JP2572753B2 - Unspecified speaker consonant identification device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Overview] The syllable patterns related to the voices of a large number of speakers are registered in a dictionary in advance, and the input voice is analyzed. When a device that performs speech recognition by selecting the object with the smallest distance by comparing the
The spectrum showing the articulation position of the consonant itself is unstable, and the spectrum of the transitional part that plays an important role in the identification of the consonant (the spectrum between the consonant and the vowel constituting the syllable)
However, even if the speakers are different, even if the vowels are of the same type, the vowel part patterns are different, so that they change according to the change of the subsequent vowels, so that it is difficult to realize a device with a high recognition rate. There was a problem that said. In order to solve such a conventional problem, the present invention extracts vowels from input speech and identifies their types, and from the dictionary, has a vowel of the same type as this and has a similar vowel spectrum. A technique is disclosed in which syllables are selected as candidates, and a high consonant identification rate can be obtained by comparing these candidates with the analysis results of the input speech.

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a speech recognition device for an unspecified speaker, and more particularly to an unspecified speaker consonant identification device capable of obtaining high recognition efficiency.

2. Description of the Related Art A speech recognition device is useful when performing speech creation for a document or operation of a device that requires a large number of commands. Since there is a difference in voice for each speaker, it is difficult to recognize a voice of an unspecified speaker, especially a consonant. For this reason, a recognition device that registers an enormous amount of speech for each speaker has been put into practical use, but the enormous amount of registration has prevented the spread of the speech recognition device.

FIG. 2 is a block diagram showing an example of the configuration of a conventional speaker-independent speech recognition apparatus.
Represents a syllable pattern dictionary (multi-template).

In FIG. 2, an analysis unit 50 converts an input voice into a digital signal, performs a short-period frequency analysis every several tens of milliseconds, and outputs a resultant frequency spectrum. Collation unit 51
Matches this with all syllable patterns stored in the syllable pattern dictionary 52, and outputs the syllable that is closest in distance as the identification result.

[Problems to be Solved by the Invention] FIG. 3 is a diagram showing an example of a syllable pattern, in which the vertical axis represents the formant frequency, the horizontal axis represents time, and 53 and 54 represent different syllable patterns. The solid lines indicate the voice input of speaker A, and the dotted lines indicate the voice input of speaker B.

As can be seen in FIG. 3, the characteristics of the consonants in the syllables include the spectrum representing the articulation position of the consonant itself (hereinafter referred to as the articulation position spectrum. Both in the spectrum of the transition to the subsequent vowel. The articulation position spectrum is
It is generally unstable and cannot always be observed, especially in continuous speech. Therefore, it is desired to use the transition part spectrum. However, since the transition spectrum significantly changes depending on the succeeding vowel, it is necessary to change the identification method for each succeeding vowel even for a specific speaker. In the case of an unspecified speaker, even if the vowels of the same type are different from one another, the transition portion spectrum having a different vowel pattern changes, which has a significant effect on consonant identification.

Under such conditions, as described above, in the conventional speaker-independent speech recognition apparatus, it is necessary to register an enormous amount of speech, and it may not always be possible to clearly identify the consonant of the speaker. Therefore, there is a problem that it is difficult to realize a device having a high recognition rate.

In view of such a conventional problem, the present invention provides an unspecified speaker consonant discrimination device that can obtain a high recognition rate by preventing the identification of consonants from being affected by the difference in speakers of vowel parts. It is intended to provide.

[Means for Solving the Problems] According to the present invention, the above-mentioned object is achieved by the means described in the claims.

That is, the present invention provides an analysis unit that converts an input voice into a digital signal and performs a short interval frequency analysis, a vowel extraction unit that extracts a vowel portion pattern of the input voice from an output of the analysis unit, and an extracted vowel portion pattern. A vowel identification unit for identifying which vowel corresponds to, a syllable pattern dictionary storing syllable patterns of a number of speakers prepared in advance, and a syllable pattern having a vowel part pattern corresponding to the vowel identified by the vowel identification unit The first vocal part pattern in each syllable pattern selected by the first dictionary selecting part, and the vowel part pattern in the syllable pattern of the input speech extracted by the vowel extracting part are extracted by extracting the vowel part pattern from the syllable pattern dictionary. A second dictionary selection unit for selecting a candidate syllable pattern from the syllable patterns selected by the first dictionary selection unit, and an output of the analysis unit and a second dictionary selection unit Therefore unspecified speaker consonant identification apparatus and a collation unit for collating the syllable pattern of selected candidate.

[Operation] In the above-described unspecified speaker consonant identification device according to the present invention, not only the type of a vowel following the consonant is considered, but also the present invention considers the speaker characteristics of the vowel part, so that the transient part and the vowel part can be obtained. Eliminate the effects of differences between different speakers. Specifically, among syllable pattern dictionaries created by many speakers, only those having a vowel part spectrum that is very similar to the vowel part spectrum of the input syllable are used for identification. Thus, a high consonant recognition rate can be obtained for the voice of the unspecified speaker.

FIG. 1 is a block diagram of an embodiment of the present invention.
Denotes an analysis unit, 2 denotes a vowel extraction unit, 3 denotes a vowel identification unit, 4 denotes a syllable pattern dictionary, 5 denotes a first dictionary selection unit, 6 denotes a second dictionary selection unit, and 7 denotes a collation unit.

The input voice is digitized by the analysis unit 1 and frequency-analyzed (spectral analysis) by FFT or the like every several tens of ms.

The vowel extractor 2 extracts a portion of the input voice where the power of the voice is large and the spectrum change is small as a vowel portion, and the spectrum of the vowel portion is converted into a vowel identification portion and a second dictionary selection portion 6.
Sent to

The vowel identification unit 3 identifies a vowel using the vowel spectrum of the input voice. In the case of Japanese, identification of five vowels (aiueo) is performed.

The syllable pattern dictionary 4 stores syllable names, vowel names, syllable patterns, vowel patterns, and the like, registered by many speakers.

The first dictionary selection unit 5 selects a corresponding item from the syllable pattern dictionary 4 according to the vowel identification result, and sends it to the second dictionary selection unit 6.

The second dictionary selection unit 6 calculates a distance (similarity calculation) between the vowel pattern of each item selected by the first dictionary selection unit 5 and the vowel pattern of the input voice obtained by the vowel extraction unit 2.
A plurality of the same syllable names are sent to the matching unit 7 from the smaller distance (larger similarity).

The matching unit 7 calculates the distance (similarity calculation) between the syllable transition pattern of the item selected by the second dictionary selection unit 6 and the transition pattern of the input syllable (similarity calculation). One or a plurality of shortest distances (larger similarities) are output as identification results.

Here, the articulation position spectrum of a consonant of a syllable may be included in the syllable transition part pattern.

[Effects of the Invention] As described above, according to the present invention, the information of the syllable transition part can be used for consonant discrimination in consideration of the difference in the spectrum of vowels between speakers, so that a high recognition rate can be achieved. it can.

[Brief description of the drawings]

FIG. 1 is a block diagram of one embodiment of the present invention, FIG. 2 is a block diagram showing a configuration of a conventional speaker-independent speech recognition device, and FIG. 3 is a diagram showing an example of a syllable pattern. 1 ... analysis unit, 2 ... vowel extraction unit, 3 ... vowel identification unit,
4 ... Syllable pattern dictionary, 5 ... First dictionary selection unit, 6 ...
... Second dictionary selection unit, 7 ... Collation unit

Claims (1)

(57) [Claims] An analysis section for converting an input voice into a digital signal to perform a short-period frequency analysis, and a vowel extraction section for extracting a vowel pattern of the input voice from an output of the analysis section. A vowel identification unit (3) for identifying which vowel the extracted vowel pattern corresponds to; a syllable pattern dictionary (4) storing syllable patterns of a large number of speakers prepared in advance; A first lexical selector (5) for extracting a syllable pattern having a vowel part pattern corresponding to the vowel identified by (3) from the syllable pattern dictionary (4), and each syllable pattern selected by the first dictionary selector (5) Is compared with the vowel pattern of the syllable pattern of the input voice extracted by the vowel extraction unit (2), and candidate syllables are selected from the syllable patterns selected by the first dictionary selection unit (5). Pa Second dictionary selection section (6), the analysis unit (1) output and the matching portion for matching the syllable pattern of selected candidate by the second dictionary selection unit (6) for selecting over emissions (7)
And an unspecified speaker consonant identification device.