JPS6377098A - Unspecified speaker consonant identifier - Google Patents

Abstract

(57) [Abstract] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] [Summary] Syllable patterns related to the voices of many speakers are registered in a dictionary in advance, and after analyzing the input voice, this and each syllable pattern in the dictionary are registered in advance. When attempting to recognize the speech of an unspecified speaker using a device that performs speech recognition by comparing the two consonants and selecting the one with the smallest distance, the spectrum representing the articulatory position of the consonant itself is unstable. ,
The transition spectrum (the spectrum between the consonant and vowel t that makes up the syllable), which plays an important role in distinguishing between consonants, is
It is said that it is difficult to realize a device with a high recognition rate because the vowel part pattern differs even for the same type of vowel between different speakers, and it changes according to the change in the following vowel. In order to solve these conventional problems, the present invention extracts a vowel from input speech, identifies its type, and searches the dictionary for vowels of the same type. This paper discloses a technique that can obtain a high consonant identification rate by selecting syllables with similar vowel spectra as candidates and comparing these candidates with analysis results of input speech.

"Field of Industrial Application" The present invention relates to a speech recognition device for unspecified speakers, and is capable of obtaining a particularly high identification rate. This relates to speaker-independent consonant identification ViW.

BACKGROUND OF THE INVENTION Speech recognition devices are useful when creating documents or operating devices that require a large number of commands using voice. Since the voices of each speaker are different, it is difficult to recognize the voices of unspecified speakers, especially consonant recognition. For this reason, recognition devices that register a huge amount of speech for each speaker have been put into practical use, but this huge number of calls, N, has prevented the widespread use of speech recognition devices.

FIG. 2 is a block diagram showing an example of the configuration of a conventional speaker-independent speech recognition device, in which 50 is an analysis section, 51 is a collation section,
52 represents a syllable pattern dictionary (multi-template).

In FIG. 2, an analysis unit 50 converts input audio into a digital signal, performs short-term frequency analysis every several tens of milliseconds, and outputs the resulting frequency spectrum. Collation section 5
1 compares this with all the syllable patterns stored in the syllable pattern dictionary 52 and outputs the closest syllable as the identification result.

B. 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 formant frequency and the horizontal axis represents time.
4 are different syllable patterns, with the solid line representing speaker A's voice input and the dotted line representing speaker B's voice input.

As seen in Figure 3, the characteristics of a consonant within a syllable are determined by the spectrum representing the articulatory position of the consonant itself (hereinafter referred to as the articulatory position spectrum; for example, in the case of a plosive, this is the plosive part spectrum) and the consonant's position of articulation. Both sides of the spectrum transition to the am vowel. The articulatory position spectrum is −
It is unstable to the touch and cannot always be observed, especially during continuous speech. Therefore, it is desirable to utilize the transition spectrum. However, the transition spectrum changes significantly depending on the following vowels, so even in the case of a specific speaker, it is necessary to change the classification method for each subsequent vowel. In the case of unspecified speakers, even if the vowel is of the same type, the vowel pattern will differ between different speakers, and the transition spectrum will also change, which will have a significant impact on consonant identification.

Due to these conditions, as mentioned above, conventional speaker-independent speech recognition devices need to register a huge amount of speech, and it may not always be possible to clearly identify consonants of any speaker. Therefore, there was a problem in that it was difficult to realize a device with a high recognition rate.

In view of these conventional problems, the present invention provides a speaker-independent consonant identification device that can obtain a high recognition rate by eliminating the influence of differences between speakers in the vowel part on the identification of consonant parts. is intended to provide.

[Means for Solving the Problems] According to the present invention, the above-mentioned object is as described in the claims, by converting input audio into a digital signal and converting it into a short-range frequency signal every several tens of milliseconds. An analysis section that performs analysis, a vowel extraction section that extracts the vowel pattern of input speech from the output of the analysis section, a vowel identification section that identifies the type of vowel in the extracted vowel section, and a large number of pre-prepared speech patterns. a syllable pattern dictionary that stores syllable patterns for each syllable divided into vowel part patterns and transition part patterns, and a first dictionary that extracts syllable patterns having vowel parts that correspond to the identification results of the vowel identification section from the syllable pattern dictionary. The selection unit calculates the distance between the vowel part pattern of each syllable pattern dictionary selected by the first dictionary selection part and the vowel part pattern of the input speech extracted by the vowel extraction part, and calculates the distance for each syllable name with a small distance. a second dictionary selection section that selects a plurality of syllables as candidates from the second dictionary selection section; and a matching section that determines an identification result by calculating the distance between the output of the analysis section and the candidate syllable pattern selected by the second dictionary selection section. This is achieved by a speaker-specific consonant identification device characterized by comprising:

[Function] In the speaker-independent consonant identification device according to the present invention described above, it goes without saying that the type of vowel following the consonant is taken into account.
In the present invention, the influence of differences between speakers in the transitional part and the vowel part is removed by considering the speaker characteristics of the vowel part.

Specifically, this is achieved by using only those syllable pattern dictionaries created by a large number of speakers that have a vowel spectrum that is very similar to the vowel spectrum of the input syllable. Therefore, it is possible to obtain a high consonant identification rate for speech of unspecified speakers.

[Embodiment] FIG. 1 is a diagram of an embodiment of the present invention, in which 1 is an analysis section, 2 is a vowel extraction section, 3 is a vowel identification section, 4 is a syllable pattern dictionary, and 5 is a first dictionary. The selection section, 6 represents a second dictionary selection section, and 7 represents a collation section.

After the input voice is digitized by the analysis unit 1, it is converted into a number +
Frequency analysis (spectrum analysis) is performed using FFT or the like for each IIS.

The vowel extractor 2 extracts a portion of the input voice with large power and little spectral change as a vowel portion, and sends the spectrum of the vowel portion to the vowel identifier and second dictionary selector 6.

The vowel identification unit 3 identifies vowel parts using the vowel part spectrum of the input speech. In the case of Japanese, five vowels (aiweo) are identified.

The syllable pattern dictionary 4 includes the following patterns registered by many speakers:
Syllable names, vowel names, syllable patterns, vowel patterns, etc. are stored in the format shown in Table 1.

The first dictionary selection section 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 section 6.

Table 1 The second dictionary 9 selection unit 6 calculates the 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 speech obtained by the vowel extraction unit 2. Then, for each same syllable name, 11 or more names with small distances (high degrees of similarity) are sent to the matching unit 7.

The matching unit 7 performs a distance meter X (11 similarity calculation) between the syllable transition pattern of the item selected in the second dictionary selection unit 6 and the input syllable transition pattern to find the smallest distance (highest similarity). One item or a plurality of items with a small distance (high degree of similarity) are output as identification results.

Here, the articulatory position spectrum of the consonant of the syllable may be included in the syllable transition pattern.

[Effects of the Invention] As explained above, according to the present invention, information on syllable transition parts can be used for consonant identification in consideration of differences in vowel spectra depending on the speaker, so a high recognition rate can be achieved. can.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of an embodiment of the present invention, FIG. 2 is a block diagram showing the configuration of a conventional speaker-independent speech recognition device, and FIG. 3 is a diagram showing an example of a syllable pattern.

Claims (1)

[Claims] An analysis section (1) that converts input speech into a digital signal and performs short-range frequency analysis every several tens of milliseconds, and a vowel part pattern of the input speech from the output of the analysis section (1). a vowel extraction unit (2) that extracts the vowel part; a vowel identification unit (3) that identifies the type of vowel in the extracted vowel part; and a vowel identification part (3) that identifies the vowel type of the extracted vowel part. a syllable pattern dictionary (4) that is stored separately from the vowel part patterns; and a first dictionary selection part (5) that extracts from the syllable pattern dictionary (4) a syllable pattern having a vowel part that corresponds to the identification result of the vowel identification part (3). ), and the distance between the vowel part pattern of each syllable pattern dictionary selected by the first dictionary selection part (5) and the vowel part pattern of the input speech extracted by the vowel extraction part (2) is calculated, and the distance is calculated from the one with the smaller distance. a second dictionary selection unit (6) that selects a plurality of syllables from as candidates; and an output of the analysis unit (1) and a second dictionary selection unit (6).
1. A speaker-independent consonant identification device comprising: a collation unit (7) that calculates a distance from a candidate syllable pattern selected by the above method to determine an identification result.