KR950009331B1 - Voice recognizing method for using vector quantizer and two step regonition - Google Patents

Abstract

The sound recognizing method using vector quantization and two stage recognition of a sound recognition system comprises the steps of storing reference sound data via a vector quantizer by an LBG algorithm after detecting sound characteristic; obtaining the distortion by vector-quantizing the input sound after designating the range of the code book; performing a pre-recognition function; determining the entry having the minimum distance as recognition data.

Description

Speech Recognition Method Using Vector Quantization and Two-stage Recognition of Speech Recognition System

1 is an overall configuration diagram of a voice recognition system of the present invention.

2 is a signal flow diagram for the voice storage process of the present invention.

Figure 3 (a), (b) is a signal flow diagram for the speech recognition process of the speech recognition system of the present invention.

* Explanation of symbols for main parts of the drawings

1: Microphone 2: Microphone / Speaker Interface

3: low pass sampling unit 4: A / D converter

5 digital signal processing element 6 address decoder

7: Buffer 8: Data ROM

9: Data RAM 10: Program ROM

11: input / output decoding logic 12: voice recognition device

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition system, and more particularly, to reduce the amount of reference data by performing vector quantization during speech storage, and to pre-recognition by distortion according to vector quantization during speech recognition. The present invention relates to a speech recognition method using vector quantization and two-stage recognition of a speech recognition system, which performs dynamic program matching as needed and reduces computation time by dynamic program matching.

The speech recognition methods of the conventional speech recognition system include linear matching and dynamic program matching. Linear matching is a method of linearly correcting a change in the time axis of an input speech signal and a stored speech signal. Dynamic program matching is a method of matching by optimizing nonlinearly the change on the time axis.

However, in the linear matching method of the conventional speech recognition system as described above, there is a defect that the recognition performance is deteriorated. In the dynamic program method, the performance is good but the computational amount increases, so that the reaction time is long.

The present invention reduces the reference data amount through vector quantization during voice storage in consideration of such a conventional defect, and performs dynamic program matching as necessary after performing preliminary recognition by distortion according to vector quantization during voice recognition. The present invention has been devised a speech recognition method using vector quantization and two-stage recognition of speech recognition system that not only improves recognition performance but also shortens the calculation time.

는 인터럽트단자,

는 인터럽트 인지단자,

는 데이타 선택단자,

는 입력, 출력 선택단자,

는 프로그램 선택단계이다.1 is an overall configuration diagram of the speech recognition system of the present invention, as shown in FIG. 1, through which the audio output passing through the microphone 1 and the microphone / speaker interface 2 is passed through the low-pass filter sampling unit 3 to the analog A. / Digital (D) converter 4 is converted into a digital signal, and applied to the digital signal processing element 5, and then through the address decoder 6, the buffer 7 through the data ROM (8), data RAM ( 9), the microphone 10 is stored in the program ROM 10 and the microphone / speaker interface 2 is controlled through the input / output decoding logic 11, and the voice recognition device 12 is driven. 5) XF is the clock signal generation terminal,

Is the interrupt terminal,

Is the interrupt acknowledge terminal,

Is the data selection terminal,

Is the input and output selection terminals,

Is the program selection phase.

Referring to the effects of the present invention configured as described above are as follows.

First, the voice recording process is performed through the microphone 1, the microphone / speaker interface 2 is divided into a frame through the low pass filter sampling unit 3, and then converted into a digital signal through the A / D converter 4. Is converted.

The digital signal obtained in this way is applied to the digital signal processing element 6 and then extracted into a frequency spectrum through a fast fourier transform (FFT), which is made into a constant group, and the pattern thereof is a data ROM 8 and a data RAM 9. )

In addition, the voice recognition process is the voice signal applied to the microphone / speaker interface (2) through the microphone (1) is divided into a frame through the low-pass filter sampling unit (3) and then converted into a digital signal through the A / D converter (4) After the signal thereof is applied to the digital signal processing element 5, the input pattern is divided into a predetermined group through FF7.

The pattern thus obtained is compared with the reference patterns stored in the data ROM 8 and the data RAM 9 one by one to extract the most similar ones, and the recognized information thereof is inputted through the input / output decoding logic 11 to the speech recognition apparatus 12. To drive.

Here, the speech recognition data storage process using vectorization will be described with reference to FIG. 2, which is a signal flow diagram.

First, the voice section is detected using the energy and zero crossing ratio by receiving the sampled digital voice signal. In this case, the voice interval has a larger energy than the surrounding background, and the zero crossing ratio is large when the voice section starts with an unvoiced sound.

In this case, since the power of the high frequency component is generally small, the pre-emphasis subtracting the previous value multiplied by the constant (a) from the current value, such as H (Z) = 1-az ^-1. ) To increase the high frequency components and cover the Hamming Window to prevent distortion in the FFT.In this case, the Hamming window formula that multiplies each component

W (n) = 0.56-0.46 COS (2n / N), 0 <n <N-1

W (n) = 0, area of remaining share

Where N is the number of samplings in one frame.

Then, the FFT is performed to obtain each power spectrum, and the characteristics of the voice are detected by dividing this into a predetermined group and adding the line spectrum belonging to the group.

The characteristics of the speech obtained in this way are divided into groups of the desired number of representative vectors, the center of gravity is obtained, the initial codebook is set, and the optimal codebook is designed by the LBG algorithm. In this case, only the index of the codebook is stored for each frame of the voice reference feature designed by the codebook, and the voice reference pattern is stored as the above operation is repeated until the voice storage process is completed.

In addition, the speech recognition process using vector quantization and two-stage recognition will be described with reference to FIG. 3 (a) and (b).

The voice section is detected by receiving the sampled digital voice signal as in the voice storage process, the pre-emphasis and hamming window are performed, and the FFT is performed to detect the feature of the voice.

, 구간=스케일×상수(const)보다 작은 경우, 구간=상수(const)로 설정하여 코드북을 저장하고, 현재입력프레임수를 n=1, 기준인덱스 프레임수를 m, 기준코드북의 수를 l로 설정한다. 이후{cbno}={인덱스(m) : n×스캐일-구간＜m＜n×스케일+구간)으로 정해지는 코드북의 수(cbno)를 선택하고, 선택된 코드북으로 벡터양자화 디스토션을 계산한 후 기준코드북(l), 입력프레임수(n)로 최소로되는 유클리디언(Euclidian)거리 (M-indist)를 산출하여 디스토션을 결정한다.At this time, scale == with the current number of reference entries set to i = 1.

If the interval is smaller than the interval = scale × const, then set the interval = const to save the codebook, n = 1 for the current input frame, m for the reference index frame, and l for the reference codebook. Set it. {Cbno} = {index (m): select the number of codebooks (cbno) determined by n × scale-division <m <n × scale + division), calculate vector quantization distortion with the selected codebook, and then reference codebook (1) The distortion is determined by calculating the Euclidian distance (M-indist), which is minimized to the number of input frames (n).

Therefore, as the interval is set as described above, mismatched codebooks that do not match the n-th input frame are prevented, and distortions that are incorrectly obtained when vectorizing the vector are prevented. .

In this way, the vector quantization is performed for each reference entry in the same manner as above to obtain the sum sum (1) of the distortions at that time, and the distortion sum is obtained for all the number of input frames (n) and the reference entries (i). Dist (1), Dist (2),... , N like Dist (N).

If the difference with the smallest distortion Dist (1) is more than a certain threshold, the distortion is excluded and dynamic program matching is performed only for the corresponding entry up to that point.

Therefore, when the extracted speech feature performs vector quantization through the codebook of each reference entry, preliminary speech recognition is performed to select an entry having a small distortion.

If the selected reference entry index i is equal to sum (i) = Dist (1), then the voice recognition process is stopped, and the recognized result is distortion Dist (1). In addition, when several reference entries 1 are selected, dynamic program matching is performed, and the entry having the smallest distance is a recognition result.

As described in detail above, the present invention reduces the reference data by vector quantization, thereby saving memory, and having a two-stage matching of preliminary speech recognition and dynamic program matching.

In addition, since the range of the codebook is limited in the preliminary speech recognition process, the distortion is prevented from being incorrectly calculated, so that the computational amount is further reduced, thereby increasing the response speed.

Claims (1)

After detecting the speech feature, storing reference speech data through vector quantization by the algorithm, calculating the distortion by setting the interval of speech recognition, specifying the codebook range, and quantizing the input speech, Performing the preliminary speech recognition function of selecting only a predetermined entry by comparing the distortion with respect to the number of input frames and the reference entries, and performing dynamic program matching on the selected fixed entry to recognize an entry having a minimum distance. Speech recognition method using vector recognition and two-stage recognition of the speech recognition system, characterized in that the step consisting of determining.

Images