JPS5882300A - Voice recognition system - 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] The present invention relates to a voice recognition method.

The conventional voice weaving method has disadvantages such as a large memory capacity for voice feature extraction patterns, a long processing time for identification based on the memory contents, and an inability to increase processing speed. FIG. 1 shows a block diagram showing a conventional speech recognition system.

In FIG. 111, voice input is input to the preprocessing unit 1,
The preprocessing section 1 amplifies the input signal to a certain level, cuts out a speech section, and outputs it to the speech feature extraction section 2. The audio feature extraction unit 2 extracts the audio input signal from the audio input signal using (1) the $9 type prediction coefficient (LPC) method, (1) the frequency band power method, (
3) Partial autocorrelation coefficient method and (4) Autocorrelation coefficient method, audio feature patterns are extracted every qIf feature pattern extraction period designated by the timing control unit 3. And audio i! In order to perform this recognition, first input the output of the voice feature extraction section 2 into the standard pattern memory section 5 as shown by the dotted line in the registration mode, and then input the standard pattern of the term specified by the purpose of use into the standard pattern memory section. 5 in advance. Next, the mode is switched to the use mode, and the output of the audio feature extracting unit 2 is stored in the input audio feature pattern memory unit 4 as shown by the solid line.
and store it for voice identification. The identification processing unit 6 uses a similarity calculation method or a pattern! A comparison is made using the matching method, and the input voice is identified and output based on the matched standard pattern.

Here, in the audio feature extraction section 2, if the feature pattern extraction period is set to 10 ms@c, prediction is performed at the 10th order using the linear prediction coefficient method, and the feature pattern is displayed at 8 bits per order, then a single @gg In the case of *, 1@ is usually 0.
5-1. 0.5 to 1.0 kbyte because there is Omec
e/becomes a word. The storage capacity of the input voice feature pattern memory section 4 and the standard pattern memory S for storing these becomes larger. In particular, the storage capacity of the standard pattern memory section 50, which stores standard patterns of many words, becomes enormous.

In general, the phonetic pattern consists of the preceding consonant section and this KlE
Although there are few changes in the characteristic patterns for vowel intervals, the approximated characteristic patterns are all allocated to the input speech characteristic pattern memory section 4 and the standard pattern memory section S at each feature pattern extraction cycle. The number of memories needed to store these data has become larger and larger. Therefore, there is a drawback that the identification process takes a long time and the multi-processing speed cannot be increased.

The present invention eliminates the above-mentioned drawbacks, requires significantly less memory capacity, and improves processing speed. The purpose is to provide a job format. Zero-shot W14 that achieves this purpose is $ll! It consists of a preprocessing unit that amplifies the audio input to a certain level, an audio feature extraction unit that periodically extracts feature patterns from the output signal of the preprocessing unit, and features that are sequentially extracted by the audio feature extraction unit. a distance calculation unit that calculates whether or not the change between the pattern and the already extracted and selected feature pattern is greater than a specific value, and outputs a clock for storing the extracted feature pattern when the change is greater than a specific value; , a selected feature pattern register that stores the extracted feature pattern in place of the already extracted and selected feature pattern by the clock of the skeleton distance calculation unit; a compression number counter that counts the number of feature pattern extraction cycles, that is, the number of compressions until the feature pattern is selected; a standard pattern memory section that stores the standard pattern of each term and the number of times of compression; and a standard pattern memory section that stores the standard pattern of each term and the number of times of compression, and the characteristic pattern, the number of times of compression, and the standard pattern stored in the input voice feature pattern memory section. The present invention is characterized by comprising an identification processing section that recognizes voice input by comparing the standard pattern stored in the printing section and the number of times of compression.

Audio il according to the invention! An example of the identification method is shown in FIG. In FIG. 2, the processing in the preprocessing unit 1 and the audio feature extraction unit 2 is the same as in the case of the first factor. Therefore, the preprocessing section 1 amplifies the input voice to a certain level and detects the voice section. The audio feature extraction unit 2 performs feature extraction,
If feature extraction is performed using the linear prediction coefficient method, the feature pattern extraction period is 10m5ec, the order of the linear prediction coefficient is 10th, and each order is 8 bits, the number of bits of the feature pattern will be 0.5m as explained above. 5 to 1.0 kbytes
/becomes a word.

Audio I of the present invention! In the identification method, feature patterns with little internal carbonization extracted at each feature pattern extraction cycle are discarded and not stored, but are compared with the previously selected feature pattern and temporarily stored, and changes in both are set. The distance calculation unit T calculates whether the distance is larger than the value, and stores only the feature pattern when the distance is larger than the value. That is, the distance calculating section 7 in FIG. 2 compares the output of the audio feature extracting section 2 with the feature pattern that has already been selected and stored in the selected feature register 8, and only when the change is larger than a preset value. , outputs a clock CK for storing the output feature pattern into the selected feature register 8 and the input voice feature pattern memory section 4 or the standard pattern memory section 5.

After one feature pattern is selected in the input voice feature pattern memory section 4, a compression number counter 9 counts the number of unselected feature patterns, assuming that there is no change in the feature until a second feature pattern is selected. The contents of the compression number counter 8 are stored together with the selected feature pattern in the input voice feature pattern memory section 4 or the standard pattern mesori section decoy, and the contents of the compression number counter 9 are cleared by the output of the distance calculation section 7. The identification processing unit 6 uses similarity calculation or pattern matching method, and the input voice feature pattern memory unit 4
The feature pattern is compared with the standard pattern in the standard pattern memory section 5, and the compression number of the compression counter 9 is used as a weighting factor for the compressed voice feature pattern to recognize the voice input.

FIG. 3 shows the speech waveform when GAKKO (school) is pronounced. As shown in FIG. 3, the constant part of the vowel m, 0 occupies v14 of one occurrence interval. Since the feature patterns in the vowel section hardly change, in the method of the present invention, the feature patterns selected in the vowel section are compressed by the distance calculation unit T and are greatly reduced. Therefore, compared to the conventional case, the number of characteristic patterns stored in the memory unit 4 and SK has been reduced to 2 songs in one voice section.

As explained above, according to the present invention, it is possible to eliminate redundancy in audio, and in particular, the memory capacity of the standard pattern mailing section 5, which stores a large number of standard patterns, can be greatly reduced. In addition, compared to the conventional method that requires processing every time there is a change in the feature pattern, the t-order recognition method of the present invention processes and identifies only selected feature patterns. The processing speed increased significantly.

In the embodiment, the preprocessing section 1 detects the voice section, but the voice feature extraction section 2 may also detect the voice section.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a conventional speech recognition method, FIG. 2 is a block diagram of a speech recognition method according to the present invention, and FIG. 3 shows an example of a speech input waveform. In the drawings, 1 is a preprocessing unit, 2 is a voice feature extraction unit, 3 is a timing control unit, 4 is an input voice feature pattern memory unit, 5 is a standard pattern memory unit, 6 is a recognition processing unit, T is a distance calculation unit, 8 is a selection feature register, and 9 is a compression number counter. Patent applicant: Sumitomo Electric Industries, Ltd.

Claims (1)

[Claims] a pre-processing section that amplifies the audio input to a certain level; an audio feature extraction section that periodically extracts feature patterns from the output signal of the pre-processing section; and a feature pattern that is sequentially extracted by the audio feature extraction section. a distance calculation unit that calculates whether or not a change from the extracted and selected feature pattern is larger than a specific value, and outputs an O clock to store the extracted feature pattern when the change is larger than a specific value; A selection feature pattern register that stores the extracted feature pattern in the already extracted and selected feature pattern 0 by a clock from the section, and after one feature pattern is selected, the next feature pattern is selected. a compression number counter that counts the number of feature pattern extraction cycles, that is, the number of compressions until the feature pattern is extracted; and an input that stores the selected feature pattern and the contents of the compression number counter according to a clock from the distance calculation section. an audio feature pattern memory section, a standard pattern memory section that stores the standard pattern of each term, the number of compressions, and the number of compressions stored in the input audio feature pattern memory section; A speech recognition method characterized by having an identification processing section that recognizes speech input by comparing it with a standard pattern and the number of times of compression.