KR100794140B1 - Apparatus and Method for extracting noise-robust the speech recognition vector sharing the preprocessing step used in speech coding - Google Patents

Abstract

1. TECHNICAL FIELD OF THE INVENTION

The present invention relates to an apparatus and method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal.

2. The technical problem to be solved by the invention

According to the present invention, a speech encoding preprocessing process and a speech feature vector extraction preprocessing process are used to extract a speech feature vector (speech feature parameter) for speech recognition in a terminal such as a mobile phone equipped with a user speech coding function. It is an object of the present invention to provide an apparatus and method for extracting a robust speech feature vector to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal.

3. Summary of Solution to Invention

An apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, the apparatus comprising: a high pass filter for removing a low frequency band signal of a speech signal received from the outside; A frequency domain converter for converting a signal from which the low frequency band is removed in the high pass filter into a spectral signal in a frequency domain; A channel energy estimator for calculating a channel energy estimate for the current frame of the spectral signal converted by the frequency domain converter; A channel signal to noise ratio estimator for estimating a channel signal to noise ratio for the speech signal based on the channel energy estimate calculated by the channel energy estimator and the background noise energy estimated by the background noise estimator; The background noise estimator for calculating and updating a background noise energy estimate for the speech signal according to a command of a noise update determiner; A voice metric calculator for calculating a sum of voice metrics on a channel with respect to the voice signal based on the channel signal to noise ratio estimated by the channel signal to noise ratio estimator; A spectral deviation estimator for estimating a spectral deviation of the speech signal based on the channel energy estimate calculated by the channel energy estimator; The noise update determiner that issues a noise estimate update command based on a total channel energy estimate estimated by the spectrum deviation estimator and a difference value between an estimated value for the current power spectrum and an estimated long term power spectrum; A channel signal to noise ratio estimation value corrector for correcting the channel signal to noise ratio estimated by the channel signal to noise ratio estimator based on the sum of the voice metrics calculated by the voice metric calculator; A channel gain calculator for calculating a linear channel gain based on the channel signal-to-noise ratio modified by the channel signal-to-noise ratio estimation modifier and the background noise energy estimate calculated by the background noise estimator; A frequency domain filter applying the linear channel gain calculated by the channel gain calculator to the spectrum signal converted by the frequency domain converter; And a time domain converter for converting a spectral signal to which a linear channel gain is applied in the frequency domain filter into a voice signal in the time domain.

4. Important uses of the invention

The present invention is used in the field of distributed speech recognition.

Distributed speech recognition system, terminal, speech coding preprocessing, speech coding, speech recognition, preprocessing sharing, noise reduction, speech feature vector extraction

Description

Apparatus and Method for extracting noise-robust the speech recognition vector sharing the preprocessing step used in speech coding}

1 is a block diagram of an apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal according to the present invention.

FIG. 2 is a detailed block diagram of an embodiment of the speech coding / recognition preprocessor of FIG. 1. FIG.

3A is a diagram illustrating an embodiment of an extended speech coding / recognition preprocessor for 11 KHz speech signal processing according to the present invention;

Figure 3b is an embodiment configuration for an extended speech coding / recognition preprocessor for 16 KHz speech signal processing according to the present invention.

4 is a diagram illustrating an embodiment of a speech recognition process used in the present invention.

FIG. 5 is a flow chart of an embodiment of a learning process for generating an acoustic model in FIG. 4. FIG.

6 is a graph showing speech recognition performance using speech feature vectors extracted by applying the present invention.

* Explanation of symbols on the main parts of the drawing

100: speech feature vector extraction module

150: voice coding module

11: speech coding / recognition preprocessor 12: speech feature vector extractor

13 voice compression unit 14 bit stream transmission unit

21 high pass filter 22 frequency domain converter

23 channel energy estimator 24 channel signal to noise ratio estimator

25: Voice Metric Calculator 26: Spectral Deviation Estimator

27: Noise Update Determinant 28: Channel Signal-to-Noise Ratio Estimator Corrector

29: channel gain calculator 30: background noise estimator

31: frequency domain filter 32: time domain converter

The present invention relates to an apparatus and method for extracting a speech feature vector from a distributed speech recognition terminal, and more particularly, to a speech feature vector (voice feature parameter) for speech recognition in a terminal such as a mobile phone having a user speech encoding function. The speech feature pre-processing process and the speech feature vector extraction preprocessing process are used to extract the speech feature vectors robust to the noise caused by the communication environment. It relates to an apparatus and a method for extracting.

Distributed Speech Recognition (DSR) is a technology that enables the speech recognition function to be implemented in low-end terminals such as mobile phones, and recognizes the characteristics of voice signals in low-end terminals and provides them from low-end terminals in high-end speech recognition servers. It consists of a dual processing system that performs speech recognition based on the received speech features.

In general, in the speech recognition field as described above, MFCC (Mel-Frequency Cepstrum Coefficient) is commonly used. The MFCC is a sine wave component representing the form of the frequency spectrum expressed in mel scale, and refers to a speech feature vector (also called speech feature parameter) representing a speech received from a user.

In the terminal side, the voice feature vector is extracted from the voice input from the user through the MFCC, and the voice feature vector is loaded in the bit stream and transmitted to the voice recognition server in order to process the voice feature vector into a form suitable for transmission to the communication network. For example, a MFCC corresponding to a voice received from a user is selected from a vector having the closest distance among center vectors of a codebook of a finite number and transmitted as bit stream data. Here, the codebook has a center value for a cluster having a similar value with respect to the voice spoken by the user, and is typically the most representative value among the extracted training data after extracting the training data (learning MFCC vector) from a large number of voice data. Is selected.

The speech recognition server side dequantizes the speech feature vector carried in the bit stream received from the terminal, and recognizes a word corresponding to the speech using the HMM (Hidden Markov Model) as the speech model. Here, the HMM is a process of modeling a basic unit, for example, a phoneme, for speech recognition. The HMM combines phonemes coming into the speech recognition engine and phonemes registered in a DB in the speech recognition engine to form words and sentences.

On the other hand, in recent years, in accordance with the trend of digital convergence, a mobile phone has been in the spotlight as a distributed voice recognition terminal, and such a mobile phone is equipped with a user voice signal processing function, for example, a voice encoder.

However, as described above, in order to extract the voice feature vector for the user's spoken voice, the user's voice signal preprocessing process, in particular, the noise attenuation process is required. The user voice signal preprocessing process is performed separately.

For example, the processing of a user voice signal is the same for voice calls and voice recognition, but the mobile phone processes it as a separate preprocessing process, especially as a separate preprocessing device, which requires additional memory and computational amount in a mobile phone having low specification characteristics. This is causing a waste of resources in mobile phones.

In addition, a speech preprocessing process such as speech encoding has a delay internally in the terminal, which causes a delay in switching between a voice call process and a speech recognition process in a mobile phone. For example, if a call is received while a user is performing a voice recognition function through a mobile phone, there is a time delay in receiving the call.

Hereinafter, a preprocessing process for a voice call and a preprocessing process for voice recognition in a conventional mobile phone will be described.

The conventional mobile phone is equipped with a speech coding module (speech coding module) and a speech feature vector extraction module (distributed speech recognition front-end module).

The speech coding module consists of a pre-processing for speech coding, a model parameter estimation, a speech compressor and a bit stream.

The speech feature vector extraction module comprises a pre-processing for speech recognition, a speech feature vector extractor (MFCC front-end), a speech compressor and a bit stream transmitter.

As described above, in the conventional method, as a component for attenuating noise mixed in a user's voice signal, a speech encoding preprocessor and a speech recognition preprocessor are separately installed in a mobile phone, which is used in a speech encoding process and a speech feature vector extraction process. This is due to the different preprocessing signals.

Therefore, in view of performing the same functions as the speech encoding preprocessor and the speech recognition preprocessor, there is an urgent need for a technology capable of integrating the preprocessing process in the speech encoding process and the speech feature vector extraction process. .

SUMMARY OF THE INVENTION The present invention has been proposed to solve the above problems and meet the above demands, and to extract a speech feature vector (speech feature parameter) for speech recognition from a terminal such as a mobile phone equipped with a user speech encoding function. An apparatus for extracting a robust speech feature vector to noise by sharing the preprocessing of a speech coder in a distributed speech recognition terminal by sharing a speech coding preprocessing process and a speech feature vector extraction preprocessing process to extract a speech feature vector robust to noise caused by a communication environment, etc. The purpose is to provide a method.

The apparatus of the present invention for achieving the above object is a device for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech encoder in a distributed speech recognition terminal, to remove the low frequency band signal of the speech signal received from the outside A high pass filter; A frequency domain converter for converting a signal from which the low frequency band is removed in the high pass filter into a spectral signal in a frequency domain; A channel energy estimator for calculating a channel energy estimate for the current frame of the spectral signal converted by the frequency domain converter; A channel signal to noise ratio estimator for estimating a channel signal to noise ratio for the speech signal based on the channel energy estimate calculated by the channel energy estimator and the background noise energy estimated by the background noise estimator; The background noise estimator for calculating and updating a background noise energy estimate for the speech signal according to a command of a noise update determiner; A voice metric calculator for calculating a sum of voice metrics on a channel with respect to the voice signal based on the channel signal to noise ratio estimated by the channel signal to noise ratio estimator; A spectral deviation estimator for estimating a spectral deviation of the speech signal based on the channel energy estimate calculated by the channel energy estimator; The noise update determiner that issues a noise estimate update command based on a total channel energy estimate estimated by the spectrum deviation estimator and a difference value between an estimated value for the current power spectrum and an estimated long term power spectrum; A channel signal to noise ratio estimation value corrector for correcting the channel signal to noise ratio estimated by the channel signal to noise ratio estimator based on the sum of the voice metrics calculated by the voice metric calculator; A channel gain calculator for calculating a linear channel gain based on the channel signal-to-noise ratio modified by the channel signal-to-noise ratio estimation modifier and the background noise energy estimate calculated by the background noise estimator; A frequency domain filter applying the linear channel gain calculated by the channel gain calculator to the spectrum signal converted by the frequency domain converter; And a time domain converter for converting the spectral signal to which the linear channel gain is applied in the frequency domain filter into a voice signal in the time domain.

Meanwhile, another apparatus of the present invention is a device for extracting a speech feature vector that is robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal. Speech coding means for being transmitted; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; And a speech coding / recognition preprocessor for attenuating noise mixed in the speech signal received from the outside, wherein the speech signal inputted to the speech coding means and the speech signal inputted to the speech feature vector extracting means are speech coding / recognition. It is characterized in that the pretreatment in the pretreatment unit.

Meanwhile, another apparatus of the present invention is a device for extracting a speech feature vector that is robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal, wherein the speech signal encoded in the voice call mode is externally transmitted through a speech traffic channel. Speech coding means for being transmitted to the network; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; A frequency down sampler for down sampling the frequency band of the audio signal received from the outside; And a speech coding / recognition preprocessor for attenuating noise mixed in the down-sampled speech signal by the frequency down sampler, wherein the speech signal inputted to the speech coding means and the speech signal inputted to the speech feature vector extracting means include: It is characterized in that the pre-processing in the speech coding / recognition preprocessor.

Meanwhile, another apparatus of the present invention is a device for extracting a speech feature vector that is robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, wherein the speech signal encoded in the voice call mode is externally transmitted through a speech traffic channel. Speech coding means for being transmitted to the network; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; A low pass quadrature mirror filter for passing a low frequency band signal to a voice signal received from an outside; A high pass quadrature mirror filter for passing a high frequency band signal to the voice signal received from the outside; And a speech coding / recognition preprocessor for attenuating noise mixed in the speech signal passed by the low pass quadrature mirror filter, wherein the speech signal inputted to the speech coding means and the speech inputted to the speech feature vector extracting means. The signal is preprocessed by the speech coding / recognition preprocessor.

In the method of the present invention, a method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal, the method comprising: removing a low frequency band signal of a speech signal received from an external device; Converting the signal from which the low frequency band has been removed into a spectral signal in a frequency domain; Calculating a channel energy estimate for the current frame of the transformed spectral signal; Estimating a spectral deviation of the speech signal based on the calculated channel energy estimate; Issuing a noise estimate update command based on the estimated total channel energy estimate and the difference between the estimated value for the current power spectrum and the estimated value for the average long-term power spectrum; Calculating and updating a background noise energy estimate for the speech signal when the noise estimate update command is received; Estimating a channel signal to noise ratio for the speech signal based on the calculated channel energy estimate and background noise energy estimate; Calculating a sum of speech metrics on a channel with respect to the speech signal based on the estimated channel signal to noise ratio; Modifying the estimated channel signal to noise ratio based on the sum of the calculated speech metrics; Calculating a linear channel gain based on the modified channel signal to noise ratio and the calculated background noise energy estimate; Applying the calculated linear channel gain to the transformed spectral signal; And converting the spectral signal to which the linear channel gain is applied to an audio signal in the time domain.

The above objects, features and advantages will become more apparent from the following detailed description taken in conjunction with the accompanying drawings, whereby those skilled in the art may easily implement the technical idea of the present invention. There will be. In addition, in describing the present invention, when it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the gist of the present invention, the detailed description thereof will be omitted. Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the accompanying drawings.

1 is a configuration diagram of an apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal according to the present invention (hereinafter, referred to as a "voice feature vector extraction apparatus").

A distributed speech recognition terminal (eg, a mobile phone, etc.) equipped with a speech feature vector extracting apparatus according to the present invention includes a speech coding module 150 and a speech feature vector extraction front. -end module (100), as shown in FIG. 1, that the speech coding preprocessor of the speech coding module and the speech recognition preprocessor of the speech feature vector extraction module are shared with each other by the speech coding / recognition preprocessor (11). Able to know.

That is, in the mobile phone according to the present invention, the voice feature vector extraction module 100 includes a voice coding / recognition preprocessor 11, a voice feature vector extraction unit (MFCC front-end) 12, and a voice compression unit 13. ) And a bit stream transmission unit (bit stream) 14.

In the mobile phone according to the present invention, the voice coding module 150 includes a voice coding / recognition preprocessor 11, a speech coder 15, a voice compressor 16, and a bit stream transmitter ( bit stream).

Of course, the mobile phone according to the present invention is equipped with a switch 50 for transitioning between the voice call mode and the voice recognition mode, and according to the operation of the switch 50, the encoded user voice signal in the voice call mode is The voice traffic channel is transmitted to the mobile communication system, and in the voice recognition mode, the extracted user voice feature vector is transmitted to the voice recognition server through the wireless data channel.

In particular, the speech coding / recognition preprocessor 11 performs a function of attenuating noise mixed with a user speech signal (input speech 8 KHz). In the present invention, the speech feature vector extraction module is not provided with a separate noise reduction device. It is replaced by a noise attenuator used as a speech encoding preprocessor.

For example, in the present invention, the speech coding / recognition preprocessor 11 performs a noise attenuation function so that the speech feature vector extraction module can extract MFCCs robust to noise. (11) is implemented as a specification capable of performing both signal preprocessing for voice calls and signal preprocessing for voice recognition.

The speech coding / recognition preprocessor 11 proposed in the present invention as described above will be described in detail later with reference to FIG. 2, and reference numerals 12, 13, 14, 15, Detailed description of components corresponding to each of 16 and 17 "will be omitted.

FIG. 2 is a detailed block diagram of an embodiment of the speech coding / recognition preprocessor of FIG. 1.

As shown in FIG. 2, the speech coding / recognition preprocessor 11 according to the present invention includes a high pass filter 21, a frequency domain conversion 22, and a channel energy estimator. channel energy estimation (23), channel SNR estimation (24), voice metric calculation (25), spectral deviation estimation (26), noise update determiner ( noise update decision (also known as background noise estimate updater) 27, channel SNR modifier 28, channel gain computation 29, background noise estimator estimation (30), frequency domain filter (31) and time domain conversion (32).

In the present invention, speech coding / recognition is applied by applying a specification that can satisfy both a speech coding preprocessing for a voice call and a speech feature vector extraction preprocessing for speech recognition, for example, IS-127 (EVRC) used in the CDMA method. The preprocessor 11 is implemented.

Meanwhile, the user voice signal s _LFB (n) input to the voice coding / recognition preprocessor 11 is data in the form of a uniform PCM format (16 bits) having an 8 KHz sampling frequency.

In general, before the speech coding process and before the speech feature vector extraction process, the noise mixed in the speech signal must be attenuated to improve the signal quality. The speech coding / recognition preprocessor 11 proposed in the present invention also provides such a noise reduction function. As a main function, it can be seen that the user voice signal s _LFB (n) inputted as shown in the figure is output as s' (n) in which noise is attenuated.

Hereinafter, each component of the speech coding / recognition preprocessor 11 will be described.

The high pass filter 21 removes the low frequency band signal of the user voice signal s _LFB (n) received through a microphone and the like, and the high pass filter 21 has a cutoff frequency band of 120 Hz. do.

Hereinafter, a signal filtered by the high pass filter 21 is defined as s _hp (n), and this signal s _hp (n) is a noise reduction target signal, and the frame size of the noise reduction target signal is 10 ms, where Let's define the current frame as m.

The frequency domain converter 22 converts the signal s _hp (n) filtered by the high pass filter 21 onto the frequency domain, using s _hp (n) using smoothed trapezoid windows. Is transformed into the frequency domain (windowing). The frequency domain conversion process is as follows.

In the flattened trapezoidal window, the first D samples on the input frame buffer {d (m, n)} of the mth frame overlap with the last D samples of the previous frame. Equation 1].

Where m is the current frame, n is the sample index of the buffer {d (m)}, L with a value of 80 is the frame length, and D with a value of 24 is the overlap (or delay) within the samples. Indicates degree. Here, the remaining samples of the input buffer are pre-emphasis as shown in Equation 2 below.

는 프리-엠퍼시스 계수이다. 상기 [수학식 1]에서 입력 버퍼는 104(L+D=104) 개의 샘플을 갖으며, 이 샘플들 내에서 첫번째 D 샘플들은 이전 프레임으로부터 프리-엠퍼사이즈드 오버랩(pre-emphasized overlap)된 부분이고, 다음의 샘플들은 현재의 프레임으로부터 프리-엠퍼사이즈드 입력(pre-emphasized input)이 되는 부분이다. 이를 토대로, 평탄화된 사다리꼴 윈도우를 입력 버퍼에 사용하여 다음의 [수학식 3]과 같은 윈도윙된 신호를 획득한다.Where a value of "-0.8"

Is the pre-emphasis coefficient. In Equation 1, the input buffer has 104 (L + D = 104) samples, and the first D samples in these samples have a pre-emphasized overlap from the previous frame. The following samples are the parts that become pre-emphasized input from the current frame. Based on this, a flattened trapezoidal window is used as an input buffer to obtain a windowed signal as shown in Equation 3 below.

Here, M having a value of 128 is a discrete Fourier transform (DFT) length, and through this M, a spectral signal G (k) as shown in Equation 4 can be obtained.

As described above, the spectral signal G (k) transformed on the frequency domain by the frequency domain converter 22 is used as an input signal of the channel energy estimator 23.

The channel energy estimator 23 calculates the channel energy estimate of the current frame m of the spectral signal G (k) received from the frequency domain converter 22 as shown in Equation 5 below.

은 다음의 [수학식 6]에 정의된 채널 에너지 평탄 인수이고, 16의 값을 갖는 N_c는 결합된 채널 수를 나타낸다. 그리고, f_L(i) 및 f_H(i) 각각은 i번째 채널의 저주파수의 DFT bin 및 i번째 채널의 고주파수의 DFT bin을 나타낸다. 이러한 f_L(i) 및 f_H(i) 각각은 다음과 같다. Here, E _min with a value of 0.0625 represents the minimum allowable channel energy value,

Is a channel energy flatness factor defined in Equation 6 below, and N _c having a value of 16 represents the number of combined channels. And, f _L (i) and f _H (i) each represent a low frequency DFT bin of the i-th channel and a high frequency DFT bin of the i-th channel. Each of these f _L (i) and f _H (i) is as follows.

f _L = {2,4,6,8,10,12,14,17,20,23,27,31,36,42,49,56},

f _H = {3,5,7,9,11,13,16,19,22,26,30,35,41,48,55,63}.

가 첫번째 프레임(m=1)에 대한 값이 영(0)인 경우에는 그 채널 에너지 추정값이 첫번째 프레임의 필터링되지 않은 채널 에너지값으로 초기화되는 것을 허용하는 의미한다.In the channel energy estimate calculated by Equation 5, the channel energy flatness factor

If the value for the first frame (m = 1) is zero, it means that the channel energy estimate is allowed to be initialized to the unfiltered channel energy value of the first frame.

The channel signal-to-noise ratio estimator 24 estimates the signal-to-noise ratio (SNR) present on the channel, which is estimated by the channel energy estimator 23 and the background noise estimator 30. Based on the background noise energy estimate, the quantized channel SNR indices are calculated as shown in Equation 7 below.

의 값은 0에서 89 사이의 값을 갖게 된다.Meanwhile, E _n (m) calculated by the background noise estimator 30 represents a current channel noise energy estimate, which is calculated through Equation 7 based on this.

Has a value between 0 and 89.

를 토대로 현재 채널 상의 음성 메트릭들의 합을 다음의 [수학식 8]과 같이 계산한다.The voice metric calculator 25 uses the signal to noise ratio estimated by the channel signal to noise ratio estimator 24, e.g., the quantized channel SNR index.

The sum of speech metrics on the current channel is calculated as shown in Equation 8 below.

Here, V (k) is a negative metric composed of 90 elements as follows.

V (k) = {2,2,2,2,2,2,2,2,2,2,2,3,3,3,3,3,4,4,4,5,5,5, 6,6,7,7,7,8,8,9,9,10,10,11,12,12,13,13,14,15,15,16,17,17,18,19,20, 20,21,22,23,24,24,25,26,27,28,28,29,30,31,32,33,34,35,36,37,37,38,39,40,41, 42,43,44,45,46,47,48,49,50,50,50,50,50,50,50,50,50,50}.

The spectral deviation estimator 26 estimates the spectral deviation of the signal on the current channel based on the channel energy estimate E _ch (m) estimated by the channel energy estimator 23. This spectral deviation estimation process is as follows.

First, the log power spectrum on the current channel is calculated by the following Equation 9 using the channel energy estimate E _ch (m) as a parameter.

Then, the difference value between the estimated value for the current power spectrum and the estimated value for the average long-term power spectrum calculated in Equation 9 is calculated through Equation 10 below.

은 이전 프레임 동안에 계산된 평균 장구간 전력 스펙트럼에 대한 추정값이다. 다만, 이러한 평균 장구간 전력 스펙트럼

은 스펙트럼 편차 추정 과정을 통해 갱신되기 전, 예컨대

초기값은 다음의 [수학식 11]과 같이 첫번째 프레임의 로그 전력 스펙트럼에 대한 추정값으로 설정된다.here,

Is an estimate of the average long-term power spectrum calculated during the previous frame. However, these average long-term power spectrum

Before is updated through the spectral deviation estimation process,

The initial value is set as an estimate of the log power spectrum of the first frame as shown in Equation 11 below.

은 하기의 잡음 업데이트 결정기(27)에서 배경 잡음 추정값을 갱신하는데 사용되는 파라미터로서 입력된다.The channel energy estimate E _ch (m) is used as a parameter to calculate the total channel energy estimate E _tot (m) for the current frame m using Equation 12 below. The total channel energy estimate E _tot (m) calculated as above and the difference between the estimate for the current power spectrum and the estimate for the average long-term power spectrum

Is input as a parameter used to update the background noise estimate in the following noise update determiner 27.

은 총 채널 에너지 추정값 E_tot(m)에 관한 함수로서 다음의 [수학식 13]으로 계산된다. Meanwhile, exponential window function argument

Is a function of the total channel energy estimate E _tot (m) and is calculated by Equation 13 below.

은 다음의 [수학식 14]에 의해

사이의 값으로 제한된다.Here, the exponential window function factor calculated through Equation 13

Is given by Equation 14 below.

Limited to values in between.

로 제한된

으로 변환된 E_tot(m)의 선형 내삽치에 대한 dB로 표현되는 에너지 끝점들이며, 이러한 세 상수들의 값은

로 정의된다. 예를 들어, 40 dB의 상대적인 에너지를 갖는 어떤 신호는 다음에 이어지는 계산을 위해서

의 지수 창함수 인수를 사용하게 된다.In addition, E _H and E _L

Limited to

Are the energy endpoints expressed in dB for the linear interpolation of E _tot (m) converted into.

Is defined as For example, some signals with a relative energy of 40 dB may be used for subsequent calculations.

We will use the exponential window function argument of.

,

초기값 등을 파라미터로 하여 다음의 [수학식 15]를 통해 다음 프레임에 대한 평균 장구간 전력 스펙트럼 추정값을 갱신한다.Finally, the exponential window function argument calculated as above

,

Using the initial value as a parameter, the average long-term power spectrum estimation value for the next frame is updated through Equation 15 below.

을 토대로, 다음의 의사코드로 표현된 로직을 통해 배경 잡음 추정기(30)에서 추정한 잡음 추정값에 대해 어떠한 추정 값으로 갱신하라는 명령, 예컨대 update_flag를 내린다.The noise update determiner 27 determines the total channel energy estimate E _tot (m) estimated by the spectrum deviation estimator 26, the difference between the estimate for the current power spectrum and the estimate for the average long-term power spectrum.

On the basis of the following, the logic expressed by the following pseudo code gives an instruction, for example, update_flag, to update any estimated value with respect to the noise estimate estimated by the background noise estimator 30.

, DEV_THLD=28, UPDATE_CNT_THLD=50, HYSTER_CNT_THLD=6이다.On the logic represented by the pseudo code, the values of the usage constants are UPDATE_THLE = 35 and NOISE_FLOOR_DB =, respectively.

, DEV_THLD = 28, UPDATE_CNT_THLD = 50, and HYSTER_CNT_THLD = 6.

의 값을 수정한다. 이러한 수정된 채널 SNR 인덱스

은 채널 이득 산출기(29)의 입력 파라미터로서 사용된다. 다음의 의사코드로 표현된 로직은 채널 신호대잡음비 추정값을 어떻게 수정하는지에 대해 보여주고 있다.The channel signal to noise ratio estimation modifier 28 estimates the quantized channel SNR index estimated by the channel signal to noise ratio estimator 24 based on the sum V (m) of the voice metrics on the current channel estimated by the voice metric calculator 25.

Modify the value of. These modified channel SNR indexes

Is used as an input parameter of the channel gain calculator 29. The logic represented by the following pseudo code shows how to modify the channel signal to noise ratio estimate.

On the logic represented by the pseudo code, the values of the constants and thresholds are as follows.

과 배경 잡음 추정기(30)에서 추정한 배경 잡음 에너지 추정값 E_n(m)을 토대로 선형 채널 이득

을 계산한다. 이러한 선형 채널 이득 산출 과정에 대해 설명하면 다음과 같다.The channel gain calculator 29 adjusts the channel SNR index modified by the channel signal to noise ratio estimation modifier 28.

And a linear channel gain based on the background noise energy estimate E _n (m) estimated by the background noise estimator 30

Calculate The linear channel gain calculation process will be described below.

First, the overall gain factor for the current frame is calculated using Equation 16 below.

은 -13이고, 잡음 에너지 최저 한도(noise floor energy)

는 1이며, 배경 잡음 에너지 추정값 E_n(m)은 배경 잡음 추정기(30)에서 추정한 값을 따른다.Where minimum overall gain

Is -13, and noise floor energy

Is 1, and the background noise energy estimate E _n (m) follows the value estimated by the background noise estimator 30.

Next, the channel gains (in dB) are calculated through Equation 17 below.

는 0.39이며, 이 이득 기울기를 다음의 [수학식 18]과 같이 선형 채널 이득들로 바꾸는 것이 바람직하다.Where gain slope

Is 0.39, and it is preferable to change this gain slope to linear channel gains as shown in Equation 18 below.

을 다음의 [수학식 19]와 같이 적용한다. The frequency domain filter 31 calculates the linear channel gain calculated by the channel gain calculator 29 for the spectral signal G (k) converted by the frequency domain converter 22.

Is applied as in Equation 19 below.

In Equation 19, the spectral signal G (k) is changed to the signal H (k) by the channel gain, of which signals in the frequency band not changed by the channel gain, for example, H (k) = G ( k) is also present. The signal in the frequency band that is not changed by the channel gain is expressed by Equation 20 below. It can be seen that the magnitude of H (k) should be even and the phase should be odd. .

Here, a complex conjugate is needed to find an inverse DFT at H (k).

In particular, the background noise estimator 30 estimates its energy estimate E _n (m) for the noise signal present on the current channel, as described above, in accordance with the command issued by the noise update determiner 27, e.g. update_flag. Update the corresponding noise estimate.

That is, when the noise estimate update update_flag is true, the background noise estimator 30 updates the noise estimate on the channel for the next frame through Equation 21 below.

은 0.9이다. 한편, 잡음 추정값은 다음의 [수학식 22]와 같이 처음의 네개의 프레임들 각각에 대해 채널 에너지 추정값으로서 초기화되어 있다.Here, the minimum channel energy E _min is 0.0625 and the channel noise smoothing factor

Is 0.9. On the other hand, the noise estimate is initialized as a channel energy estimate for each of the first four frames as shown in Equation 22 below.

Here, the minimum channel noise initial energy E _init is 16.

The time domain converter 32 converts a user voice signal in which the noise received through the frequency domain filter 31 is attenuated, for example, a user voice signal expressed in the frequency domain into a user voice signal in the time domain. This time domain transformation process is as follows.

First, the signal filtered by the frequency domain filter 31 is converted into a time domain using an inverse DFT through Equation 23 below.

Then, overlap-and-add is applied to Equation 23 to perform a frequency domain filtering process as shown in Equation 24 below.

Finally, de-emphasis is applied to Equation 24 to convert the user voice signal in the time domain as shown in Equation 25 below.

는 0.8이며,

는 320개의 샘플들을 담을 출력 버퍼를 나타낸다.Where the de-emphasis factor

Is 0.8,

Denotes an output buffer that will contain 320 samples.

As described above, the user voice signal S '(n) whose noise is attenuated through the voice coding / recognition preprocessor 11 may be obtained, and the user voice signal S' (n) may be a voice call mode or a voice. According to the recognition mode, it is input to the voice feature vector extraction unit 12 of the voice feature vector extraction module 100 or to the voice coder 15 of the voice coding module 150.

이다. 물론, 잡음 감쇄 대상 신호의 프레임 사이즈의 그 설정 크기에 따라 사용자 음성신호 S'(n)를 다르게 출력할 수도 있음을 당업자라면 쉽게 이해할 수 있을 것이다.In addition, since the frame size of the noise reduction target signal is set to 10 ms as defined in the description of the speech coding / recognition preprocessor 11, since the noise reduction is performed every 10 ms frame, the above-described user voice signal S ' (n) [output signal of speech coding / recognition preprocessor]

to be. Of course, it will be readily understood by those skilled in the art that the user voice signal S '(n) may be output differently according to the set size of the frame size of the noise reduction target signal.

Meanwhile, a method corresponding to the speech coding / recognition preprocessor, which is a preprocessing device for the apparatus of the present invention, for example, the speech feature vector extraction module and the speech coding module described above with reference to FIG. It is a time series process corresponding to the apparatus, and the method flow of the present invention will not be described separately.

FIG. 3A is a diagram illustrating an embodiment of an extended speech coding / recognition preprocessor for 11 KHz speech signal processing according to the present invention, and FIG. 3B illustrates an extended speech coding / 16 KHz speech signal processing according to the present invention. One embodiment configuration diagram for the recognition preprocessor.

In the speech coding / recognition preprocessor 11 of the present invention described above with reference to FIG. 2, the user speech signal of 8 KHz is the noise reduction target signal. Thus, the present invention also provides a speech coding / recognition preprocessor for processing a 11 kHz user speech signal with reference to FIG. 3a, and a speech coding / recognition preprocessor for processing a 16 kHz user speech signal with reference to FIG. 3b. do.

3A is provided with a frequency down sampler 41 for converting a user voice signal of 11 KHz into a user voice signal of 8 KHz in front of the voice coding / recognition preprocessor shown in FIG. 2 for 11 KHz voice signal processing. . The user voice signal down-sampled by the frequency down sampler 41 is input to the present voice coding / recognition preprocessor.

FIG. 3B shows a low pass quadrature-mirror filter (QMF LP) in front of the speech coding / recognition preprocessor shown in FIG. 2 for processing 16 KHz speech signals [DEC by 2] (46). ) And a high pass quadrature-mirror filter (QMF HP) [DEC by 2 and SI] 47.

The low pass quadrature mirror filter 46 divides the received 16 KHz voice signal into a low frequency band of 0 KHz to 4 KHz, and the high pass quadrature mirror filter 47 divides the received 16 KHz voice signal into 4 Divide into high frequency band of KHz ~ 8KHz.

In particular, the low frequency signal processed by the low pass quadrature mirror filter 46 is input to the speech coding / recognition preprocessor, while the high frequency signal processed by the high pass quadrature mirror filter 47 extracts the speech feature vector. The voice feature vector extracting unit (MFCC front-end) 12 of the module 100 is input. This allows the voice feature vector extractor 12 to use 26 Mel-filter banks to extract the voice feature vector MFCCs.

As described above, the low frequency signal processed by the low pass quadrature mirror filter 46 is input to the voice feature vector extractor 12 through the voice coding / recognition preprocessor and processed by the high pass quadrature mirror filter 47. The received high frequency signals are combined with each other as one signal on the voice feature vector extractor 12 while being input to the voice feature vector extractor 12. For example, this high frequency signal is combined with the low frequency signal before the log filter bank energy is converted to the cepstrum coefficient. In addition, log-energy of all frequency bands is also calculated using both high and low frequency signals.

In addition, the extended speech coding / recognition preprocessor shown in FIGS. 3A and 3B is subject to the frequency extension scheme specification of the ETSI DSR standard (ETSI ES 202 050 v1.1.3) so that a sampling frequency signal of 11 KHz or 16 KHz can be used. Can be implemented accordingly.

4 is a diagram illustrating an embodiment of a speech recognition process used in the present invention, FIG. 5 is a flowchart illustrating an embodiment of a learning process for generating an acoustic model in FIG. 4, and FIG. 6 is applied to the present invention. It is a graph showing speech recognition performance using extracted speech feature vectors.

The present invention described with reference to Figs. 1, 2, 3A and 3B can be applied to a distributed voice recognition terminal such as a mobile phone, and it is necessary to verify how the present invention affects the voice recognition performance.

Thus, the speech recognition performance to which the present invention is applied will be described through a speech recognition process and a learning process for generating an acoustic model.

FIG. 4 illustrates a speech recognition process using a Hidden Markov Model (HMM). The speech feature is extracted from the user's speech (301), the acoustic model (303), and the language model (304) for the extracted speech feature. In addition, the phonetic dictionary 305 is searched and pattern matching 302 is performed to recognize words and sentences corresponding to the voices.

In the speech feature extraction 301, the method proposed by the European Telecommunication Standards Institute (ETSI) standard proposal "ETSI ES 201 108" is used. That is, the speech feature is extracted from the speech data through the MFCC to form a speech feature vector as a higher order coefficient. The speech feature vector uses an acoustic model 303, a language model 304, and a pronunciation dictionary 305. Pattern matching is used to find a word string with the highest probability.

In particular, to verify the speech recognition performance of the conventional method, the signal or the IS of which the noise is attenuated through the preprocessing process defined in the ETSI DSR standard (ETSI ES 202 050 v1.1.3) is used for the speech signal used for speech feature extraction as described above. The signal is attenuated using the preprocessing process defined in -127.

On the other hand, to verify the speech recognition performance of the present invention, a signal whose noise is attenuated through the speech coding / recognition preprocessor 11 that uses the speech signal used for speech feature extraction as described above is used.

In the speech feature extraction process, the 13th MFCCs and log-energy are extracted using the MFCC front-end as the input signal. Among them, 12th order MFCCs (c0, ..., c12), log-energy, and their delta and delta-delta are used as parameters for learning and recognizing acoustic models.

On the other hand, the HMM is used in the acoustic model 303, and the phoneme model according to the language is used as the acoustic model in the present invention. The learning process for generating the phoneme model is described below with reference to FIG. 5.

First, a monophone based model, which is a phoneme independent model, is generated using a speech feature vector extracted from training data (401).

Then, a new phoneme label file is generated by performing forced alignment based on the generated monophone-based model (402).

Meanwhile, the generated monophone based model is extended to generate a triphone based model, which is a phoneme dependent model (403).

Then, state sharing is performed in consideration of the small amount of training data for the generated triphone-based model (404).

Then, the model obtained as a result of the state sharing is increased by mixing density to generate a final acoustic model (405).

Meanwhile, the language model 304 illustrated in FIG. 4 uses a statistic-based method. Here, the statistics-based method refers to a method of statistically estimating the probability value of possible word strings from a DB of speech spoken in a given situation. A representative of such statistical-based language models is n-gram. This en-gram method approximates and uses the word string probability as the product of the previous n conditional probabilities, and uses bigram in FIG.

In the pronunciation dictionary 305, in Korean, the pronunciation dictionary provided by SiTEC's "CleanSent01" is used, and in English, "CMU dictionary V.0.6" provided by Carneige Mellon University is used. In addition, for the pronunciation of a word not provided in "CleanSent01", a pronunciation converter produced in-house using the "pronounciation of standard language definition" was used. The total number of words in the pronunciation dictionary provided by "CleanSent01" is 36,104, and the total number of words in the pronunciation dictionary used in the speech recognizer is 223,857.

On the other hand, in the speech DB used in FIG. 4, in the case of Korean, the "reading sentence speech DB (CleanSent01)" provided by the Voice Information Industry Technology Center (SiTEC) is used, and in the case of English, "AURORA" constructed by ETSI. 4 DB (Wall Street Journal).

Meanwhile, in order to generate a language model, 5,000 sentences of text data used for training, and 3,000 sentences of selected text data among “text DB for speech recognition language model” provided by ETRI may be used. HTK (Hidden Markov Model Toolkit) v3.1 was used to generate these language models. Finally, the generated model consists of a network of 31,582 words.

Figure 6 shows the result of using the noise attenuated signal processed according to the conventional method in the speech recognition process (68, 51%), and the result of using the noise attenuated signal processed according to the present invention in the speech recognition process The word recognition rate (69.31%) is shown, and as can be seen from FIG. 6, it can be understood that the speech recognition performance to which the present invention is applied is improved compared to the conventional method.

For example, in the terminal of the present invention, by extracting a speech feature vector robust to noise due to a communication environment by sharing a speech encoding preprocessing process and a speech feature vector extraction preprocessing process, the memory of the low specification terminal and the amount of calculations are reduced. It will be readily apparent to those skilled in the art that improving speech recognition performance.

As described above, the method of the present invention may be implemented as a program and stored in a recording medium (CD-ROM, RAM, ROM, floppy disk, hard disk, magneto-optical disk, etc.) in a computer-readable form. Since this process can be easily implemented by those skilled in the art will not be described in more detail.

The present invention described above is capable of various substitutions, modifications, and changes without departing from the technical spirit of the present invention for those skilled in the art to which the present invention pertains. It is not limited by the drawings.

The present invention as described above, by sharing the pre-processing for voice calls and speech recognition, there is an effect that can improve the speech recognition performance while consuming less memory and calculation amount of the low-end terminal.

In addition, the present invention has an effect of preventing the switching delay between the voice call process and the voice recognition process due to the speech encoding preprocessing process and the speech feature vector extraction preprocessing process in the terminal.

In addition, the present invention has the effect of properly attenuating noise mixed in the user's voice signal in speech coding and speech feature vector extraction.

Claims (25)

An apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, A high pass filter for removing the low frequency band signal of the voice signal received from the outside; A frequency domain converter for converting a signal from which the low frequency band is removed in the high pass filter into a spectral signal in a frequency domain; A channel energy estimator for calculating a channel energy estimate for the current frame of the spectral signal converted by the frequency domain converter; A channel signal to noise ratio estimator for estimating a channel signal to noise ratio for the speech signal based on the channel energy estimate calculated by the channel energy estimator and the background noise energy estimated by the background noise estimator; The background noise estimator for calculating and updating a background noise energy estimate for the speech signal according to a command of a noise update determiner; A voice metric calculator for calculating a sum of voice metrics on a channel with respect to the voice signal based on the channel signal to noise ratio estimated by the channel signal to noise ratio estimator; A spectral deviation estimator for estimating a spectral deviation of the speech signal based on the channel energy estimate calculated by the channel energy estimator; The noise update determiner that issues a noise estimate update command based on a total channel energy estimate estimated by the spectrum deviation estimator and a difference value between an estimated value for the current power spectrum and an estimated long term power spectrum; A channel signal to noise ratio estimation value corrector for correcting the channel signal to noise ratio estimated by the channel signal to noise ratio estimator based on the sum of the voice metrics calculated by the voice metric calculator; A channel gain calculator for calculating a linear channel gain based on the channel signal-to-noise ratio modified by the channel signal-to-noise ratio estimation modifier and the background noise energy estimate calculated by the background noise estimator; A frequency domain filter applying the linear channel gain calculated by the channel gain calculator to the spectrum signal converted by the frequency domain converter; And A time domain converter for converting a spectral signal to which a linear channel gain is applied in the frequency domain filter into a speech signal in the time domain. Apparatus for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal comprising a. The method of claim 1, The speech signal output from the time domain converter is noise attenuated, and is distributed to a speech feature vector extracting unit of the speech feature vector extracting module mounted in the terminal or to a speech coder of the speech coding module. A device for extracting speech feature vectors robust to noise by sharing the preprocessing of speech coders in. The method of claim 1, And a frame size of the signal filtered by the high pass filter is 10 ms. The apparatus for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal. The method of claim 1, The frequency domain transformer uses a smoothed trapezoid window to convert the input signal into a spectral signal in the frequency domain, thereby sharing the preprocessing of the speech coder in a distributed speech recognition terminal. Device to extract it. The method of claim 1, In calculating the channel energy estimate, if the channel energy flatness factor is zero for the first frame, the channel energy estimate is allowed to be initialized to the unfiltered channel energy value of the first frame. A device for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal. The method of claim 1, And a signal-to-noise ratio for the speech signal includes a quantized channel SNR index as in the following Equation. [Equation]

A method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, Removing the low frequency band signal of the voice signal received from the outside; Converting the signal from which the low frequency band has been removed into a spectral signal in a frequency domain; Calculating a channel energy estimate for the current frame of the transformed spectral signal; Estimating a spectral deviation of the speech signal based on the calculated channel energy estimate; Issuing a noise estimate update command based on the estimated total channel energy estimate and the difference between the estimated value for the current power spectrum and the estimated value for the average long-term power spectrum; Calculating and updating a background noise energy estimate for the speech signal when the noise estimate update command is received; Estimating a channel signal to noise ratio for the speech signal based on the calculated channel energy estimate and background noise energy estimate; Calculating a sum of speech metrics on a channel with respect to the speech signal based on the estimated channel signal to noise ratio; Modifying the estimated channel signal to noise ratio based on the sum of the calculated speech metrics; Calculating a linear channel gain based on the modified channel signal to noise ratio and the calculated background noise energy estimate; Applying the calculated linear channel gain to the transformed spectral signal; And Converting the spectral signal to which the linear channel gain is applied to an audio signal in a time domain A method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal including a. The method of claim 7, wherein The spectral deviation estimation step, Calculating a log power spectrum on a current channel for the speech signal using the channel energy estimate as a parameter; Calculating a difference value between the estimated value for the calculated current power spectrum and the estimated value for the average long-term power spectrum; Calculating a total channel energy estimate for a current frame on a channel related to a speech signal using the channel energy estimate as a parameter; Calculating an exponential window function factor using the calculated total channel energy estimate as a parameter; Updating the average long-term power spectrum estimation value for the next frame on the channel with respect to the speech signal using the calculated exponential window function factor and the initial power spectrum value as parameters; A method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal comprising a. The method of claim 8, The average long-term power spectrum is initialized as an estimate of the log power spectrum of the first frame on the channel for the speech signal. The method of extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal. . The method of claim 7, wherein In the noise update determining step, the background noise estimate updating parameter includes a total channel energy estimate calculated in the spectrum deviation estimating step and a difference between an estimate for the current power spectrum and an estimate for the average long-term power spectrum. A method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal. The method of claim 7, wherein The linear channel gain calculating step, Calculating an overall gain factor for the current frame on the current channel for the speech signal; And The process of calculating the channel gain on the current channel for the voice signal A method for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal comprising a. An apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, Voice coding means for transmitting the coded voice signal to the outside through a voice traffic channel in a voice call mode; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; And It includes a speech coding / recognition pre-processing unit for attenuating noise mixed in the speech signal received from the outside, The speech signal inputted to the speech coding means and the speech signal inputted to the speech feature vector extracting means are preprocessed by the speech coding / recognition preprocessor. Device for extracting speech feature vectors. The method of claim 12, And a speech signal input to the speech coding / recognition preprocessor, wherein the speech signal has a frequency band of 8 KHz. The method of claim 12, The speech feature vector extracting means, A speech feature vector extractor for extracting a speech feature vector from the speech signal preprocessed by the speech coding / recognition preprocessor; A speech compressor for compressing the speech feature vector extracted by the speech feature vector extractor; And Bit stream transmission unit for transmitting the speech feature vector compressed by the speech compression unit to the bit stream data to be transmitted to the outside Apparatus for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal comprising a. The method of claim 12, The speech coding means, A speech coder for coding a speech signal preprocessed by the speech coding / recognition preprocessor; A voice compressor for compressing a voice signal coded by the voice coder; And Bit stream transmission unit for transmitting the voice signal compressed by the voice compression unit to the bit stream data for transmission to the outside Apparatus for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal comprising a. The method of claim 12, An apparatus for extracting a robust speech feature vector to noise by sharing the preprocessing of a speech encoder in a distributed speech recognition terminal further comprising a switch for transitioning between a voice call mode and a speech recognition mode. An apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, Voice coding means for transmitting the coded voice signal to the outside through a voice traffic channel in a voice call mode; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; A frequency down sampler for down sampling the frequency band of the audio signal received from the outside; And A speech coding / recognition preprocessor for attenuating noise mixed in the down-sampled speech signal by the frequency down sampler; The speech signal inputted to the speech coding means and the speech signal inputted to the speech feature vector extracting means are preprocessed by the speech coding / recognition preprocessor. Device for extracting speech feature vectors. The method of claim 17, The speech signal input to the frequency down sampler has a frequency band of 11 KHz, the apparatus for extracting a speech feature vector robust to noise by sharing the preprocessing of the speech coder in a distributed speech recognition terminal. An apparatus for extracting a speech feature vector robust to noise by sharing preprocessing of a speech coder in a distributed speech recognition terminal, Voice coding means for transmitting the coded voice signal to the outside through a voice traffic channel in a voice call mode; Speech feature vector extraction means for causing the extracted user speech feature vector to be transmitted to the outside in the speech recognition mode; A low pass quadrature mirror filter for passing a low frequency band signal to a voice signal received from an outside; A high pass quadrature mirror filter for passing a high frequency band signal to the voice signal received from the outside; And And a speech coding / recognition preprocessor for attenuating noise mixed in the speech signal passed by the low pass quadrature mirror filter. The speech signal inputted to the speech coding means and the speech signal inputted to the speech feature vector extracting means are preprocessed by the speech coding / recognition preprocessor. Device for extracting speech feature vectors. The method of claim 19, The voice signal input to each of the low pass quadrature mirror filter and the high pass quadrature mirror filter has a frequency band of 16 KHz. Device for extracting feature vectors. The method of claim 20, The low pass quadrature mirror filter divides a 16 KHz voice signal input from the outside into a low frequency signal of 0 KHz to 4 KHz, thereby sharing the preprocessing of the voice coder in a distributed speech recognition terminal. Device for extracting vectors. The method of claim 20, The high pass quadrature mirror filter divides a 16 KHz voice signal received from the outside into a low frequency signal of 4 KHz to 8 KHz, thereby sharing the preprocessing of the voice coder in a distributed speech recognition terminal. Device to extract it. The method according to any one of claims 19 to 22, The low frequency signal processed by the low pass quadrature mirror filter is input to the speech coding / recognition preprocessor, and the high frequency signal processed by the high pass quadrature mirror filter is input to the speech feature vector extracting means. A device for extracting a speech feature vector robust to noise by sharing preprocessing of a speech encoder in a distributed speech recognition terminal. The method of claim 23, The low frequency signal and the high frequency signal are combined into a single signal on a speech feature vector extracting means, and the high frequency signal is combined with the low frequency signal before the log filter bank energy is converted into a cepstrum coefficient. A device for extracting speech feature vectors robust to noise by sharing preprocessing of speech coders in a recognition terminal. In a terminal equipped with a processor, A function of removing a low frequency band signal of an audio signal input from the outside; Converting the signal from which the low frequency band is removed into a spectral signal in a frequency domain; Calculating a channel energy estimate for the current frame of the transformed spectral signal; Estimating a spectral deviation with respect to the speech signal based on the calculated channel energy estimate; Issuing a noise estimate update command based on the estimated total channel energy estimate and the difference between the estimated value for the current power spectrum and the estimated value for the average long-term power spectrum; Calculating and updating a background noise energy estimate for the speech signal when the noise estimate update command is received; Estimating a channel signal-to-noise ratio for the speech signal based on the calculated channel energy estimate and background noise energy estimate; Calculating a sum of speech metrics on a channel with respect to the speech signal based on the estimated channel signal to noise ratio; Modifying the estimated channel signal to noise ratio based on the sum of the calculated speech metrics; Calculating a linear channel gain based on the modified channel signal to noise ratio and the calculated background noise energy estimate; Applying the calculated linear channel gain to the transformed spectral signal; And A function of converting the spectral signal to which the linear channel gain is applied to an audio signal in the time domain A computer-readable recording medium having recorded thereon a program for realizing this.

Images