JPS58102999A - Voice recognition equipment - 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 speech recognition device.

The speech recognition device uses a so-called nokuturn pine that extracts the characteristics of the input speech by frequency analysis, creates an input pattern, and identifies the input speech based on the degree of similarity between the input pattern and a pre-registered standard pattern. There is a ching method.

In such a speech recognition device, input speech is identified by calculating the degree of similarity between an input turn and a standard pattern, that is, performing a pattern matching calculation. This machining calculation is usually processed by a microconverter, but when recognizing the audio of a given word, even if the audio is the same word, the length of the word can be adjusted by varying the speaking speed. The pattern length and the temporal variation of the pattern corresponding to the frequency distribution structure of the voice are not the same. For this reason, in the matching calculation, it is necessary to perform non-linear or linear time transformation on the human cover pattern and compare the degree of similarity with the standard pattern. However, if the difference in word length between the input speech and the registered speech, that is, the difference in the pattern length between the input speech and the standard pattern, is large, the processing time will be longer and the recognition rate will be poor. There was a problem.

SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a speech recognition device that speeds up the processing time for pattern matching calculations and improves the recognition rate.

The speech recognition apparatus according to the present invention is configured to perform 77 checking only when the magnitude of the difference between the word duration of input speech and the word duration of speech used as a standard pattern is within a predetermined time.

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a schematic diagram of a speech recognition device according to the present invention. In FIG. 1, a voice to be registered or identified is input to a microphone 1, and the output signal of the microphone 1, that is, the voice signal, is amplified by an amplifier 2-C and supplied to a feature extraction circuit 3. The feature extraction circuit 3 consists of a filter circuit 4 consisting of a plurality of oBPFs (band pass filters) each connected to the output end of the amplifier 2, and a multiplexer 5, and the output of the filter circuit 4 is The multiplexer 5 sequentially samples the signals in a time-division manner. The output of the multiplexer 5 is sent to the microcomputer 7 via a ~width (analog/digital) converter 6.
supplied to Microcomputer tough is a processor,
Clock generator, memory.

Consists of input/output interfaces, etc., and A /
A standard pattern is created and stored from, for example, an 8-bit digital signal that is the output signal of the D converter 6.

Next, the operation of the microcomputer 7 during recognition will be explained with reference to the operation flow diagram of FIG.

First, the microcomputer 7 requests the input of the voice for recognition using, for example, a display lamp provided on the device, and waits until the voice for recognition is input (11).
). Then, such voice is inputted from the microphone 1, and the characteristics of the voice are extracted by frequency analysis by the feature extraction circuit 3, and then supplied to the microcomputer 7 as a digital signal via the A/D converter 6. Then, an input cover pattern is created from the digital signal (12).

Next, the word duration Tw of the input speech is calculated from the pattern length of the input pattern (13). This word duration Tw is determined based on the power level of the input speech. by the way,
The multiplexer 5 scans the output signal of the BPF consisting of a predetermined number of filter circuits 40, for example eight, at a constant cycle of 19T and supplies the scanned signal to the A/D converter 6. Therefore, the power level of the human voice can be determined by adding a predetermined number of digital signal data generated by the A/D converter 6 forming the input pattern in the order of input. When a voice level with characteristics as shown in FIG. 3 is obtained, normally a time Δ1. The start of the word is set at time 11, which is 10 minutes earlier, and the end of the word is set at time t2, which is [lH'lΔt2 later than the time ty when the voice level falls below the Tokohata level vr2. Therefore, the word time length Tw is t2-61. Next, Microcomputer Tough calculates the magnitude of the difference between the word duration Tw of the input speech obtained as described above and the word duration Twn (n is a natural number) of the speech registered as a certain standard pattern. △T'wn with n set to 1 (14
) (15). Then, it is determined whether the magnitude of the difference iwn is greater than a predetermined value Tr1, for example, 200 m5ec (16). In the case of ΔTwn ) Tr, no matching calculation with that standard pattern is performed, and the degree of similarity with that standard pattern is set to 0 (17).

If △'rwn≦'rr, perform pattern matching calculation between the input cover pattern and the standard pattern, and calculate the similarity 5n (n is a natural number) - (18) o Then, next, the natural number n is the standard pattern. is equal to the number np (19). In the case of n4np, add 1 to n and calculate (2
Return to step (15). In the case of n=np, the maximum value SM is calculated from the respective degrees of similarity Sn calculated by matching calculation.
AX is detected21), and its maximum value SMAX is the predetermined value S.
It is determined whether it is greater than r (22). SMAX≧Sr
In the case of , the maximum value SM is assumed to be the same word as the input voice among the voices registered as a standard pattern.
The identification signal corresponding to the standard pattern that has become AX is output to the recognition output port (23). However, SMAX <
In the case of Sr, an unidentified signal is generated from the recognition output port because the human voice is a pronunciation of a word different from the registered voice (24).

As described above, according to the speech recognition device according to the present invention, pattern matching calculation is performed only when the magnitude of the difference between the word duration of input speech and the word duration of speech registered as a standard pattern is within a predetermined time. As a result, the difference between the word duration of the input speech and the word duration of the registered speech becomes much faster, and mismatching between words with extremely different lengths is eliminated, improving the recognition rate. It is.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of the present invention, FIG. 2 is an operation flow diagram of the microcomputer shown in FIG. 1, and FIG. 3 is a characteristic diagram showing temporal fluctuations in audio level. Explanation of the symbols of the main part 1...Microphone 3...Feature extraction circuit 4...Filter circuit 5...Multiplexer 6...A/D conversion Vessel 7...
...Microcomputer applicant Pioneer Co., Ltd. agent Patent attorney Motohiko Fujimura

Claims (2)

[Claims] (1) A speech recognition device that extracts the features of input speech, creates an input pattern, and matches the input pattern with a pre-registered standard pattern, wherein 1. A speech recognition device characterized in that the matching is performed only when the magnitude of the difference between the word length of the speech and the word duration is within a predetermined time. (2) The speech recognition device according to claim 1, wherein the word duration of the input speech is determined from the level of the input speech.