JPS63223693A - Phoneme pattern registration - 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] [Table of contents] Overview Industrial field of application Prior art (Fig. 5) Problems to be solved by the invention Means for solving the problems (Fig. 1) Working examples ( a) Description of one embodiment (Figures 2, 3, and 4)
Description of other embodiments Effects of the invention [Summary] In a phonological pattern registration method for extracting and registering phonological patterns used for continuous speech recognition from a speaker's speech, By changing the word to be read out next, the registered phoneme distribution is controlled.

[Industrial application field]

The present invention relates to a phonological pattern registration method for registering phonological utterance patterns required in continuous speech recognition based on phonological recognition, and particularly relates to a phonological pattern registration method capable of controlling the appearance frequency of registered phonemes. .

Phonological recognition is attracting attention as one of the speech recognition technologies, and in phonological recognition, it is characterized by a large number of phonological learning patterns as a dictionary.In particular, for continuous speech recognition, it is necessary to extract continuous speech uttered by one person. A pattern is needed.

For this reason, it is effective for K to automatically extract phoneme sections from word utterances and register phoneme patterns, and it is desired that patterns for each phoneme can be obtained with a small number of voice inputs.

[Conventional technology]

FIG. 5 is an explanatory diagram of the prior art.

The speaker S speaks each word into the microphone MP while referring to a predetermined word list shown on the instruction section 9.

The uttered speech is converted into a time series of spectrum and power in the second analysis section 1, segmented into phoneme units by the segmentation section 2 with reference to power dips and spectral change points, and sent to the dictionary reference section 3.

The dictionary reference unit 3 compares the segmentation results with a word dictionary 4 that stores the number of phonemes and the standard length of phonemes for each word, and confirms whether each phoneme of the word has been uttered correctly and whether the segmentation has been performed normally. .

If the segmentation is successful, the phoneme has been extracted from the word speech, and the phoneme pattern and base of the word are given to the phoneme registration unit 5.

The phoneme pattern is registered in memory.

Conversely, if the segmentation is not successful, the phonological pattern of the word is not registered.

[Problem that the invention seeks to solve]

In the conventional phonological pattern registration method, words are uttered in the order of a word list created in advance, and in order to prevent defective phonological patterns from being registered, phonological patterns are not registered from words that cannot be successfully segmented. I was playing K.

The word list is created to include all phonemes in order to register each phoneme pattern, but since the speech patterns of the nine speakers S are all different, there are times when segmentation fails. A pattern emerges.

When a bad phonological pattern occurs like this.

Previously, we had the same speaker repeat the word, but segmentation often failed again.
A problem has arisen in that certain phonological patterns cannot be registered sufficiently.

Also, taking this into consideration, it is conceivable to have two speakers input a wide variety of words aloud.

The problem of inefficient registration was common.

An object of the present invention is to provide a phoneme pattern registration method that can register sufficient phoneme patterns with a relatively small number of words.

[Means for solving problems]

FIG. 1 is a diagram explaining the principle of the present invention.

The present invention separates phonemes from the input speech of the microphone MP that reads out the indicated word list. Extracted by dictionary reference section 3.

In the registration of the phoneme pattern in which the phoneme pattern extracted by the phoneme registration unit 5 is registered, the appearance frequency distribution (distribution of the number of registered phonemes) of the phonemes registered in the phoneme appearance frequency unit 6 is obtained, and the obtained phoneme appearance frequency is referred to. Then, the word list creation section 7 generates a list of words to be uttered next, and the generated word list is displayed on the instruction section 8.

[For production]

In the present invention, the appearance frequency distribution of registered phonemes is monitored,
An evaluation standard is set for the appearance frequency distribution, and words that increase the evaluation value of already registered phonemes during phoneme registration are selected as words for primary registration.

The word is displayed and one speaker is prompted to say it.

That is, by monitoring the registration status of phonemes based on the frequency distribution of registered phonemes and generating a primary word list according to the frequency distribution, it is possible to control the frequency of appearance of registered phonemes. .

〔Example〕

(a) Description of one embodiment FIG. 2 is a detailed explanatory diagram of the present invention, and shows a functional block diagram of a phoneme pattern registration device.

In the figure, the same components as those shown in FIGS. 1 and 5 are indicated by the same symbols, and COMP is a computer that executes the functions of blocks 1 to 7 by calculation;
1 is a spectrum analysis section.

A device that performs spectrum analysis of the audio input through the microphone MP and outputs a spectrum pattern time series; 12 is a power extractor; a component that converts the spectrum pattern time series into a power time series; 21 is a dip unit; This is an extraction section that extracts from the power time series pattern the section in which a dip exists (the position of the dip and the length of the dip) and provides it to the dictionary reference section 3.

61 is a phoneme counter which adds the registered phoneme to the phoneme appearance frequency distribution Pi of the storage unit 62 and calculates the phoneme appearance frequency distribution P.
i is updated, 71 is a word list creation section, which creates a word list from the unused word list section described below so as to have a desired phoneme appearance frequency distribution according to the word list creation flow described below; 72; is the unused word list part.

It stores unused words.

Incidentally, the instruction section 8 is constituted by a display section, and a list of words to be uttered is displayed.

FIG. 3 is a word list creation flowchart of the word list creation section 71 in FIG. 2, and FIG. 4 is an operation explanatory diagram.

First, a list of nine words to be uttered is displayed on the display section 8, and one speaker S sequentially utters the words in the displayed word list.

The uttered word is input to the analysis unit 1, the spectrum analysis unit ll obtains a spectral pattern time series, and the power extraction unit 12 converts it into a power time series pattern. In this embodiment, the power time series pattern is By comparing the size of the dips, that is, the length, depth, number and shape of the dips, with the data in the word dictionary 4, it is evaluated whether the segmentation has been performed normally.

For this reason, the segmentation unit 2 includes the dip extraction unit 2.
1 is provided to extract the section (position, length) where the dip exists from the power time series pattern and perform segmentation.

The word dictionary 4 contains the number of dip intervals for each word.

Shape and size are stored.

The dictionary reference section 3 compares the dip sections extracted by the dip extraction section 21 with the number, shape, and size of the dip sections stored in the word dictionary 4.

Based on the results of this verification, the segmentation results are finally evaluated to determine whether correct segmentation has been performed.

For words determined to have been correctly segmented, the phoneme name and phoneme pattern are given to the phoneme registration unit 5, and the phoneme pattern is registered in correspondence with the phoneme name.

On the other hand, phoneme patterns are not registered for words whose segmentation has been determined to be incorrect.Next, the registered phonemes are passed to the phoneme color/retrieval unit 61 and added to the phoneme appearance frequency distribution Pi in the storage unit 62. added and updated.

When all the words displayed on display section 1 8 have been uttered in this way, word list creation section 7 starts creating a list of words to be read out next.

The list creation unit 7 creates a word list that increases the evaluation value of the distribution of the number of registered phonemes.

A word list to be read aloud at the next registration is created from the phoneme appearance frequency distribution Pi via the phoneme counter 61.

Here, in order to avoid rereading words that have already been read, the primary word list is selected from the unused word list section 72, which is a list of words that have not yet been read.

Generation of this word list will be explained using FIGS. 3 and 4.

In FIG. 3, selecting a word from the unused word list section 72 is called "inputting a word", and selecting an input word as a word to be added to the next word list is called "inputting a word".
We will call it "Adopt a word".

Also, i is the phoneme number i = l...N, nPi is Pi plus the phoneme distribution included in the selected word list, wpi is the phoneme distribution included in the input word, and X
i is the distribution that nPi takes when the input word is adopted, 81
．． 82 is the evaluation value of the phoneme distribution 0. Note that the evaluation value 81.82 indicates that the larger the value, the higher the evaluation 0. First, the list creation section 71 obtains the actual phoneme appearance frequency distribution Pi from the phoneme appearance frequency section 6. , this is set as nPi.

As shown in FIG. 4, the phoneme appearance frequency distribution P1 stores the number of registrations (occurrences) of each phoneme.

(2) The list creation section 71 checks whether all the words have been input from the word list section 72, and when all the words have been input, the process proceeds to step (2).

On the other hand, if all words have not been input, one word is input from the word list section 72.

(2) Next, the list creation unit 71 creates a phoneme distribution wPi (i=1...N) in the input word.

The list creation unit 71 adds wPi to the phoneme appearance frequency distribution npi to obtain Xi.

To control the appearance frequency distribution of registered phonemes so that all phonemes have an equal distribution, use the evaluation function (variance) val
(Xi) is defined as follows.

From the above equation, find val(Xi) and set it to 81.0 Next, find the evaluation function val(np').

And let this be 82.

■ The list creation unit 71 compares 81 and 82, and
If >82, the input word is selected because it is a word with equal phonemes.

Return to step ■.

On the other hand, if 81≦82, all the phonemes are not equal depending on the input word, so the input word is not selected and the process returns to step (2).

(2) When the word list section 72 is checked as to whether or not to be adopted in this manner, the selected words are outputted to the display section 8 as the next word list.

Therefore, as shown in FIG. 4, if there are variations in the phoneme distribution Pi actually registered at the time when a certain word list is finished reading, the value of val(Xi) is not the maximum value. Here, if a word list with a phoneme distribution that cancels out this variation is generated as shown in Figure 4, and all the phonemes in the word obtained here are registered, the appearance frequency distribution of the registered phonemes will be , nPi, etc., the phoneme distribution becomes uniform and the value of val(Xi) increases.

In this way, the registered phoneme distribution can be controlled.

(b) Description of other embodiments In the embodiments described above, 1 variance val(Xi) is performed using equations (1)2 to (2) so that the registered phoneme distribution is uniform. In order to control the intended distribution Ri, the function may be determined as follows.

val(Xi)=-1/(N-z'), Σ(Zt-Z
)”Zi = Xi /Ri)
(3) Z=(1c)ΣZi The same applies to val(nPi).

Furthermore, in order to control the phoneme distribution so that all phonemes appear at least once, a function may be determined as follows.

Let val(Xi)=Σval2(Xi).

Note that the same applies to val(nPi).

Furthermore, the number of words in the word list can be selected arbitrarily, and the number of words to be read at one time is determined, and the number of words can be shown to two speakers. It may also be shown as

Although the present invention has been described above with reference to embodiments, the present invention can be modified in various ways according to the gist of the present invention.

These are not excluded from the present invention.

〔Effect of the invention〕

As explained above, according to the present invention, when automatically extracting phonemes from word utterances and registering phoneme patterns, it is possible to control the registered phoneme distribution, thereby achieving a desired phoneme distribution. A phonological pattern dictionary is obtained.

It is also possible to reduce the number of words used for registration.

This has the effect of improving the efficiency of registration, and is also useful for improving the reliability of phoneme patterns.

[Brief explanation of drawings]

FIG. 1 is a diagram explaining the principle of the present invention. FIG. 2 is a detailed explanatory diagram of the present invention. FIG. 3 is a flowchart for creating a word list in FIG. 2. FIG. 4 is an explanatory diagram of the operation of one embodiment of the present invention. FIG. 5 is an explanatory diagram of the prior art. In the figure, '1... analysis section. 2...Segmentation section. 3...Dictionary reference section. 5... Phonological registration section. 6...Phonological appearance frequency part. 7...Word list creation section.

Claims (4)

[Claims] (1) In a phoneme pattern registration method that extracts phonemes from input speech that reads out a list of words and registers the extracted phoneme pattern, the frequency of appearance of the registered phoneme is determined, and the frequency of appearance of the phoneme is referred to. 1. A phoneme pattern registration method comprising: generating a list of words to be uttered next; and displaying the generated word list. (2) The phoneme pattern registration method according to claim (1), wherein the word list is generated so that the frequency distribution of the phonemes to be registered approaches an intended distribution. (3) The phoneme pattern according to claim (1), wherein the word list is generated such that the appearance frequency distribution of the phonemes to be registered is an even distribution for all phonemes. How to register. (4) The word list is generated such that the frequency distribution of registered phonemes is such that all phonemes appear at least once. Phonological pattern registration method.