JPS6295598A - Voice recognition apparatus - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device that is applied to, for example, recognizing a speaker's speech word by word.

[Summary of the invention]

This invention provides a speech recognition device that converts a speech signal into a frequency spectrum to produce time-series frame data, performs binarization processing for each frame, obtains direct data regarding the spectrum data, and also Binary data indicating the characteristics of one frame data as a whole, such as the distance between adjacent frames and phoneme properties (vowel, consonantity, etc.), is added to the binary data of the spectral data to obtain direct feature data. By performing pattern matching based on this binary feature data, the recognition rate is improved, memory capacity is reduced, and calculation processing time is shortened.

[Conventional technology]

Tree + t, I! The voice recognition device previously proposed by the applicant (Special Patent No. 1988-106177) includes a microphone as a voice input section, a preprocessing circuit, an acoustic analyzer,
It is composed of a feature data extractor, a registered pattern memory, a pattern matching determiner, etc.

This audio section and conference device limits the audio signal manually input from the microphone to the band required for speech recognition in a preprocessing circuit, converts it into a digital audio signal using a /D converter, and converts this digital audio signal into a digital audio signal. Supplies the acoustic analyzer.

Then, in Otoya Σ JJr h, the audio signal is converted into a frequency spectrum, and the frequency spectrum is normalized using N frequencies as representative values at regular intervals on the logarithmic axis. Frame data consisting of N channels of spectral data is fed to the feature data extractor.

The feature data extractor calculates the distance between adjacent frame data, calculates the trajectory length of the N-dimensional vector from the start frame to the end frame of the audio signal by summing the distances between each frame, and extracts the longest audio with the largest number of words. In this case, the trajectory length is equally divided by the predetermined number of divisions necessary to extract the features, and only the frame data corresponding to the division points are extracted as feature data, thereby adjusting for fluctuations in the rate of speech generation of the speaker. Normalize and output the time axis so that it is not affected.

When registering this feature data, it is supplied to the registered pattern memory and used as a registered feature data block (standard pattern).
At the time of recognition, the input audio signal is processed as described above to become an input waiting data block (human cover turn), and is supplied to a pattern matching judger. And in the pattern matching judger, the input feature data block and ? Perform pattern matching between 21 Feature Data Pro and Ri.

The pattern matching determiner calculates the inter-frame distance between the frame data constituting the registered feature data block and the frame data constituting the input feature data block, and sets the sum of the inter-frame distances as the matching distance; The matching distance is calculated in the same way for Ike's registered feature data pro-tsuku, and the word corresponding to the data block whose matching distance is the minimum and is determined to be sufficiently close is selected as the recognition result. Output.

[Problem that the invention seeks to solve]

However, in the conventional audio part n (device), the frame data output from the acoustic analyzer is directly stored in the registered pattern memory as a recording data block via the feature data extractor. Therefore, there was a problem that the amount of memory of the registered pattern memory became enormous.

Along with this, there is also a problem that the calculation processing time becomes longer depending on the number of data in pattern matching.

Therefore, an object of the present invention is to reduce the capacity of registered pattern memory by binarizing frame data.
Another object of the present invention is to provide a speech recognition device that can reduce matching processing time.

Additionally, the applicant has developed a speech recognition device that accurately normalizes the tendency of the spectrum that varies due to various causes, converts each piece of spectral data constituting frame data into 24W, and performs pattern matching based on this binary data. Japanese Patent Application No. 166191/1983 has been proposed as such.

However, in this speech recognition device, due to the binarization process, the characteristics of one frame data as a whole are weakened.
There was a drawback that the difference with the pond frame data became smaller and the similarity became stronger. For example, if the frame data shown in Figure 3A is binarized using the reference level in the figure, rl
, 1,0.0,1,0,1゜1, .1,0,0,0.
0.1. '1.1 j] Although they are binary data and are clearly different frame data, there is almost no difference. When the pattern matching determiner calculates the distance between these binary frame data, the interframe distance is calculated as a small value, and there is a possibility that there will be no large difference in the matching distance and the recognition rate will decrease. .

Therefore, in particular, the present invention aims to improve the recognition rate without sacrificing the advantage of binarization, which is the reduction in memory capacity and the high speed of matching processing.

[Means for solving problems]

The present invention provides a speech recognition device in which an input audio signal is converted into N-channel frequency spectra, and time-series data of the N-channel frequency spectra is input, and a spectral cheater binarization process is performed for each frame of the time-series data. In addition to obtaining binary data of the spectral data, features related to one frame such as voicing/silence, distance between adjacent frames, and phoneme characteristics are extracted to obtain binary data. This is a speech recognition device characterized in that binary data regarding a frame is added to the input speech signal to generate binary feature data, and the binary feature data is used to recognize an input speech signal.

(Operation port: In the mixed binarization pattern extractor 10, the spectral data of the time series frame data compressed by the time axis normalization process is binarized, and the voice/silence extractor 7
．． Binary data indicating the characteristics of each frame supplied from the adjacent distance calculator 8 and the phonetic pattern extractor 9 is added to the binary data of the spectral data of the temporally corresponding frame to obtain binary feature data, Pattern matching is performed using this binary feature data.

[Example] An embodiment of the present invention will be described below with reference to the drawings.

FIG. 1 shows an embodiment of the present invention, and in FIG. 1, reference numeral 1 indicates a microphone as a voice input section.

An analog audio signal from microphone 1 is supplied to filter 2 . Filter 2 has a cutoff frequency of 7, for example.
．． It is a 5kHz low-pass filter, and the audio signal is passed through the filter 2 at a frequency of 7.5kll, which is required for speech recognition.
This audio signal is limited to a band below z and is supplied to an A/D converter 4 via an amplifier 3.

The A/D converter 4 has a sampling frequency of 12.5, for example.
The audio signal is analog-to-digital converted in the A/D converter 4 to an 8-bit digital signal, which is then supplied to the spectrum converter 5.

The spectrum converter 5 converts the audio signal into a frequency spectrum and generates, for example, an N-channel spectrum data string. In the spectrum converter 5, the audio signal is converted into a frequency spectrum by arithmetic processing,
For example, a spectral data string having representative values of N frequencies at constant intervals on the logarithmic axis is obtained. Therefore, the audio signal is expressed by the magnitude of the discrete frequency spectrum of N channels. Then, unit time (frame period 8
J]), a spectral data string of N channels is output as one frame data. That is, the audio signal is extracted every frame period as a parameter expressed by an N-dimensional vector, and the spectral pattern extractor 6.
The voiced/unvoiced extractor 7° is supplied to the adjacent distance calculator 8 and the phonemic pattern extractor 9, respectively.

The spectral pattern extractor 6 compresses time-series frame data by normalizing the time axis.

For example, in the spectral pattern extractor 6, the absolute value of the difference between the spectral data for each channel of adjacent frame data is determined, and the sum thereof is taken as the inter-frame distance between the adjacent frame data. Furthermore, the sum of the interframe distances is determined, and the trajectory length of the N-dimensional vector from the start frame to the end frame of the audio signal is determined. Then, in the case of the longest voice with the largest number of words, the trajectory length is equally divided by a predetermined number of divisions necessary to extract the features. Only frame data corresponding to each division point is extracted, the time axis is normalized so as not to be affected by variations in the speech production rate of the speaker, and this extracted frame data is mixed and converted into 1σ. A pattern extractor 10 is provided.

In the voiced/unvoiced extractor 7, voiced sections and unvoiced sections in the audio signal are detected based on the presence or absence of pitch waves.

In other words, voiced sounds include pinch waves because the exhaled air sent out from the lungs is interrupted by the vibration of the vocal cords, while unvoiced sounds are created by the articulatory organs, especially the tip of the tongue.
This is caused by the turbulent flow of air that occurs when exhaled air from the lungs passes through the narrow space formed by teeth, lips, etc. Does not include pitch waves. Therefore, for example, the presence or absence of a pitch wave is detected by determining the correlation regarding the low frequency channels of the time-series frame data sequentially supplied from the spectrum converter 5, and the section including the pinch wave is determined as the Yoimiya section. For example "1"
The interval that does not include a pitch wave is defined as a silent interval, and is expressed, for example, as "0", and voiced/unvoiced data is generated. This voiced/unvoiced data is supplied to a mixed binary pattern extractor 10.

In the adjacent distance calculator 8, the distance between adjacent frames of the time-series frame data supplied from the spectrum converter 5 is calculated by, for example, the sum of the absolute values of differences in spectrum data for each channel. This distance between adjacent frames is quantized with 2 bits and converted into adjacent distance data.

This adjacent distance data is supplied to the mixed binary pattern extractor 10.

In the phoneme pattern extractor 9, the phoneme 1 raw of each frame data, that is, the characteristic of the spectral shape is detected and expressed by, for example, 2-bit data. For example, the spectral shape of voiced vowels ra, ruj, and roj is characterized by a large level occurring in the lower channel, and in this case, data rl and Oj are generated. Also,
The shape of the spectrum of the voiced vowels riJ and reJ is characterized by large levels occurring in the low-frequency and high-frequency channels, and in this case, data of ro and Oj are generated. Also, the voiceless consonant rs j.

The shape of the spectrum such as itj has the characteristic that a large level is generated in the channel on the high frequency side, and in this case, data of rO and IJ are generated.

Furthermore, in the case of frame data that does not match the other three characteristic spectral shapes described above, data of "1°1" is generated. These phonemic data are supplied to the mixed binary pattern extractor 10.

In the mixed binarization pattern extractor 10, the frame data extracted by the spectral pattern extractor 6 is binarized. For example, the trend value for correcting trend fluctuations regarding the spectral data constituting each frame data is the average value of the spectral data from channel 0 to a predetermined channel n (0≦n≦N-1), and the predetermined channel n
It is obtained by multiplying the average value of the spectral data from 1 to the maximum channel N-1 by an appropriate coefficient. Subtraction is performed between the trend value determined for the spectral data of each channel and the corresponding spectral data to flatten the spectral trend,
The spectral tendency is normalized so as not to be influenced by individual differences among speakers, ambient noise, and the like. The spectral data whose spectral tendency has been normalized is compared with the standard 1 value set to an appropriate value, and the spectral data with the value larger than the reference value is set as "1", and the spectrum data with the value smaller than the base 4λ value is The data is set to "0" and binarized.

Also, voiced/unvoiced data (for example, 1 bit) is supplied from the voiced/unvoiced extractor 7.

The adjacent distance data (for example, 2 bits) supplied from the adjacent distance calculator 8 and the phonemic data (for example, 2 bits) supplied from the phonetic pattern extractor 9 are converted into binarized frame data (for example, 2 bits) that temporally correspond to each other. (for example, 7 bits) to form binary feature data, and this binary feature data is supplied to the mode switching circuit 11.

At the time of registration, this binary feature data is supplied to the registered pattern memory 12 via the mode switching circuit 11. For example, when M frames are extracted by the spectral pattern extractor 6, as shown in FIG. A data block is stored as a registered feature data block. At the time of recognition, the input audio signal is converted into binary feature data through the above-described processing, and this binary feature data is supplied to the pattern matching determiner 13 to be used as an input feature data block. Input feature data block and all entries 2. ★
Pattern matching is performed between the feature data blocks.

That is, the pattern matching determiner 13 calculates the interframe distance between the corresponding frames between the registered feature data block supplied from the registered pattern memory 12 and the input feature data block. For example, the inter-frame distance is determined by the sum of the absolute values of the data differences of the same frame, and the sum is taken as the matching distance. Then, the word corresponding to the registered feature data block that is determined to be the smallest and sufficiently close among the matching distances determined for all the registered feature data blocks is taken as the recognition result.

In one embodiment of the present invention, a configuration has been described in which voiced/unvoiced data, adjacent distance data, and phoneme data are added to binarized frame data. The configuration may be such that the information is added, or the configuration may be such that data such as audio intensity (power) is further added.

Moreover, this invention is not limited to a hard-wired configuration.
The processing may be performed by software using a microcomputer or a 71° microprogram method.

〔Effect of the invention〕

In this invention, in the mixed binarization pattern extractor, spectrum data of compressed time series frame data is binarized by regular processing on the time axis, and at the same time, the voiced/unvoiced extractor, the adjacent distance calculator, and the phoneme Binary data indicating the characteristics of each frame supplied from the pattern extractor is added to the binary data of the spectral data of the temporally corresponding frame to create binary feature data, and this binary feature data is used to create a pattern. Matching is performed.

Therefore, one advantage of this invention is that the binary data representing the characteristics of the entire frame data of 11 In addition, since binary feature data is used, the capacity of the registered pattern memory can be reduced, and the matching processing time can be shortened.

[Brief explanation of drawings]

FIG. 1 is a block diagram of the configuration of one embodiment of the present invention, and FIG.
FIG. 3 is a route map showing the data structure of a binary feature data block in one embodiment, and FIG. 3 is a route map used to explain a conventional speech recognition device. Microphone, 5-varnish vector converter in the explanation of the main symbols in the drawing. 6: Spectral pattern extractor, 7: Voiced/unvoiced extractor, 8: Adjacent distance calculator, 9: Phonemic pattern extractor, 10: ? M-combined binary pattern extractor,
11: Mode switching circuit, 12: Registered pattern memory, 13: Pattern matching determiner. Fig. 1 Fig. 3 A Fig. 3 B 2. Direct special episode reflow/7 Fig. 2

Claims (1)

[Scope of Claims] A speech recognition device in which an input audio signal is converted into N-channel frequency spectra, and time-series data of the N-channel frequency spectra is input, wherein two of the spectrum data are input for each frame of the time-series data. Value processing is performed to obtain binary data of the spectral data, and features related to one frame such as voicedness/silence, distance between adjacent frames, and phoneme characteristics are extracted to obtain binary data, and the above spectral data is Adding binary data regarding the frame to the binary data to obtain binary feature data,
A speech recognition device characterized in that the input speech signal is recognized using the binary feature data.