cepstrum output from the cepstrum mean and variance normalization unit A weighted autoregressive moving average filter for filtering the weighted autoregressive moving average of the normalized cepstrum time series by obtaining a weight and a training voice column output from the weighted autoregressive moving average filter are input. An acoustic model training unit for updating the HMM acoustic model to maximize the likelihood of the HMM corresponding to the rum, and to the recognition speech cepstrum output from the weighted autoregressive moving average filter through Viterbi decoding. Select a sentence of which maximum likelihood of the HMM acoustic model outputted from the acoustic model training unit There is composed by including a speech recognition.

Features of the feature compensation method using the weighted autoregressive moving average filter and the global spectral mean and variance normalization according to the present invention for achieving the above object are (A) receiving a plurality of input training voice and recognition voice, respectively Extracting a training speech column and a recognition speech column for the speech signal of each frame, and (B) a global average of the extracted training speech column and the recognized speech column; Normalizing the training speech cepstrum and the recognized speech cepstrum with each other and (C) normalizing the recognized speech cepstrum in the normalization, and inputting the previous sentence from the first sentence to the last (n) sentence, respectively. Continuously updating the global mean and variance of the currently input cepstrum on the basis of the normalized value of the cepstruum of (d), and (D) the normalized training voice and phosphorus. Performing weighted autoregressive moving average filtering on the normalized cepstrum time series by obtaining weights of the presence or absence of speech for each frame of the equation speech cepstrum, and (E) training voice output by filtering the weighted autoregressive moving average When the cepstrum is input, the acoustic model training step of updating the HMM acoustic model to maximize the likelihood of the HMM corresponding to the training speech cepstruum, and (F) through Viterbi decoding And a speech recognition step of selecting a recognition sentence of which the likelihood of the HMM acoustic model is maximum with respect to the weighted autoregressive moving average filtered recognition speech cepstrum among the updated HMM acoustic models.

Preferably, the step (B) comprises (b1) taking k (constant constants of 1 or more) samples for each of the input training voice and the recognition voice column, and configuring each frame into one frame; Discrete Fourier Transform (DFT) is performed for each frame, and then passed through a mel-scale triangular filter bank to yield N (constants of 1 or more) filter banks, which are discrete cosine transforms. Discrete Cosine Transform (DCT) to calculate the value of the cepstrum, and (B3) normalizing each value of the calculated cepstrum using cepstrum mean and variance normalization, respectively. It is done.

)을 수식

를 이용하여 정규화하는 것을 특징으로 한다.Preferably, the step (B3) is a value of the calculated cepstrum (

) Formula

It is characterized by using the normalization.

을 수식

로 갱신하고, 이를 다시 수식

로 변환하여 켑스트럼의 전역 평균을 산출하고, 수식

로 변환하여 켑스트럼의 전역 표준편차를 산출하는 단계와, (C2) 상기 산출된 켑스트럼의 평균과 표준편차를 이용하여 n번째 인식 대상 문장의 t번째 프레임에서의 k번째 켑스트럼

를 수식

를 이용하여 정규화하는 단계로 이루어지는 것을 특징으로 한다.Preferably, the step (C) is a global mean using the first to the last (n) sentences used for (C1) recognition

Formula

And update it again

To obtain the global mean of the cepstrum,

Calculating the global standard deviation of the cepstrum, and (C2) a k-th cepstrum in the t-th frame of the nth sentence to be recognized using the calculated mean and standard deviation of the cepstrum.

Formula

Characterized by using the step of normalizing.

를 적용하여 필터링을 수행하는 것을 특징으로 한다.Preferably in step (D) the weighted autoregressive moving average filtering is

It is characterized by performing filtering by applying.

As described above, the feature compensation apparatus and method using the weighted autoregressive moving average filter, the global spectral mean, and the variance normalization according to the present invention can be applied to the degree of presence of speech in the conventional autoregressive moving average filter based TMF scheme. By applying weights, we compensate for the disadvantages of TMF based on autoregressive moving average filter. It also has the effect of reducing the problem of environmental inconsistency by using global averages and variances from the entire spectrum of data.

In addition, the present invention can operate sufficiently in real time due to the small amount of calculation required, and also has an effect of greatly reducing the degradation of recognition performance in various environments in which additional noise and echo are present.

1 is a block diagram showing the structure of a feature compensation apparatus using a weighted autoregressive moving average filter and global spectral mean and variance normalization according to an embodiment of the present invention.
2 is a flowchart illustrating a feature compensation method using a weighted autoregressive moving average filter, global spectral mean, and variance normalization according to an embodiment of the present invention.

Other objects, features and advantages of the present invention will become apparent from the detailed description of the embodiments with reference to the accompanying drawings.

Referring to the accompanying drawings, a preferred embodiment of a feature compensation apparatus and method using a weighted autoregressive moving average filter according to the present invention, global cepstrum average and variance normalization will be described below. The present invention may, however, be embodied in many different forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided so that this disclosure will be thorough and complete, and will fully convey the scope of the invention to those skilled in the art. It is provided to let you know. Therefore, the embodiments described in this specification and the configurations shown in the drawings are merely the most preferred embodiments of the present invention and do not represent all the technical ideas of the present invention. Therefore, It is to be understood that equivalents and modifications are possible.

1 is a block diagram illustrating a structure of a feature compensation apparatus using a weighted autoregressive moving average filter, global spectral average, and variance normalization according to an exemplary embodiment of the present invention.

As shown in FIG. 1, the feature compensator includes a MFCC (Mel-Frequency Cepstral Coefficients) feature extractor 100, a cepstrum mean and variance normalizer 110, an updater 120, and a weighted autoregressive moving average. The filter 130, the acoustic model training unit 140, and the speech recognition unit 160 are configured.

The MFCC feature extractor 100 receives clean voices (hereinafter referred to as "training voices") and noise-mixed voices (hereinafter referred to as "recognition voices") as voice signals and separates them into a plurality of frames. Fourier transform is performed on the speech signal to extract the MFCC features of the training speech string and the recognized speech string, and output a frame sequence over time.

The cepstrum mean and variance normalization unit 110 trains the training speech cepstrum and the recognized speech 으로 as the global mean and variance of the training speech cepstrum and the recognized speech cepstrum output from the MFCC feature extractor 100. Normalize each strum.

The updater 120 normalizes the recognized speech cepstrum by the cepstrum mean and variance normalization unit 110, respectively, from the first sentence to the last (n) sentence, the cepstrum from previous input sentences, respectively. After updating the current global mean and variance based on the global mean and variance and the spectral mean and variance of the current input sentence, the updated global mean and variance of the updated recognized speech cepstrum is converted into the spectral mean and variance. Re-enter the normalization unit 110. The re-entered global mean and variance are used to normalize the cepstrum of the next input statement. This is because it is difficult to obtain the global mean and variance of the cepstrum in real time in real speech recognition, and it is possible to efficiently calculate the global average of the recognized speech cepstrum to normalize the characteristics of the recognized speech (recognition speech cepstrum). Do it.

The weighted autoregressive moving average filter 130 obtains weights for the presence or absence of speech for each frame of the normalized training speech cepstrum and the recognized speech cepstrum output from the cepstrum mean and variance normalization unit 110. Weighted autoregressive moving average filtering of the normalized cepstrum time series.

The acoustic model training unit 140 corresponds to the training speech cepstrum input from the acoustic models stored in the HMM acoustic model DB 150 when the training speech cepstrum output from the weighted autoregressive moving average filter 130 is input. The HMM acoustic model is updated to maximize the likelihood of the HMM acoustic model.

The speech recognition unit 160 outputs the HMM sound output from the acoustic model training unit 140 for the recognized speech cepstrums output from the weighted autoregressive moving average filter 130 through Viterbi decoding. Output the recognition sentences with the maximum likelihood of the model.

The operation of the feature compensation apparatus using the weighted autoregressive moving average filter, the global spectral mean, and the variance normalization according to the present invention configured as described above will be described in detail with reference to the accompanying drawings. The same reference numerals as those in Fig. 1 designate the same members performing the same function.

2 is a flowchart illustrating a feature compensation method using a weighted autoregressive moving average filter, a global spectral mean, and variance normalization according to an embodiment of the present invention.

Referring to FIG. 2, first, the MFCC feature extractor 100 receives a training voice, which is a clean voice, and a recognition voice, which is a noise-mixed voice, as a voice signal, and divides the voice into a plurality of frames. Fourier transform is performed to extract the MFCC features of the training speech cepstrum and the recognized speech cepstrum and output a frame sequence over time (S10).

Subsequently, the cepstrum mean and variance normalization unit 110 normalizes and outputs the MFCC features by the global mean and the variance of the training speech cepstrum and the recognized speech cepstrum output from the MFCC feature extractor 100. S20).

The normalization process of the cepstrum is described in detail as follows.

Prior to normalizing the training speech and recognition speech columns, k samples are taken for each training speech and recognition speech column to form one frame each. In this case, k / 2 samples are overlapped between neighboring frames. K is a constant of 1 or more.

Subsequently, a Discrete Fourier Transform (DFT) is performed for each successive frame, and then passed through a mel-scale triangular filter bank to produce N filter bank outputs, which are discrete cosine transforms. Transform: DCT) to calculate the spectral value.

는 다음 수학식 1과 같이 구한다. At this time, the value of the k th spectrum in the t th frame

Is obtained as in Equation 1 below.

는 t번째 프레임에서 b번째 멜-스케일의 삼각 필터뱅크 출력값이고, N은 필터뱅크의 차수이다.Where above

Is the triangular filterbank output of the bth mel-scale in the tth frame, and N is the order of the filterbank.

는 다음 수학식 2와 같이 구한다.And said

Is obtained as in Equation 2 below.

는 b번째 필터뱅크의 i번째 성분의 값을 나타내며,

는 b번째 필터뱅크가 가지는 성분의 개수이다. 그리고 상기

는 b번째 필터뱅크의 i번째 성분에 대응되는 입력음성의 t번째 프레임 DFT 계수 크기이다.At this time,

Denotes the value of the i th component of the b th filter bank,

Is the number of components of the b th filter bank. And said

Is the t-th frame DFT coefficient magnitude of the input voice corresponding to the i-th component of the b-th filter bank.

는 켑스트럼 평균 및 분산 정규화를 사용하여 다음 수학식 3과 같이 정규화시킨다.The value of the cepstrum

Is normalized as follows using Equation Mean and Variance Normalization.

)과 표준편차(

)는 각각 다음 수학식 4, 수학식 5와 같다.And the mean of the cepstrum (

) And standard deviation (

) Are as shown in Equations 4 and 5, respectively.

은 n번째 문장의 프레임 수이다. 따라서 상기

은 훈련 또는 인식 음성에서 사용한 전체 프레임의 개수이다.Where N is the total number of sentences used in training or recognition speech,

Is the number of frames in the nth sentence. Thus above

Is the total number of frames used in training or recognition speech.

On the other hand, since the entire sentence cannot be used in the actual recognition environment, when normalizing through the cepstrum mean and variance normalization unit 110, it is determined whether the input voice is a training voice or a recognized voice (S30). Normalization according to the cepstrum is further performed by updating the mean and the variance of the cepstrum in the same manner as in Equation 4 and Equation 5 using the sentences used in the previous recognition experiment through the update unit 120. Perform (S40).

The global mean and variance update method of the cepstrum in speech recognition are described in detail as follows.

을 다음 수학식 6과 같이 갱신할 수 있다.In other words, it is difficult to find the global mean and variance of the cepstrum in actual speech recognition, so the global mean using the first to last (n) sentences used for recognition

May be updated as in Equation 6 below.

In addition, by changing Equation 6 to Equation 7, it is possible to efficiently calculate the global average of the Cepstrum.

는 n번째 문장을 추가함에 따라 수학식 8과 같이 구한다.Where above

Is obtained as in Equation 8 by adding the nth sentence.

Where n 'and α are predefined constants obtained through experiments.

In addition, the global standard deviation can be obtained as in Equation 9 in the same manner as the above-described global mean.

를 아래 수학식 10과 같이 정규화시킬 수 있다. Using the mean and standard deviation of the cepstrum thus obtained, the k th cepstrum in the t th frame of the n th sentence to be recognized

Can be normalized as in Equation 10 below.

와

는 상기 수학식 4, 수학식 5에서 구한 값으로 초기화한다. At this time, the equations (7) and (9)

Wow

Is initialized to the values obtained in Equations 4 and 5 above.

When normalization is completed by the global mean and the variance of the training speech cepstrum and the recognized speech cepstrum through the cepstrum mean and variance normalization unit 110, the weighted autoregressive moving average filter 130 is normalized. Weighted autoregression moving average filtering is performed on the normalized cepstrum time series by obtaining weights of the presence or absence of the voice for each frame of the training speech and the recognized speech cepstrum as input signals (S50).

In this case, the weighted autoregressive moving average filter performs filtering by applying Equation 11 below.

는 가중 자동회귀 이동평균 필터(130)를 적용하여 t번째 프레임의 k번째 정규화된 켑스트럼을 보상한 결과이다.K = 1, ..., 12 denotes the order of the cepstrum, and

Is a result of compensating for the k th normalized spectral of the t th frame by applying a weighted autoregressive moving average filter 130.

W (t) is a weight for the presence or absence of speech in the t-th frame, and is represented by Equation 12 as a sigmoid function having a range of [0,1].

로서, 상기

는 t번째 프레임의 0번째 켑스트럼으로 에너지의 크기를 나타내며,

는 0번째 켑스트럼의 평균이다.In addition,

As above

Denotes the magnitude of energy as the 0 th cepstrum of the t th frame,

Is the mean of the zeroth cepstrum.

And α is a constant and determined through experiments.

As such, when the weighted automatic regression moving average filtering is completed (S50), it is determined whether the weighted automatic regression moving average filtered voice is a training voice or a recognized voice (S60).

As a result of the determination (S60), in the case of the training voice, the similarity of the HMM acoustic model corresponding to the training voice cepstrum input from among the acoustic models stored in the HMM acoustic model DB 150 through the acoustic model training unit 140. The HMM acoustic model is updated such that is maximized (S70).

The voice recognition unit 160 performs maximum Viterbi decoding to maximize the likelihood of the HMM acoustic model with respect to the weighted autoregressive moving average filtered recognition speech cepstrum among the updated HMM acoustic models. The recognition sentence is output (S80).

As such, by applying weights according to the presence or absence of speech in the TMF method based on the existing autoregressive moving average filter, the disadvantages of the TMF based on the autoregressive moving average filter are compensated.

Although the technical spirit of the present invention described above has been described in detail in a preferred embodiment, it should be noted that the above-described embodiment is for the purpose of description and not of limitation. In addition, those skilled in the art will understand that various embodiments are possible within the scope of the technical idea of the present invention. Accordingly, the true scope of the present invention should be determined by the technical idea of the appended claims.

Claims (7)

Mel-Frequency Cepstral Coefficients (MFCC) for receiving training speech and recognition speech, separating them into a plurality of frames, and performing Fourier transform on the speech signal of each frame to extract the training speech and recognition speech columns. Feature extraction section,
A spectral mean and variance normalizer which normalizes the training speech cepstrum and the recognized speech cepstrum by the global mean and variance of the training speech cepstrum and the recognized speech cepstrum output from the MFCC feature extractor;
When normalizing the recognized speech column in the normalization, the global average and the variance of the currently input column are continuously based on the normalized value of the column in the previous input from the first sentence to the last (n) sentence. An update unit to update,
A weighted autoregressive moving average filtering of the normalized cepstrum time series is performed by obtaining a weight of the normalized training speech cepstrum and the recognized speech cepstrum present in each frame of the cepstrum mean and variance normalization unit. A weighted autoregressive moving average filter,
An acoustic model training unit for updating the HMM acoustic model so that the likelihood of the HMM corresponding to the input training speech column is maximized when the training speech column is output from the weighted autoregressive moving average filter;
Viterbi decoding selects a sentence that maximizes the likelihood of the HMM acoustic model output from the acoustic model training unit with respect to the recognized speech cepstrum output from the weighted autoregressive moving average filter. A feature compensation device using a weighted autoregressive moving average filter, global spectral averaging and variance normalization, comprising a speech recognition unit. The method of claim 1,
The feature compensating device includes a weighted autoregressive moving average filter and a global cepstrum, comprising a HMM acoustic model database including HMMs (Hidden Markov Models) of phoneme units corresponding to word-by-word phonemes for recognition words. Feature compensation device using mean and variance normalization. (A) receiving input training voice and recognition voice, respectively, and separating them into a plurality of frames, and extracting the training voice string and the recognition voice string for the voice signal of each frame;
(B) normalizing the training speech cepstrum and the recognized speech cepstrum with the global mean and variance of the extracted training speech cepstrum and the recognized speech cepstrum, respectively;
(C) When normalizing the recognized speech cepstrum in the normalization, the global mean and the variance of the cepstrum currently input based on the normalized value of the cepstrum of the previous input from the first sentence to the last (n) sentence, respectively. Continuously updating
(D) performing weighted autoregressive moving average filtering on the normalized cepstrum time series by obtaining weights for the presence or absence of speech for each frame of the normalized training speech and the recognized speech cepstrum;
(E) updating the HMM acoustic model so that the likelihood of the HMM corresponding to the input training voice column is maximized when the weighted autoregressive moving average filtered and outputted training voice column is input;
(F) selecting a recognition sentence that maximizes the likelihood of the HMM acoustic model to the weighted autoregressive moving average filtered recognition speech cepstrum among the updated HMM acoustic models through Viterbi decoding; A feature compensation method using a weighted autoregressive moving average filter, global spectral mean, and variance normalization, comprising: a step; The method of claim 3, wherein step (B)
(B1) taking k (constant constants of 1 or more) samples for each of the input training speech and recognition speech cepstrum, and configuring each frame into one frame,
(B2) Discrete Fourier Transform (DFT) for each contiguous frame configured, and then pass through a mel-scale triangular filter bank to calculate N (constants of 1 or more) filter bank outputs. Calculating a value of the cepstrum by using a discrete cosine transform (DCT);
(B3) normalizing each value of the calculated cepstrum using a cepstrum mean and variance normalization, respectively, and performing a weighted autoregressive moving average filter and a global cepstral mean and variance normalization. Feature compensation method used. The method of claim 4, wherein step (B3)
The calculated cepstrum value (

) Formula

Normalize using,
At this time, the average of the cepstrum (

) ≪ / RTI &

, The standard deviation (

) Is the formula

Calculated using
Where N is the total number of sentences used in training or recognition speech,

Is the number of frames in the nth sentence, so

The weighted autoregressive moving average filter and the global spectral mean and variance normalization are characterized in that the number of total frames used in training or recognition speech. The method of claim 3, wherein step (C)
(C1) Global mean using the first to last (n) sentences used for recognition

Formula

And update it again

To obtain the global mean of the cepstrum,

Calculating the global standard deviation of the cepstrum,
(C2) k th cepstrum in the t th frame of the n th sentence to be recognized using the calculated cepstrum mean and standard deviation

Formula

Normalizing using
At this time,

Adds the nth sentence

And a weighted autoregressive moving average filter, a global spectral mean, and variance normalization, wherein n 'and are predefined constants. The method of claim 3, wherein
In step (D), the weighted autoregressive moving average filtering is

To apply the filtering,
In this case, k = 1, ..., 12 denotes the order of 켑 strum,

Is the result of compensating for the k th normalized spectral of the t th frame through weighted autoregressive moving average filtering.
The w (t) is a weight of the presence or absence of speech in the t-th frame, and is expressed as a sigmoid function having a range of [0,1].

Lt; / RTI >
remind

As above

Denotes the magnitude of energy as the 0 th cepstrum of the t th frame,

Denotes an average of the 0 th cepstrum, and α is a predefined constant, a weighted autoregressive moving average filter, and a global cepstrum mean and variance normalization.

Images