JPS63279297A - Discrete word 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 [Field of Industrial Application] The present invention relates to a discrete word speech recognition device that improves a method for detecting speech sections of input speech.

[Conventional technology]

Conventionally, template matching methods, syntactic recognition methods, statistical determination methods, and the like are known as methods for discrete word speech recognition in discrete word speech recognition devices.

FIG. 4 is a block diagram showing the configuration of a conventional discrete word speech recognition apparatus using the template matching method disclosed in, for example, "Speech Recognition" (Chapter 3 Preprocessing for Speech Recognition) by Higashi Niimi. In the figure, 1 is an acoustic analysis circuit that analyzes input speech and converts it into features and parameters, 2 is a speech section detection circuit that detects speech sections of input speech, 3 is a standard pattern memory that stores standard patterns, and 4 is a standard pattern memory that stores standard patterns. A pattern matching circuit performs pattern matching between a standard pattern that is already stored and a newly input input pattern; 5 is an input voice; 6
are feature parameters, 7 is an input cover turn, 8a and 8b are standard/signature turns, and 9 is a recognition result.

In the conventional discrete word speech recognition device configured as described above, before recognizing an input pattern, a standard pattern is first registered, and a newly input input pattern is recognized based on this standard pattern. . That is, the standard pattern is registered by inputting the standard pattern 8a into the standard pattern memory 3. When recognizing,
This is performed by inputting the input cover turn 7 to the pattern matching circuit 4 and performing pattern matching with the standard pattern 8b, and the word of the standard pattern 8b having the highest degree of similarity is selected as the recognition result 9. Also when registering standard pattern 8a,
Even when recognizing the introductory pattern 7, the voice interval detection circuit 2 performs the same voice interval detection process, and the detection method is based on the detection performance of the phonemes that exist at the beginning and end of the voice that uttered the word, and the noise. This system simultaneously takes into consideration the discrimination performance between the speech interval and the speech interval.

[Problem that the invention seeks to solve]

Since the above conventional discrete word speech recognition device is configured as described above, it is possible to simultaneously achieve the function of detecting phonemes existing at the beginning and end of speech and the function of discriminating between noise and speech sections. It is difficult to detect all phonemes, and if you try to detect all the phonemes, you will end up mistaking noise for a speech interval, and conversely, if you try to absolutely prevent noise from being mistaken for a speech interval, you will end up with some phonemes that cannot be detected at all. For this reason, if we try to achieve this by detecting a fixed speech interval as in the conventional discrete word speech recognition device, there will be a trade-off between the phoneme detection performance and the discrimination performance between noise and speech intervals. If there are phonemes that are difficult to distinguish and the pronunciation is weak or unstable, words with these phonemes at the beginning or end will cause errors in the detection of phonetic intervals, and this will cause However, when the standard pattern is registered, the standard pattern cannot be created correctly, and when the input cover pattern is recognized, the input cover pattern cannot be detected correctly, so there is a problem that the recognition rate deteriorates.

The present invention has been made to solve this problem, and is capable of stably detecting voice intervals by changing the voice interval detection method for each standard pattern to be registered or for each standard pattern to be pattern matched. It is an object of the present invention to provide a discrete word speech recognition device in which the recognition rate can be improved.

[Means for solving problems]

The discrete word speech recognition device according to the present invention includes an acoustic analysis circuit that analyzes input speech and converts it into feature parameters, a speech section detection circuit that detects a speech section of the input speech, and a feature of the speech section of the input speech. In addition to the recognition means that recognizes the time series of parameters, it is equipped with a pronunciation table that stores pronunciation information of the words to be recognized, and the above-mentioned speech interval detection circuit uses the pronunciation information on the pronunciation table according to a predetermined rule. The method for detecting the voice section is changed.

[Effect]

In the discrete word speech recognition device of the present invention, before registering a standard pattern and recognizing an input pattern, pronunciation information of the word to be recognized is stored in a pronunciation table, and based on the pronunciation information on this pronunciation table, the standard pattern is The method of detecting a voice section of the voice section detection circuit is changed for each standard I-turn that is to be registered in the turn memory or for each standard pattern that is to be pattern matched by the pattern matching circuit.

It is now possible to detect phonemes that are difficult to detect using the detection method of a fixed speech interval in conventional devices by using the detection method of the speech interval corresponding to the phoneme number, resulting in stable speech interval detection processing. Since this is carried out, an improvement in gas can be achieved.

〔Example〕

FIG. 1 is a block diagram showing the configuration of a discrete word speech recognition device which is an embodiment of the present invention. In the figure, 1 is an acoustic analysis circuit that analyzes input speech and converts it into feature parameters; 2 is a speech section detection circuit that detects speech sections of input speech; 3 is a standard pattern memory that stores standard patterns;
4 is a pattern matching circuit that performs pattern matching between the already stored standard pattern and the new 6ζ input input A turn, 5 is the input voice, 6 is the stick parameter, 7 is the input turn, 8m, 8b are Standard pattern, 9 is recognition result, 1
0 is a pronunciation table that stores pronunciation information of words to be recognized;
lla and llb are pronunciation information.

FIG. 2 is a flowchart showing the operation of the speech section detection circuit in the discrete word speech recognition device shown in dX1.

FIG. 3 is a characteristic diagram showing characteristic parameters obtained by the acoustic analysis circuit in the discrete word speech recognition device 1ζ of FIG. 1.

Next, the operation of the discrete word speech recognition apparatus shown in FIG. 1, which is an embodiment of the present invention, will be explained. In the discrete word speech recognition device according to the present invention, prior to recognizing the input pattern 7, the pronunciation information 11a and the elongated pattern 8a are registered. Registration of the pronunciation information 111L involves inputting and storing the pronunciation information 11a on the pronunciation table 10. Further, when registering the standard pattern 8a, the standard pattern 8a is stored in the standard pattern memory 3, similar to the conventional discrete word speech recognition device. However, the method for detecting a speech section in the speech section detection circuit 2 according to the present invention is not a method for detecting a fixed speech section as in the conventional discrete word speech recognition device, but a standard method based on the pronunciation information 11b on the pronunciation table 10. It is determined by the voice section detection rule based on the standard pattern 8a1 to be registered in the pattern memory 3.

In addition, when recognizing the input pattern 7, the input pattern 7 is input to the pattern matching circuit 4 and is pattern-matched with the standard pattern 8b, similar to the conventional discrete word speech recognition device. Standard pattern 8b
is selected as the recognition result 9. However, the method of detecting a speech section in the speech section detection circuit 2 is not a method of detecting a fixed speech section as in the conventional discrete word speech recognition device, but a method of detecting a speech section in the pattern matching circuit 4 based on the pronunciation information 11b on the pronunciation table IO. It is determined by the detection rule of the voice section based on the standard pattern 8b for which pattern matching is to be performed.

Next, an example of the voice interval detection rule in the voice interval detection circuit 2 is based on the standard pattern ``pear'' (pronunciation information is ``NA3-ko'').
) will be explained. A flowchart showing the operation of the speech section detection circuit 2 according to the speech section detection rules is shown in FIG. ing. As shown in Figure 3, (3-1) is the threshold for detecting voice sections, (3-2) is the full-band spectrum normalized by the threshold, and (3-3) is the threshold for detecting voice sections. ) is the vector power in the 200 Hz to 800 Hz band normalized by the threshold, (3-4) is 2 normalized by the threshold,
The 5KHz to 5Kllz band is the lowest. First, from the full-band switch power (3-2), the third
Section BC in the figure is extracted as a voice section (step 2-1 in FIG. 2). Next, the beginning of the word is rNJ (step 2-2m), and rNJ is the full-band spectral power (3-
2) alone may be missing/ratio, so 20011
z ~ 800 is corrected using the band squictor power (3-3), and section AC in Figure 43 is extracted (Step 2-
2b).

In addition, the ending of the word is "sIJ" (step 2-4a), and rsIJ pronounced low due to accent may be devoiced, so 2.5 K
Hz ~ 5K
4) to correct the ending of the word and extract the section AD in FIG. 3 (step 2-4b). The voice section detection circuit 2 finally detects the section AD in FIG. 3 as a voice section. The voice interval detection rule in the voice interval detection circuit 2 described above is merely an example, and the present invention is not limited thereto.

In the above embodiment, the template matching method was used as an example of a method for discrete word speech recognition, but it can also be applied to other methods for related discrete word speech recognition, such as syntactic recognition and statistical decision methods. , the same effect as the above embodiment is obtained.

〔Effect of the invention〕

As explained above, the present invention provides a discrete word speech recognition device that includes a pronunciation table that stores pronunciation information of words to be recognized, and uses a standard pattern to be registered in a standard pattern memory based on the pronunciation information on the pronunciation table. Alternatively, the method for detecting speech sections of the speech section detection circuit is changed for each standard pattern that the pattern matching circuit attempts to match, so that it is difficult to detect using the fixed method of detecting speech sections in conventional devices of this type. It is now possible to detect different phonemes at any time by using the method for detecting the speech interval corresponding to that phoneme, and this has the excellent effect of improving the recognition rate because stable speech interval detection processing is performed. It is something.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a discrete word speech recognition device that is an embodiment of the present invention, and FIG. 2 shows the discrete word speech recognition device shown in FIG. 1! Flowchart showing the operation of the speech interval detection circuit in IT, Figure 3 is a characteristic diagram showing the characteristic parameters obtained by the acoustic analysis circuit in the discrete word speech recognition device of Figure 1, and Figure g4 is the conventional discrete word speech recognition AGC. FIG. 2 is a block diagram showing the configuration of the device. In the figure, 1...Speech analysis circuit, 2...Speech section detection circuit, 3...Standard pattern memory, 4...Pattern matching circuit, 5...Input voice, 6...Characteristic parameters , 7... Input cover pattern, 8a, 8b... Yanagi quasi pattern, 9... Recognition result, 10... Pronunciation table,
lla, 11b... pronunciation information, (2-1), (
2-2a), (2-2b), (2-31), (2-3
b), (2-41L), (2-4h)-, step, (
a - t)...Threshold value for detecting voice sections, (3-2)...The entire band normalized by the threshold value is a vector, (3-3)...Threshold value 200tlz ~ 800112 band spectral power, normalized by the value
(3-4)...2,5K normalized by threshold
It has high power in the Hz to 5KHz band. In each figure, the same reference numerals indicate the same or equivalent parts.

Claims (1)

[Claims] an acoustic analysis circuit that analyzes input speech and converts it into a feature parameter; a speech section detection circuit that detects a speech section of the input speech; and a recognition means that performs recognition on a time series of feature parameters of the speech section of the input speech. and a pronunciation table for storing pronunciation information of words to be recognized, and a switching means for switching the method of detecting a speech section of the speech section detection circuit at any time based on the pronunciation information on the pronunciation table. Speech recognition device.