JPS6410080B2 - - Google Patents

Description

[Detailed Description of the Invention] (1) Technical Field of the Invention The present invention provides a standard pattern that can register a standard pattern that can recognize an input speech pattern regardless of whether the last syllable of the input speech pattern is devoiced or unvoiced. Regarding registration method.

(2) Background of the technology As one of the recognition methods in speech recognition systems, a technique is used in which a standard pattern is registered in advance, this pattern is compared with an input speech pattern, and the input speech is recognized based on the match. . In this technique, the last syllable of the input speech pattern becomes devoiced depending on the word, which is an obstacle in speech recognition.

(3) Prior Art and Problems In order to solve these inconveniences, for words whose final syllables tend to be devoiced, the words are uttered several times and the averaged pattern is registered as a standard pattern. Techniques are used, such as matching the input speech pattern to be recognized, or registering a plurality of patterns obtained by the above-mentioned utterances and matching them for speech recognition.

However, these methods not only impose a burden on the user when registering the standard patterns but cannot expect to improve the recognition rate, and in the case of the latter technique, the storage capacity required to register the standard patterns also increases. Inevitable.

(4) Purpose of the Invention The present invention was devised in view of the drawbacks of the conventional techniques as described above, and its purpose is to improve the recognition rate without imposing any burden on the user, and without increasing the storage capacity. The objective is to provide a standard pattern registration method.

(5) Structure of the Invention The purpose of this invention is to register the standard pattern of a word whose final syllable is easily devoiced, and to convert the final syllable of the uttered speech pattern of the word into the devoiced final syllable obtained from the average speech pattern. First and second synthesized speech patterns are generated by replacing the tail syllable and the last syllable that has not been devoiced, and these synthesized speech patterns are compared with the uttered speech pattern to generate a synthesized speech pattern with a large matching distance and a synthesized speech pattern. This is achieved by registering the uttered voice pattern as a standard pattern.

(6) Embodiments of the invention Hereinafter, embodiments of the invention will be described with reference to the accompanying drawings.

FIG. 1 shows a speech recognition system implementing the method of the present invention. 1 is a microphone, 2 is a frequency analysis section, 3 is a parameter extraction section, 4 is a segmentation section, 5 is a pattern buffer, 6 is a switching means, 7 is a pattern matching section, and 8 is a standard pattern storage section. It constitutes a conventional speech recognition system.

A standard pattern creation section 9 is interposed between the pattern buffer 5 and the standard pattern storage section 8, and the present invention is implemented within the above-mentioned conventional speech recognition system. The pattern buffer 5 also constitutes a part of the structure of the present invention, and a specific structure of the present invention is shown in FIG.

In FIG. 2, 10 is a pattern buffer (same as reference numeral 5 in FIG. 1).
Also, the last syllable is indicated by 0. 11,1
2 are in synthesis sections that receive input speech patterns from the pattern buffer 10, respectively. The synthesis section 11 is connected to a devoiced syllable pattern supply section 13 .

This supply section stores a pattern of the last syllable that is not devoiced, which is determined from an average speech pattern obtained from the utterances of a plurality of people, and supplies it to the synthesis section 11 .

Furthermore, the synthesis section 12 is connected to a devoiced speech pattern supply section 14 . This supply section 14 stores a devoiced last syllable pattern obtained from the average speech pattern obtained from the utterances of a plurality of people, and supplies it to the synthesis section 12.

15 and 16 are synthetic speech pattern storage units, respectively.

17 and 18 are matching units, respectively, and the matching unit 17 includes the synthesized speech pattern storage unit 15 and the pattern buffer 1.
0, matches the first synthesized speech pattern from the storage section 15 and the input speech pattern from the pattern buffer 10, and outputs the matching distance between them. The matching section 18 is also connected to the synthetic speech pattern storage section 16 and the pattern buffer 10, and matches the second synthetic speech pattern from the storage section 16 with the input speech pattern from the pattern buffer 10. , output the matching distance between them.

19 is a pattern selection section, which is configured to select a synthesized speech pattern and an input speech pattern with a large matching distance and supply them to the standard pattern storage section 20 (same as reference numeral 8 in FIG. 1). There is.

Next, the process of registering a standard pattern to be matched with an input voice pattern using the apparatus shown in FIG. 2 will be explained.

The syllables consist of a ₁ , a ₂ , ......a _N , and the last syllable is a _N
A word A that is likely to be devoiced is generated toward the microphone 1, and its output signal is processed by the frequency analysis section 2, parameter extraction section 3, and segmentation section 4 as in the conventional case, and the input speech pattern A _P
is placed in the pattern buffer 10.

Prior to this, the devoicing pattern supply unit 13
, the pattern a ^T _N of the unvoiced last syllable a _N described above is stored, and the devoicing pattern supply unit 14 stores the pattern a ^U _N of the devoiced final syllable a _N described above. The respective patterns are supplied from the respective supply sections to the synthesis sections 11 and 12 when the input audio pattern ^AP is output from the pattern buffer 10.

The following processing is performed in the combining units 11 and 12 that receive these supplies. That is, the synthesis section 11
, the last syllable of the input speech pattern A ^P
The pattern of a _N and the unvoiced pattern a ^T _N are exchanged to generate a first synthesized speech pattern A ^T [the last syllable of which is indicated by "1"],
The signal is supplied to the synthesized speech pattern storage section 15 and stored there. In addition, in the synthesis unit 12, the pattern of the last syllable a ^N of the input speech pattern A _P is replaced with the devoiced pattern a ^U _N to form a second synthesized speech pattern A ^U [the last syllable is "2"]. ] is generated, and the synthesized speech pattern storage unit 1
6 and stored there.

The first synthesized speech pattern A ^T in the synthesized speech pattern storage section 15 and the input speech pattern A ^P in the pattern buffer 10 are supplied to the matching section 17, these two patterns are matched, and the matching distance between them is output. Ru. In parallel with this, the second synthesized speech pattern A ^U of the synthesized speech pattern storage section 16 and the input speech pattern A ^P of the pattern buffer 10 are supplied to the collation section 18, and these two patterns are collated. The matching distance of is output.

Both of these matching distances are supplied to the pattern selection section 19, where the synthesized speech pattern A ^U or ^AT with the larger matching distance is selected and output, and the input speech pattern ^AP is also output.

Both of these patterns are stored in the standard pattern storage section 20 as standard patterns.

Such registration is completed in a single operation. In addition, the standard pattern includes the sound pattern uttered for the word and the last syllable for the word, regardless of whether or not the word uttered is devoiced for registration as described above. Since the standard pattern includes a speech pattern in which the last syllable is not devoiced or a speech pattern in which the last syllable is devoiced, after registration as described above, the word corresponding to the registered standard pattern is devoiced and uttered. Since the speech can be recognized regardless of whether the speech is uttered without devoicing, the recognition rate can be improved. In addition, compared to the case where a large number of standard patterns are registered for each word, not only does the storage capacity become smaller, but the processing is also simplified.

(7) Effects of the invention As is clear from the above explanation, according to the present invention, the following effects can be obtained.

(1) Standard patterns can be registered in one operation.

(2) Therefore, the registration process can be simplified.

(3) In registering such standard patterns, the speech recognition rate can be improved.

[Brief explanation of drawings]

FIG. 1 is a diagram showing the configuration of a speech recognition system implementing the present invention, and FIG. 2 is a diagram showing an embodiment of the present invention. In the figure, 10 is a pattern buffer, 11 and 12 are synthesis units, 13 is a devoiced syllable pattern supply unit, and 1
4 is a devoiced syllable pattern supply unit, 15 and 16 are synthesized speech pattern storage units, 17 and 18 are collation units, 1
9 is a pattern selection section.

Claims (1)

[Claims] 1. When registering a standard pattern for a word whose final syllable is likely to be devoiced, the final syllable of the vocalized speech pattern of the word is determined by devoicing the final syllable obtained from the average speech pattern and the final unvoiced syllable. , generate first and second synthesized speech patterns, compare these synthesized speech patterns with the uttered speech pattern, and use the synthesized speech pattern with a large matching distance and the uttered speech pattern as a standard pattern. A standard pattern registration method characterized by registration.