JPS6083093A - Word voice recognition equipment - 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] This invention is based on the characteristics of the beginning of the word sound.

The present invention relates to a word speech recognition device that performs primary selection before the end of word speech utterance, thereby shortening the entire preliminary selection process in terms of time.

Features used for preliminary selection of word sounds include ease of processing, high preliminary selection ability (the amount by which the amount of recognition processing is reduced by preliminary selection), and little effect on the final recognition rate.9
It is common to use global features of the entire word sound from the viewpoint of FIG. 1 shows an embodiment of a conventional word speech recognition device that performs preliminary selection using global features of the entire word speech.

The audio waveform (2) input from the microphone (11) is subjected to frequency spectrum analysis for each analysis frame, for example, in the audio analysis circuit (3), and the time series of feature vectors A:a1a
2...at...al. Below, the time series of this feature vector is input feature pattern/(7 frames in total)
It is called (4). The global feature extraction circuit (5) calculates a Fourier expansion coefficient F1j as shown in the following equation from the input feature pattern (4), and outputs it as a global feature pattern (6).

π(t-1)j F1j=(Σa tlcos −) /L 1°=°(
1) t=1'L Here, ati is the input feature (first feature parameter at time t during turn (4), L is the number of frames in the speech section, Flj is the first % feature parameter j These are the following Fourier expansion coefficients.The global feature registration memory (8) stores the global features (turn 01j) of the registered word sounds W1W2, . . . that have been registered in advance by the speaker in use. In the selection circuit (7).

Global feature pattern 'F1j (6) of input word sound and global feature pattern of registered word sound WI W2... Turn 0
For example, calculate the inter-pattern distance a(y, c) with 1j using the following formula.

The word numbers of several registered word sounds having small values of a(P, c) are output as preliminary selection results (9). Recognition word dictionary registration memo IJ (Lυ is a time series of feature vectors of registered word sounds w1, w2, etc. registered by the speaker using them 5 = e1s2...8M (hereinafter referred to as "registered features" turns) is stored in the recognition processing circuit uI.

Pattern matching is performed between the registered feature pattern of the registered word voice outputted in the preliminary selection result (9) and the input # continuation pattern (4) to calculate the inter-pattern distance.

Output as the word category recognition result az of the registered word sounds with the minimum distance between patterns.

Global feature patterns of human word sounds necessary for preliminary selection (
The calculation in step 6) cannot be started until after the input word voice has been uttered and all frames have been analyzed. On the other hand, it has the disadvantage that the preliminary selection calculation takes up a large amount of time.

In order to eliminate such drawbacks, the present invention performs primary selection based on the initial feature of the input word sound before the end of speech production, thereby reducing the amount of calculation required for preliminary selection based on the primary global feature. The drawings will be explained in detail below.

FIG. 2 shows an embodiment of the present invention, and since the components from the microphone (1) to the recognition result a2 in FIG. 2 are the same as those in FIG. 1, their explanation will be omitted.

Input feature butter 7 for frames at the beginning of the word calculated by the speech analysis circuit (3) Ah=a1 a2...aK(13
is immediately subjected to, for example, frame averaging processing in the word-beginning preliminary selection circuit 04, and nine word-beginning features) are determined.

The word beginning feature registration memo IJH includes the center word W of each cluster when the registered word sounds are divided into a plurality of clusters based on the word beginning feature baku-/. (W5.Wl.1.

etc.) and the word numbers (3, 5, 8, etc.) belonging to each cluster are described. The word-initial preliminary selection circuit αe is a word-initial feature detector.
h<t'3 and the word-head feature registration memory (between the word-head feature pattern e of the cluster center word W in lη) is calculated, and several cluster center words with small inter-pattern distances are calculated. Select. Then, all the word numbers in the cluster to which each selected cluster center word belongs are read out from the word-head feature registration memo 'J (17), and outputted as the -next child selection result 0 sand.

The above pre-selection process using word beginnings consists of (a) using the word beginning of the input word as the feature quantity, and (b) calculating the distance between turns using the word beginning feature between the input word audio and the registered word audio. Since it is performed between several cluster center words, the calculation time is short.
For this reason, it is completed before the end of the word vocalization.

In the preliminary selection circuit (7), the distance calculation of the global feature pattern is performed only for the registered word sounds of the primary child selection result a8 output before the end of the word sound utterance. Therefore, according to the number of words to be selected, the preliminary selection circuit (7)
The amount of computation performed within is significantly reduced.

In the above, the words to be recognized are registered in advance by the speaker. Although the case of speaker-specific word speech recognition has been described, the present invention may also be used for speaker-independent speech recognition in which speakers are not limited. In this case, the nine word-initial feature patterns, global feature patterns, and registered feature patterns are created in advance by statistical methods from word speech data from a plurality of speakers.

As described above, according to the present invention, primary selection is performed before the end of word vocalization based on the word-initial feature pattern of the word voice, so the first-order global 4? The amount of calculation for preliminary selection based on characteristic patterns can be reduced, and this has the effect of shortening the time for preliminary selection processing in word speech recognition VC.

[Brief explanation of drawings]

Figure 1 is a diagram showing an example of a conventional word speech recognition device;
The figure is a diagram showing an embodiment of a word speech recognition device according to the present invention. In the figure, (1) is the microphone, (2) is the input audio waveform, (
3) is the speech analysis circuit, (4) is the input feature pattern (all frames), +51 is the global feature extraction circuit, (6) is the global feature pattern, (power is the preliminary selection circuit, (8) is the global feature Registration memo'), (91 is the preliminary selection result,
(II is the recognition processing circuit, συ is the recognition word dictionary registration memory, α is the recognition result, α snoring input feature pattern (word beginning), Q4 is the word beginning preliminary selection circuit, +1!9 is the word beginning feature pattern , σe is the word-initial preliminary selection circuit, aη is the word-initial feature registration memory. QIG is the primary child selection result. In addition, the same or equivalent parts in Figure 1 are indicated with the same symbols. Agent Oiwa Masuo No. 1

Claims (1)

[Claims] Each cluster of registered word sounds is divided into a plurality of clusters according to the characteristics of the beginning of the word in a word speech recognition device that performs preliminary selection of registered word sounds for input word sounds and performs recognition processing on one selected registered word sound. A word-beginning feature registration memory described by the feature of the word-beginning of the central word and the word number belonging to each cluster; Calculates the inter-pattern distance between the cluster center word and the word-head feature in the head feature registration memory, extracts several clusters in order from the cluster with the cluster center word with the smallest inter-pattern distance, and belongs to that cluster. 1. A word speech recognition device comprising: a word-initial preliminary selection circuit that outputs a word number as a primary component selection result.