JPH0222700A - voice recognition device - 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 control of a preliminary selection method used in a speech recognition device, particularly a speaker-specific speech recognition device.

In general, speech recognition for large words is for 500 to several thousand words, and because the number of target words is large, some or all of the characteristic patterns of unknown human power are used in advance, or evaluation functions are used. This method reduces the number of target words for final recognition. However, when a user uses such a system, he or she registers words as necessary to expand his or her vocabulary. Therefore, the number of words gradually increases, and the preliminary selection at the time of short and middle words is the same as that of major words, which sometimes results in a poor recognition rate. Under these circumstances, the preliminary selection matching of the recognition device generally has multiple stages, and only words that have passed several stages of preliminary selection are subjected to final recognition. This preliminary selection method is, for example, the BTS P method (Binary tinges
pecturum pattern) etc., the X at the beginning of the word
It is designed to use frames to perform preliminary selection calculations from the middle of utterance, and to perform final recognition from the end of utterance. However, the preliminary selection method is fixed to the one with the maximum overhead cost, and in order to speed up the processing as much as possible, a unique preliminary selection method is used, so the pass rate for small overhead costs and medium amount of words is not guaranteed.

As mentioned above, the preliminary selection method of the Jingodo speech recognition device is
In general, the aim is to increase speed, and since there are many target words, comparatively simple data or feature patterns are compressed and used.

However, since there are fewer target words for small miscellaneous expenses, it is desirable to perform preliminary selection more accurately and precisely. Furthermore, the number of passing words for each preliminary selection is determined based on whether the degree of similarity of each word is greater than a certain threshold, so the number of passing words varies.

As mentioned above, preselection schemes for conventional speech recognition devices generally use a preselection scheme tailored to the situation at the maximum overhead of the device. Therefore, although a certain level of ability was obtained with the maximum miscellaneous expenses, this led to a decline in ability with the small miscellaneous expenses and Chinese language.

The present invention was developed in view of the above-mentioned circumstances.
In particular, make the most appropriate preliminary selection according to the number of target words,
This was done with the aim of accurately making preliminary selections for small and medium expenses.

In order to achieve the above object, the present invention also provides a preprocessing section that amplifies and corrects audio input from a microphone, a frequency analysis section that performs frequency analysis of the audio, and a frequency analysis section that performs frequency analysis of the audio. It has a feature extractor that obtains a feature pattern using data from the uttered voice, a dictionary registration section that registers the feature pattern of the uttered voice as a dictionary, a dictionary storage section thereof, and a dictionary storage section that matches the feature pattern of unknown input with the dictionary. It has a recognition matching unit that performs recognition, and a result output unit that selects the word with the highest degree of similarity, and during recognition, it preliminarily reduces the number of target words using some or all of the unknown input data. This is a speech recognition device having several stages of preliminary selection sections, characterized in that a recognition method for preliminary selection is controlled based on the number of registered dictionary words or the total number of words to be recognized. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining one embodiment of the present invention. In the figure, 1 is a microphone, 2 is a preprocessing section, and 3 is a group of band-pass filters (B, P).

F), 4 is a feature extraction unit, 5 is a dictionary registration unit, 6 is a preliminary selection data extraction unit, 7□, 7□, ... 7° is a preliminary selection unit,
8 is a passing word counting determination unit, 91, 9□, . . . 9 is a preliminary selection unit, 10 is a dictionary unit, 11 is a dictionary registration counting unit, 12 is a preliminary selection control unit, 13 is a main selection unit, and 14 is a result The output unit first counts the number of registered dictionaries registered in advance and determines the preliminary selection method to be used. That is, when the recognition target word is a large expense, for example, the first preliminary selection 7□ is selected, and a method that is relatively low in accuracy but can be performed at high speed is selected. In addition, in the case of small words, a method is used that requires a little longer processing time but is more accurate.

Similarly, in the next stage preliminary selection, the number of words passed through the previous stage preliminary selection is counted, and the next stage preliminary selection method is determined.

In this way, by varying the preliminary selection method according to the number of target words, it becomes possible to perform preliminary selection according to the number of recognized words possessed by the device, thereby increasing reliability.

Furthermore, the same effect can be obtained by changing the number of target patterns in the preliminary selection or changing the parameter for determining the degree of similarity according to the number of passing words.

As is clear from the above description, according to the present invention, it is possible to perform the most appropriate preliminary selection according to the number of target words, and to accurately perform the preliminary selection for small and medium words. It can be carried out.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention. 1...Microphone, 2...Preprocessing section, 3...
Bandpass filter group, 4...Feature extraction section, 5...
-Dictionary registration unit, 6...preliminary selection data extraction unit, 71
．． 7□. ... 7o... Preliminary selection section, 8... Passing word count determination section. 9□, 9□,...9. ...Preliminary selection section, 10...
Dictionary department. 11... Dictionary registration counting section, 12... Preliminary selection control section. 13... Main selection section, 14... Result output section.

Claims (1)

[Claims] 1. A preprocessing unit that amplifies and corrects the audio input from the microphone; a frequency analysis unit that performs frequency analysis of the audio;
a dictionary registration unit and its dictionary storage unit that includes a feature extraction unit that obtains a feature pattern using data from the frequency analysis unit, registers feature patterns of uttered sounds as a dictionary; and a dictionary storage unit that stores feature patterns of unknown inputs. It has a recognition matching unit that performs matching with the word, and a result output unit that selects the word with the highest degree of similarity. A speech recognition device having several stages of preliminary selection units for reducing words, characterized in that a recognition method for preliminary selection is controlled based on the number of registered dictionary words or the total number of words to be recognized. Device