JPH01138595A - Word voice recognition equipment - Google Patents

Abstract

PURPOSE: To reduce the memory capacity and to prevent erroneous recognition by storing feature parameters of word voices to be recognized and those of noise effective for noise rejection in first and second dictionaries respectively and recognizing a word voice by matching of these feature parameters. CONSTITUTION: Feature parameters of the word voice to be recognized are stored in a first dictionary 1, and a prescribed number of feature parameters of noise effective for noise rejection in circumstances of word voice recognition are stored in a second dictionary 2 including new feature parameters of noise which are extracted by a feature extraction part 3 just before word voice recognition. Feature parameters of a word voice are extracted by the feature extraction part 3 and are matched with feature parameters stored in first and second dictionaries 1 and 2 by a matching means 4 to recognize the word voice. Thus, the memory capacity is reduced, and erroneous recognition of the word voice is prevented.

Description

[Detailed Description of the Invention] [Summary] Regarding a word speech recognition device, the first dictionary aims to provide a word speech recognition device that can reduce the number of types of noise characteristic parameters to be stored in order to reject noise. The feature parameters of the word speech to be recognized are stored in the first mode, and in the environment where word speech recognition is performed, new noise features extracted by the feature extraction unit immediately before word speech recognition are performed. A predetermined number of noise feature parameters, including parameters, that are effective for noise rejection in this environment are stored in a second dictionary, and in a second mode, feature parameters of word sounds are extracted by the feature extraction unit; The matching means performs matching with feature parameters stored in the first dictionary and the second dictionary to perform word speech recognition.

[Industrial application field]

The present invention relates to improvements in word speech recognition devices.

A word speech recognition device performs matching (comparison) between a feature parameter indicating the characteristics of an input word speech and a plurality of feature parameters stored in a dictionary, calculates the distance from each feature parameter, Word speech recognition is performed by selecting the word with the smallest distance (those with similar l).

In order to perform word speech recognition, just before the speaker utters the word speech to the word speech recognition device, various noises such as the sound of dropping an object, the sound of a telephone ringing twice, the sound of a car crashing, etc., may be input. be.

At this time, the word speech recognition device extracts the feature parameters of the input sound, matches them with a plurality of feature parameters stored in the dictionary, and rejects the extracted sound if the distance exceeds a predetermined value.

However, since word speech recognition devices usually take into account speech fluctuations (word ambiguity),
The distance limit has a wide range.

Therefore, if only the characteristic parameters of the voice of the word to be recognized are stored in the dictionary, a malfunction may occur in which one word is selected from the dictionary despite the noise.

It is necessary to solve this problem, but in order to solve this problem, it is desirable that the memory capacity is small and the possibility of word speech being mistakenly recognized as noise is small.

[Conventional technology]

A conventional example will be explained below using figures.

FIG. 3 is a block diagram of a conventional word speech recognition device.

In the figure, 3 is a feature extraction unit that extracts feature parameters, 5 is a processor, and 6° is an RO that stores a processing program.
M, 7 is a parameter memory that temporarily stores the feature parameters extracted by the feature extractor 3; 10 is a dictionary memory that stores the feature parameters of speech of the word to be recognized and the feature parameters of noise; 8 is a word as a recognition result. This is an I10 unit that outputs numbers etc.

Conventionally, no matter what environment word speech recognition is performed in, when noise is input, in order to be able to clearly identify it as noise and reject it, for example, the sound of dropping an object, the sound of a telephone ringing, etc. Characteristic parameters of about 20 to 30 types of noise, such as applause and car noises, are stored in the dictionary memory 10 together with characteristic parameters of speech of the word to be recognized.

The sound input through the microphone is input to the feature extraction unit 3 shown in FIG. 3, and feature parameters are extracted at intervals of 10 m5.

Therefore, the processor 5 reads the program stored in the ROM 6' and performs the following operations.

The feature parameters are read in from the feature extraction unit 3 at Ioms intervals, stored in the parameter memory 7, a section recognized as a word is cut out, and the feature parameters of this section are matched with all the feature parameters stored in the dictionary memory 10. , calculate the distance, and if the smallest distance is the feature parameter of noise, it is rejected, and if the feature parameter of the word sound to be recognized, it is assumed that the word sound has been recognized.
10 unit 8 outputs the word number etc. of this word sound.

[Problem that the invention seeks to solve]

However, in conventional word speech recognition devices, no matter what environment the word speech recognition is performed in, when noise is input, in order to be able to clearly identify it as noise and reject it, the memorized noise There are many types of feature parameters, which increases the memory capacity, and because there are many types of noise feature parameters, there is a high possibility that word speech will be mistakenly recognized as noise when input.

SUMMARY OF THE INVENTION An object of the present invention is to provide a word speech recognition device that can reduce the number of noise characteristic parameters to be stored in order to reject noise.

[Means for solving problems]

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

As shown in FIG. 1, the first dictionary 1 stores characteristic parameters of word speech to be recognized, and in the first mode, in an environment where word speech recognition is performed, , a predetermined number of noise feature parameters effective for noise rejection in this environment, including the new noise feature parameters extracted by the feature extraction unit 3, are stored in the second dictionary 2.

Then, in the second mode, the feature parameters of the word sounds are extracted by the feature extractor 3, and
The matching means 4 performs matching with the feature parameters stored in the second dictionary 2 to perform word speech recognition.

[For production]

The type of noise when performing word speech recognition is limited depending on the location.

Among the limited noises, continuous noise is determined by the location where word speech recognition is performed and the time period (2 to 3 minutes) immediately before word speech recognition is performed.

, Non-recurring noises such as the ringing of a telephone, the sound of a chair bowing, the sound of dropping something, etc. can be identified into a small number of types depending on the environment.

Therefore, in the present invention, a few identified non-recurring noises are first stored in the second dictionary 2 as noise characteristic parameters.

Next, in mode 1, at the location where word speech recognition is performed, the number of types of characteristic parameters of noise occurring in the time period immediately before word speech recognition is set to a predetermined number, including non-recurring noise. The continuous noise is updated to a new one and stored in the second dictionary 2 as follows.

Also, the characteristic parameters of the speech of the word to be recognized are of course stored in the first dictionary 1.
Remember it.

Then, word speech recognition is performed in the second mode, but the noise input at this time is stored in the second dictionary 2, although there are fewer types than in the past, but in this environment It is rejected with effective noise characteristic parameters suitable for.

That is, since the number of types of noise characteristic parameters to be stored is smaller than in the past, the memory capacity can be reduced, and the possibility of erroneously recognizing word speech is extremely small.

〔Example〕

An embodiment of the present invention will be described below with reference to the accompanying drawings.

FIG. 2 is a block diagram of a word speech recognition device according to an embodiment of the present invention.

The difference between FIG. 2 and FIG. 3 is that the contents of the program stored in the ROM 6 have been changed, and the characteristic parameters of the speech of the word to be recognized are stored in the dictionary memory 1. The feature parameters are stored in the dictionary memory 2, and the keyboard 9 is provided to switch between modes 1 and 2.

In the case of the present invention, first, the characteristic parameters of the speech of the word to be recognized are stored in the dictionary memory 1, as in the conventional case.

Next, for example, the sound of a telephone ringing or the sound of a chair bowing.

Among non-recurring noises such as the sound of dropping an object, noises that are likely to occur at the word speech recognition location are selected and their characteristic parameters are stored in the dictionary memory 2.

Therefore, the processor 5 reads the program stored in the ROM 6 and performs the following operations.

Immediately before word speech recognition at the place where word speech recognition is performed (2
Mode 1 is designated on the keyboard 9 in order to store the characteristic parameters of the noise (mainly continuous noise) for 3 minutes) into the dictionary memory 2.

Then, the feature extraction unit 3 extracts the feature parameters lQms.
It is read at intervals, stored in the parameter memory 7, segments that appear to be words are cut out, and the characteristic parameters of these segments are
Matching is performed with all the feature parameters stored in the dictionary memory 2 to find the distance, and the one with the minimum distance is
When it is larger than a predetermined value, it is stored in the dictionary memory 2 until the predetermined number N (approximately 5 to 6) is reached, and when it exceeds N, the oldest feature parameter is deleted to keep it within N.

If the value of the minimum distance is smaller than the predetermined value, do nothing and return to step (3).

Next, mode 2 is designated using the keyboard 9, and word audio is input.

Then, ■The feature extraction unit 3 extracts the feature parameters for 10ms.
It is read at intervals, stored in the parameter memory 7, segments that appear to be words are cut out, and the characteristic parameters of these segments are
The distance is determined by matching with all the feature parameters stored in the dictionary memory 1.2.

When the distance to all feature parameters is greater than a predetermined value, the input sound is not a recognized word sound but a new noise, so it is stored in the dictionary memory 2, and if the number is N or more, it is stored as the oldest noise. Delete noise feature parameters.

If the one with the smallest distance is the one stored in the dictionary memory 2, the process returns to that step.

Output word numbers, etc.

In this case, if there is no non-recurring noise, its characteristic parameters are of course not stored in the dictionary memory 2.

In this way, when rejecting noise, a small number of noise characteristic parameters (approximately 5 to 6 types) that are suitable for the environment are stored in the dictionary memory 2, so the memory capacity is reduced. , again, the word sound is mistaken for noise! There is almost no possibility of knowing.

〔Effect of the invention〕

As explained in detail above, according to the present invention, a small number of noise characteristic parameters (approximately 5 to 6 types) that are suitable and effective for the environment are stored in the memory to reject noise, so the memory capacity is reduced. This has the effect of almost eliminating the possibility of erroneously recognizing word speech as noise.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of the principle of the present invention, FIG. 2 is a block diagram of a word speech recognition device according to an embodiment of the present invention, and FIG. 3 is a block diagram of a conventional word speech recognition device. In the figure, 1 is the first dictionary, dictionary memory, 2 is the second dictionary, dictionary memory, 3 is the feature extraction section, 4 is the matching means, 5 is the processor, 6.6" is the ROM? 7 is the parameter Memory, 8 indicates I10 unit, 9 indicates keyboard, and 10 indicates memory for dictionary.゛Light Robe 092C 2nd (Rotary) Central A Row 6) Industry 1 Nibi Island] Passenger e Chia 4 buildings have No. 1 F D, 72 daytime 3 mouths

Claims (1)

[Claims] The first dictionary (1) stores feature parameters of word speech to be recognized, and in the first mode, immediately before word speech recognition is performed in an environment where word speech recognition is performed. The noise feature parameters effective for noise rejection in this environment, including the new noise feature parameters extracted by the feature extraction unit (3), are extracted by a predetermined number of second
In the second mode, the feature parameters of the word sounds are extracted by the feature extraction unit (3), and the feature parameters of the word sounds are stored in the first dictionary (1) and the second dictionary (2). ) is matched with the feature parameters stored in the matching means (4).
), and performs word speech recognition.