JPH0224699A - voice recognition device - Google Patents

Abstract

PURPOSE:To speed up pattern collating by omitting preliminary collating which is not necessary for an input voice with short pattern length as to the voice recognition device which performs the preliminary collating from the start end of the input voice according to the similarity of a specific-length pattern. CONSTITUTION:When a voice is inputted from a microphone 1, a feature extraction part 2 extracts a feature quantity. A pattern collating part 4 collates the specific pattern length of the input voice with patterns 3 for preliminary collating which are arranged in increasing overall-length order in the arrangement order to select candidates larger than specific similarity. When the end of the input voice is detected by this preliminary collating and the all pattern lengths of the input and all pattern lengths of standard patterns are different by more than a predetermined length or rate, main collating is started without performing the collating with the preliminary collating patterns corresponding to the standard patterns. Consequently, a short input voice is collated at a high speed.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition device, and more particularly to a recognition device with preliminary matching in pattern recognition.

1st technique - In speech pattern recognition, for an input pattern,
First, recognition candidates are selected through preliminary verification, and then the selected candidates are verified again to select a target object. A known method has been known in the art. A standard dictionary is prepared in advance, which is formed by a pattern of features of a predetermined length from the start of the total pattern length of the object to be recognized, and when the input voice has passed a predetermined length from the start, the dictionary pattern is After matching the input pattern with a predetermined length pattern and selecting candidates with a certain degree of similarity or higher, and detecting the end of the input speech, the selected candidates are matched again using the entire pattern length and the largest similarity is selected. A candidate indicating the degree is output as a recognition result. Also, a method is known in which, after detecting the end of the input voice, standard patterns that deviate from the input pattern length by more than a predetermined length or ratio are removed from the candidates among the candidates selected by preliminary matching, thereby further narrowing down the candidates. ing. There is no problem with this method as long as the preliminary matching ends before the end of the input speech is detected, but if the number of recognition target words increases, the preliminary matching must be performed for all target words. There may be cases where the pattern length is short, and there may be cases where preliminary matching is still being performed even when the end of the input voice is detected. In such a case, even though the input speech has ended, preliminary matching is performed on all words, and then main matching is performed on the selected candidates and the recognition results are output. It will take a long time.

The present invention was made in view of the above-mentioned circumstances.
In particular, it was developed with the aim of enabling high-speed processing in pattern recognition.

In order to achieve the above-mentioned object, the present invention also provides a feature extraction unit that extracts the feature amount of the voice for a name input by voice, and a feature extracted by the feature extraction unit. A standard pattern storage unit that stores the quantity as a standard pattern, and a pattern matching between an input pattern formed by the feature quantity extracted by the feature extraction unit at the time of voice input and the feature quantity stored in the standard pattern storage unit stored in advance. In a speech recognition device equipped with a pattern matching unit that recognizes which standard pattern a human voice corresponds to by performing a step, when inputting a voice, when a predetermined length of the input voice has elapsed from the start point of the input pattern, A pattern formed from a feature amount of a predetermined length is compared with a pattern formed from a feature amount of a predetermined length from the starting end of the standard pattern in order from the smallest total pattern length of the standard pattern (hereinafter referred to as preliminary matching). Starting from the total pattern length of the input pattern, if the preliminary matching is still being performed at the time when the end of the input speech is detected, candidates with a certain degree of similarity or higher are selected.

The preliminary matching is not performed for the standard pattern whose total pattern length is larger than a predetermined length or ratio,
The feature is that only for the candidates selected so far, the standard pattern and the entire pattern length of the input pattern are compared again, the degree of similarity is determined for each candidate, and the candidate showing the greatest degree of similarity is output as the recognition result. This is what I did.

Hereinafter, the present invention will be explained based on examples.

As described in the section on the prior art, there may be cases where preliminary verification is still being performed when the end of the input voice is detected, but this only occurs when the pattern length of the input voice is short. This is because, in the preliminary matching method described here, the time required for preliminary matching is independent of the input pattern because matching is performed for a predetermined length from the start of the input pattern and a predetermined length from the start of the standard pattern for all target words. The present invention was made in view of these circumstances.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention,
FIG. 2 is a flowchart for explaining the operation. In the figure, 1 is a microphone, 2 is a feature extraction section, 3 is a preliminary matching pattern, 4 is a pattern matching section, 5 is a standard pattern storage section, and 6 is a In the result output section, the user first registers a standard pattern in advance. At this time, patterns formed by patterns of a predetermined length from the starting end of each standard pattern are arranged in order from the shortest total pattern length of the standard pattern,
Register it as a preliminary verification pattern.

At this time, without creating a new pattern for credit verification,
A mark indicating the order of length of each pattern length is attached to the standard pattern, and when performing preliminary matching, a predetermined length is taken out from the starting end of the standard pattern in order from shortest to longest while referring to the mark, and the predetermined length pattern of the input pattern is It is also possible to use a method that performs a comparison.

When audio is input from the microphone 1, the feature extraction unit 2
Features are extracted in . When a predetermined period of time has elapsed since the start of the input voice was detected, the pattern matching unit 4 compares the predetermined length pattern of the input voice with the preliminary matching patterns 3 arranged in descending order of total pattern length. This is done in order, and candidates with a certain degree of similarity or higher are selected. If the end of the input audio is detected during this preliminary matching, the total pattern length of the input is compared with the total pattern length of the standard pattern corresponding to the preliminary matching pattern that has not yet been matched, and the total pattern length of the input is If the total pattern length of the standard pattern differs by more than a predetermined length or ratio (for example, if it is set to 50% or more with respect to the input audio, 150% of the input audio of Looms
(in the case of a standard pattern having a pattern length of m5 or more), the standard pattern is not compared with the preliminary matching pattern corresponding to it, and main matching is started. By using such a preliminary matching method, even if the number of target words increases in the matching method described in this embodiment, high-speed matching can be performed for short input speech.

As is clear from the above explanation, according to the present invention, in a speech recognition device that performs preliminary matching based on the similarity of a pattern of a predetermined length from the beginning of input speech, for input speech with a short pattern length, By omitting preliminary matching for unnecessary standard patterns, it is possible to speed up pattern matching.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention.
FIG. 2 is a flowchart for explaining the operation. 1...Microphone, 2...Feature extraction unit, 3...
・Preliminary matching pattern, 4... pattern matching section, 5.
...Standard pattern storage section, 6...Result output section. Figure 2 Figure 1

Claims (1)

[Claims] 1. A feature extraction unit that extracts features of a name input by voice; a standard pattern storage unit that stores the features extracted by the feature extraction unit as a standard pattern; By performing pattern matching between the input pattern formed by the feature values extracted by the feature extraction unit and the feature values stored in the standard pattern storage unit stored in advance, it is possible to recognize which standard pattern the input voice corresponds to. In a speech recognition device having a pattern matching section, when a predetermined length of the input speech has elapsed from the start point during speech input, a pattern formed from a feature amount of a predetermined length from the start end of the input pattern and the start end of the standard pattern are detected. , and a pattern formed from features of a predetermined length, starting from the standard pattern with the smallest total pattern length (hereinafter referred to as preliminary matching), select candidates with a certain degree of similarity or higher, and select the candidates with a certain degree of similarity. If the preliminary matching is still being performed at the time when the end of Without performing any matching, only the candidates selected so far are compared again for the standard pattern and the entire pattern length of the input pattern, the degree of similarity is determined for each candidate, and the candidate showing the greatest degree of similarity is output as the recognition result. A voice recognition device characterized by: