JPH0558553B2 - - Google Patents

Description

[Detailed Description of the Invention] [Table of Contents] Overview Industrial Application Fields Conventional Technology Problems to be Solved by the Invention Means for Solving the Problems Examples Effects of the Invention [Summary] Pre-registered Words When recognizing input speech that has a speech pattern, find the inter-pattern distance between the speech pattern of the unknown input word and the registered word speech pattern, and select the word of the registered word speech pattern that gives the minimum distance as the recognition result. In the word speech recognition method, (1) when registering the registered word speech pattern, by providing a storage unit that stores the method of section detection for each word or for each monosyllable, Therefore, the method of detecting the section is changed for each word or single syllable to be registered. (2) The output of the section detection unit for unknown input speech is selected for each word or monosyllable read out from a group of registered patterns to be compared, and is compared and recognized.

[Industrial application field]

The present invention has registered word speech patterns in advance, and when recognizing input speech, calculates the inter-pattern distance between the speech pattern of the unknown input word and the registered word speech pattern, and the registered word giving the minimum distance. The present invention relates to a word speech recognition method that uses words of a speech pattern as a recognition result, and particularly relates to a method for registering registered word speech patterns or a method for recognizing unknown input speech.

Generally, when speaking Japanese, for example, (ki)
(ku) (shi) (su) (chi) (tsu) (hi) (fu) (pi) (
)
When a syllable (beat) with a consonant (k), (g), (t), (h), or (p) is followed by a vowel (i) or (u), such as (shiyu), the posture of the vowel alone is , actually (i) (u)
There is a phenomenon in which vowels do not sound voiced, which is called "vowel devoicing."
It is called. {“Japanese Vocal Accent Dictionary”
Refer to the NHK edition} If such a syllable that is easily devoiced is at the beginning or end of a word, the voice segment detection processing performed in general speech recognition processing may not drop the devoiced portion. In many cases, when registering a standard voice pattern, a voice pattern in which the devoiced portion is omitted may be registered.

In addition, with recent advances in computer technology, document processing systems, so-called word processors, have become popular. However, in normal word processors, the processing is limited to words or monosyllables entered using keys, so the operation is difficult. Due to the problem of poor performance, voice word processors have recently been put into practical use. In this case, the problem is how to improve the registration quality of monosyllables.

In this way, the registered standard speech pattern and the unknown input speech pattern are compared, and the most similar pattern, specifically, the above-mentioned standard speech pattern with the minimum distance between patterns is recognized as the speech. In speech recognition in a recognition device, the same problem as in the case of standard speech pattern registration exists because preprocessing includes section detection processing after extraction of feature parameters for the unknown input speech.

Under these circumstances, there has been a need for a section detection method that can register high-quality standard speech patterns and extract high-quality speech when recognizing unknown input speech.

[Conventional technology]

FIG. 7 is a diagram illustrating a conventional method for registering standard speech patterns and recognizing unknown input speech.

First, from a registered voice input from a microphone, a recognition parameter representing the characteristics of the voice pattern is extracted by a parameter extraction section 1.

In extracting this recognition parameter, for example, BPF analysis using a group of band-pass filters,
Methods such as linear predictive analysis (LPC) are known.

In the above BPF analysis, the audio input from the microphone is processed at a specific sampling period (for example, 18 ms).
The digitized spectrum intensity of the audio energy sampled in is used as the recognition parameter.

Using the extracted recognition parameters, the next section detecting section 2 detects the section where the voice truly exists, and uses the recognition parameters of that portion as data for pattern matching.

Specifically, for example, the parameter extraction unit 1
The voice power is calculated based on the recognition parameters of the input voice created in step 1, the voice power is checked against a certain threshold value, and the portion larger than the threshold value is set as the voice presence section.

In addition, there are various methods to increase the detection accuracy of the section, such as combining other parameters such as zero crossing and autocorrelation in addition to the audio power threshold, but basically, once the section is detected, Once a method was determined, the same method was used to detect intervals without making any changes.

Then, the result of the section detection was registered in the speech dictionary (registered pattern group) 6 as a standard speech pattern without performing any checks.

Next, in the method of recognizing unknown word speech, the speech pattern of the unknown input word extracted by the same method and the standard speech are extracted from the previously registered speech dictionary (registered pattern group) 6. The sequentially extracted patterns are compared using, for example, a pattern matching method, the distance between the two is calculated, and the standard speech pattern with the minimum distance is detected and used as the recognition result.

Another conventional recognition method for the above unknown input speech is to have a plurality of threshold values in advance in the section detecting section 2, find a plurality of sections, and calculate the standard speech pattern for each section. The idea was to compare all of them, select the closest combination, and use it as the recognition result.

However, in this method, each registered pattern must be compared with unknown input audio patterns corresponding to the plurality of threshold values, so there is a problem that the amount of processing increases significantly, and There was a problem with this because it contained many erroneous voice patterns, which caused erroneous recognition.

[Problem that the invention seeks to solve]

When detecting sections using a uniform conventional method like this, there are problems with voice sections being dropped and noise being added.For example, if you try to reduce the dropouts, the addition of noise increases, If you try to reduce this, the number of dropouts will increase, making it difficult to detect high-quality sections.

For words that have been omitted or added during section detection during registration, the same omissions or additions will not normally occur during recognition, so most of them will be rejected, or
There was a problem that it was misrecognized.

In view of the above-mentioned drawbacks of the conventional art, the present invention provides that the parts that are likely to be omitted in the above-mentioned section detection can be known in advance based on devoicing rules, etc., so that words or single syllables can be detected in accordance with the devoicing rules, etc. It is an object of the present invention to provide a method of improving the quality of speech detected by section detection by changing the section detection method for each section.

[Means for solving problems]

In order to register a speech pattern in the word speech recognition device according to the present invention, a parameter extraction unit extracts recognition parameters representing the characteristics of the input speech pattern, and a specific threshold value is set among the recognition parameters from the parameter extraction unit. In order to detect the recognition parameters of the section where true speech exists, which is a larger part, a plurality of section detection units each having a voice power threshold and/or a zero-crossing threshold determined for detection on a word-by-word or monosyllable basis; A section detection storage section that stores the relationship between a word or a single syllable and a section detection section having a corresponding detection function, and one of the outputs of the plurality of section detection sections is registered as a pattern under the control of the section detection storage section. In addition, in order to recognize an unknown speech pattern, a parameter extraction section extracts recognition parameters representing the characteristics of the unknown input speech pattern, and a switching section selects recognition parameters from the parameter extraction section. A plurality of section detection sections each having a voice power threshold or a zero-crossing threshold for detecting the recognition parameter of a section where true speech exists, which is a portion larger than a specific threshold, and storing the output from each section detection section. a plurality of input pattern memories, a registered pattern group that stores registered word audio patterns, a selection unit that selects an output of the input pattern memory for each word or monosyllable to be matched read from the registered pattern group; This is achieved by comprising a matching unit that matches the registered pattern from the pattern group with the output from the selection unit and uses the registered word or monosyllable that provides the minimum distance between patterns as the recognition result.

[Effect]

That is, according to the present invention, when recognizing input speech having registered word speech patterns in advance,
In a word speech recognition method that calculates the inter-pattern distance between the speech pattern of the unknown input word and the registered word speech pattern, and uses the word of the registered word speech pattern that gives the minimum distance as a recognition result, (1) the registered word When registering audio patterns, select the section detection method.
By providing a storage unit that stores information for each word or each monosyllable, the method for detecting the section can be switched for each word or monosyllable to be registered based on the stored content, and (2) ) The output of the section detection unit for unknown input speech is selected for each word or monosyllable read from the registered pattern group to be matched, and is compared and recognized, so there is no error in section detection. This has the effect of improving the quality of standard speech patterns and improving the recognition performance for unknown input speech.

〔Example〕

Embodiments of the present invention will be described in detail below with reference to the drawings.
FIG. 1 is a diagram showing a configuration example for voice pattern registration according to the present invention, and FIG. 2 is a diagram showing a configuration example for recognizing unknown input voice according to the present invention. Section detection unit 2' in
and related mechanisms (switching section 3a, selection section 3b, etc.)
are the means necessary to carry out the present invention.

FIG. 3 is a diagram showing the concept of the registration method of the present invention, where a indicates the case of a word, b indicates the case of a monosyllable, and FIG. 4 shows the concept of the registration method of the present invention. It is a figure explaining an error rate. Moreover, FIG. 5 is a diagram showing the concept of the recognition method according to the present invention,
a shows the section detection method, b shows the matching method with registered speech patterns, and FIG. 6 is a diagram for explaining the recognition rate when the recognition method according to the present invention is adopted.

Below, while referring to Figures 1 and 2, Figures 3~
The word speech recognition method of the present invention will be explained using FIG.

First, the standard voice pattern registration method will be explained.

For example, the word to be recognized is "aomori",
If the word "aichi" is included,
When registering the word speech pattern, the voice power threshold used for detecting the section of the word voice for "aomori" is raised by 3 decibels than usual to make it less likely that noise will be added. {Figure 3 a
Refer to "Section 1 of A"} Also, lower the voice power threshold used to detect the section of the word voice for "aichi" by 3 decibels than normal only for the end to make it less likely that dropouts will occur. . Alternatively, section detection for the end,
Using both the audio power threshold and the threshold for zero crossings {see “section B” in Figure 3a},
Section detection is performed using the pattern of section 2 (A+B).

This is because the ending of the word "aichi" is unvoiced in most cases, so the speech power at the end of the word is low, but the high-frequency components are relatively strong, and the zero crossing This means that section detection is carried out by focusing on the characteristic that the value increases.

Figure 3b shows the temporal changes in voice power and zero cross of the monosyllables ``su'' and ``a'' on the same axis. It can be seen that the vocal power and zero-crossing characteristics differ for each single syllable.

FIG. 4 is a diagram explaining the error rate when registering a voice pattern using the procedure of the present invention.
Many of the devoiced word endings are dropped, and although many of the devoiced word endings are detected correctly when section detection is performed using the voice power threshold and zero crossing, the error rate increases due to the addition of noise, and individual It is shown that the error rate is minimized when the interval detection method is changed.

As described above, the configuration for registering a speech pattern in the word speech recognition device according to the present invention focuses on the fact that each word or single syllable has different speech power and zero-crossing characteristics. In order to perform the most suitable interval detection for extracting syllables, a interval detection method is stored in the interval detection storage unit 3 for each word or single syllable, and the interval detection unit 3 stores the interval detection method for each word or monosyllable. (1~n)
2' is selected by the switching section 3a. (See FIG. 1) Next, the configuration for recognizing unknown input speech in the word speech recognition device of the present invention will be described.

For example, the word to be recognized is "aomori",
If the word "aichi" is included,
The threshold value of the voice power used to detect the section of the word voice for "aomori" is raised by 3 decibels compared to normal to make it difficult for noise to be added. {Refer to “Section 1 of A” in Figure 5a} In addition, the voice power threshold used to detect the section of the word voice for “aichi” was lowered by 3 decibels than normal only for the end to prevent omissions. Make it difficult to wake up. Alternatively, the section detection for the end can be performed using the audio power threshold and the threshold for zero crossing {fifth
Section detection is performed using the pattern of section 2 (A+B) using both the "Section B" in Figure a).

This is because the ending of the word "aichi" is unvoiced in most cases, so the speech power at the end of the word is low, but the high-frequency components are relatively strong, and the zero crossing This means that section detection is carried out by focusing on the characteristic that the value increases.

In this way, the interval detection method for each word or single syllable is determined based on the devoicing rule known in advance (see Figure 2, interval detection unit (1 to n) 2')), and When unknown speech is input, each section detection section (1 to n) 2' performs section detection, and the recognition of the detected speech section is stored in the input pattern memory (1 to n) 2''. Parameters are stored.

On the other hand, in the control unit 7, the standard speech pattern registered in the registered pattern group 6 is read out for each word, and the corresponding input pattern memory (1 to n) 2'' is read out.
is selected by the selection unit 3b, the inter-pattern distance from the standard speech pattern is determined in the collation unit 4, and the standard speech pattern with the smallest inter-pattern distance is output as the recognition result. (See FIG. 2) FIG. 5b schematically shows the recognition process at this time, and shows the process when recognizing the word sound "aichi".

In the section detection shown in , the "chi" part at the end of the word is dropped, and in the section detection shown in , noise is added to the beginning of the word, and the "chi" part at the end is dropped. There is.

Therefore, when calculating the inter-pattern distance between the registered pattern groups A, B, and C read out from the registered pattern group 6, as shown in the figure, the inter-pattern distance is the minimum in the registered pattern group B, and the registered pattern group By selecting the word with the minimum distance between patterns from among the words registered in B, the correct word pronunciation "aichi" can be recognized.

FIG. 6 explains the recognition rate when recognizing unknown input speech according to the present invention.
It can be seen that the error rate is similar to that of the registration shown in the figure, and that the recognition rate is maximum when the section detection method is changed individually for each word (+).

As is clear from Figure 5b, this + interval detection includes many errors,
In most of the matching results, the distance between patterns is large, so that the number of erroneous recognitions is reduced.

In the case of registering the standard speech patterns mentioned above, or in the case of recognizing unknown input speech, in either case, the selection conditions for the interval detection method must be manually set for each word or single syllable in advance. Needless to say, it may be automatically generated and set using the above-mentioned devoicing rules or the like.

For example, in the above example, we know from the above devoicing rules that the ending of the word "aichi" is devoiced, so we perform the above interval detection for the final part of the word. It is defined as follows.

Even if the interval detection method of the present invention is used, the problem of noise addition remains, but as in the case of "aichi" above, words with devoiced endings usually account for 1 to 20% of all words. Therefore, for the other 80 to 90% of words, the problem of noise addition is solved by detecting sections where noise addition is unlikely to occur, as in the case of "aomori". becomes significantly smaller.

In addition, if the words to be recognized have been determined from the beginning and there are no changes or only a few changes, the section detection method can be set manually in advance, but if there are many changes, the above-mentioned method can be used. It is best to automatically generate it using devoicing rules, etc. It goes without saying that if a single syllable is used as the standard speech registration pattern, it may be set in advance.

〔Effect of the invention〕

As described above in detail, the word speech recognition device of the present invention has registered word speech patterns in advance,
When recognizing input speech, the distance between the speech pattern of the unknown input word and the registered word speech pattern is determined, and the word of the registered word speech pattern that gives the minimum distance is used as the recognition result. In this method, (1) when registering the registered word sound pattern, by providing a storage unit that stores a method for detecting an interval for each word or for each monosyllable, the interval is detected based on the stored content; (2) The output of the section detection unit for unknown input speech is switched for each word or single syllable to be registered. Since each syllable is selected and compared and recognized,
This has the effect of reducing section detection errors, improving the quality of standard speech patterns, and improving recognition performance for unknown input speech.

[Brief explanation of drawings]

FIG. 1 is a diagram showing an example of the configuration for voice pattern registration in a word voice recognition device according to the present invention;
FIG. 2 is a diagram showing an example of the configuration for recognition of unknown input speech in the word speech recognition device according to the present invention, FIG. 3 is a diagram showing the concept of the registration method of the present invention, and FIG.
The figure is a diagram explaining the error rate when using the registration method of the present invention, Figure 5 is a diagram showing the concept of the recognition method according to the present invention, and Figure 6 is a diagram showing recognition when using the recognition method according to the present invention. FIG. 7 is a diagram for explaining the conventional standard voice pattern registration and a method for recognizing unknown input voice. In the drawing, 1 is a parameter extraction section, 2 is a section detection section, 2' is section detection sections 1 to n, 2'' is an input pattern memory, 3 is a section detection storage section, 3a is a switching section, 3b is a selection section, 5 is a collation unit, 6 is a registered pattern group, 8 is a control unit, A is section 1, A+B is section 2, and ~ is a section detection method, respectively.

Claims (1)

[Scope of Claims] 1. A parameter extraction unit that extracts recognition parameters representing the characteristics of an input speech pattern, and recognition of a section in which true speech exists, which is a portion of the recognition parameters from the parameter extraction unit that is greater than a specific threshold. a plurality of interval detection units in which voice power thresholds and zero-crossing thresholds are determined for each word or each single syllable in order to detect the parameters;
A section detection storage section that stores the relationship between each word or single syllable and a section detection section having a corresponding detection function; and one of the outputs of the plurality of section detection sections is registered under the control of the section detection storage section. A word speech recognition device comprising: a switching section that selects a pattern; and a switching section that selects a pattern. 2. A parameter extraction unit that extracts recognition parameters representing the characteristics of an unknown input speech pattern, and a parameter extraction unit that detects recognition parameters of sections where true speech exists, which is a portion larger than a specific threshold among the recognition parameters from the parameter extraction unit. a plurality of section detection sections having predetermined speech power thresholds and zero-crossing thresholds; a plurality of input pattern memories that store the output from each section detection section; a registered pattern group that stores registered word speech patterns; A selection section selects the output of the input pattern memory for each word or monosyllable to be matched read out from the pattern group, and the registered pattern from the registered pattern group is compared with the output from the selection section, and the minimum distance between patterns is determined. 1. A word speech recognition device comprising: a matching section whose recognition result is a registered word or monosyllable that gives a .