JPS58169199A - Voice recognition system employing vowel information - 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 (1) Technical field of the invention The present invention relates to the voice gw& method, and in particular, it is used to find the difference between the vowel voice data of the user and the vowel voice data of the dictionary content speaker. A voice g-weaving method using vowel information that corrects the dictionary contents using the difference information and adapts it to the user's voice (2) Conventional technology and problems Conventional speech recognition devices do not calculate the g-fiber ratio. In order to improve this, dictionary files were pre-registered with the voices of specific speakers, tb users. However, such a specific speaker method has the disadvantage that if there are a large number of single candies to be recognized, the user must register them all, making it very time-consuming to create a dictionary.

Therefore, if voice recognition could be performed using a dictionary registered with someone else's voice, it would be very convenient since there would be no need to create a dictionary for each user.

However, currently, when performing speech recognition using a registered dictionary with someone else's voice, it is generally more difficult to perform voice recognition than when using a registered dictionary with one's own voice due to individual differences in voice. Only a very low recognition rate can be obtained.

Therefore, as a method to increase the recognition rate when using a dictionary registered with someone else's voice, an adaptive process is used that either brings the input voice closer to the voice of the person who is speaking the dictionary contents, or brings the dictionary contents closer to the user's voice. Is required.

However, when considering shortening the speech recognition ability formula, there is a problem with the method of modifying the input speech to make it closer to the speech of the dictionary content speaker.

That is, when input speech is modified and processed, there is a problem in that the recognition speed is slowed down by the amount of time required for the processing, and, for example, it becomes difficult to respond with a laugh time response.

On the other hand, in the method of modifying the contents of the dictionary to bring it closer to the user's voice, it is sufficient to modify the contents in advance before use, so there is no problem even if it takes time to modify the contents before use.

However, when the number of dictionary items is large, time compression is performed for each dictionary item due to the relationship between the 11Cd dictionary memory size and the total n+h during recognition, and as a result, detailed correction becomes difficult.

For example, as a method for adapting to individual differences, there is a method of finding the average spectrum of each of the dictionary entry speaker and the input speaker and using the difference spectrum to adapt, but if you try to do this for each sound, It would be fine if the time compression was done at about 4cli, but if not, it would be strange.

(3) Purpose of light The purpose of the present invention is to improve Qτ,
When adapting the voice fiffi +Jji^ to the input speaker and the user, the recognition rate is adjusted by adapting it to each vowel type, and the vowel volume at the time of compression is memorized. It is easy to write Cτ without causing a large increase in the number of new entries even for a dictionary that has already been compressed.
The purpose of this invention is to provide a speech recognition formula using vowel information that can be adapted.

(4) Structure of the Invention Ninthly, in order to achieve this object, the friend voice recognition method using the vowel information of the present invention uses a word dictionary, a vowel information storage means of the #word dictionary creator, and a speech recognition system based on the user's voice in advance. A word dictionary modification means is provided for modifying the single dictionary according to the vowel information created and the vowel information of the word dictionary creator, and the vowel information of the word dictionary is modified according to the user's voice. It is characterized by adapting the dictionary to the user's voice.

(5) Embodiment of the Invention An embodiment of the present invention will be explained based on FIGS. 1 to 3.

First, the present invention will be explained in detail with reference to FIGS. 1 and 2.

In the present invention, the vowel spectra of each of the dictionary creator and the input speaker are determined.For example, for the vowel ■, as shown in FIG. The voice spectrum of person B is determined, and the difference A-B between the two spectra is determined as shown in FIG. 1 (b).

To this end, as shown in FIG. 2, in order to compress information, the channel is divided into, for example, eight channels in the frequency direction, and the difference A-B is determined for each section. In this way, by finding the difference between the two for each vowel and correcting the word dictionary based on this difference, a word dictionary adapted to the input speaker AK can be obtained.

An embodiment of the present invention will be described with reference to FIG.

In FIG. 3, lti microphone, 2 is a vowel spectrum collection unit, 3 is a vowel difference spectrum calculation unit, 4 is a word dictionary, 5 is a dictionary speaker vowel spectrum holding unit, 6 is a word dictionary correction unit, and 7#i corrected word Dictionary, 8td Microphone, 9 is a recognition unit.

The vowel spectrum collection unit 2 collects the vowel spectrum of the input speaker to create the curved person shown in FIG. It is also possible to extract from a standard single peak, but in this example, the vowel itself is uttered.

The vowel difference spectrum calculation unit 3 calculates the difference for each channel segment of each vowel between the input speaker A and the dictionary content speaker B, as shown in FIG. The vowel spectrum of speaker A is transmitted from the vowel spectrum collection section 2, and the vowel spectrum of speaker B in the dictionary is transmitted from the dictionary speaker vowel spectrum holding section 5.

Word Dictionary 4#'i Dictionary Contents This dictionary was created by speaker B, and contains the essential characteristics of each word in a file. Each vowel spectrum of speaker B in the dictionary is held in dictionary speaker vowel spectrum holding section 5.

This vowel spectrum, like that of the input speaker, may be obtained by uttering the vowel itself or from a step word, but in this example, the one obtained by uttering the vowel itself is retained. There is.

The word dictionary correction unit 6 corrects the vowel information of each word stored in the word dictionary 4 based on the difference for each channel category of each vowel as shown in FIG. 2, which is obtained by the spinning vowel difference spectrum calculation s3. The words corrected in this way are stored in the corrected word dictionary 7.

Microphone 8Fi input This is the voice input section when the speaker performs voice recognition, and the unknown input input from this microphone 8 is giii! Matching with the corrected word dictionary 7 is performed by the unit 9.

Next, the operation shown in FIG. 3 will be briefly explained.

(2) When creating the word dictionary 4, the dictionary contents Vowel information of speaker B is held in the dictionary speaker vowel spectrum holding section 5.

■ When voice recognition is performed using this word dictionary 4, first, the input speaker who is the user inputs each vowel through the microphone 1. The input spectrum of each vowel is held in the vowel spectrum collection unit 2, and based on this and each vowel spectrum of the dictionary content speaker B held in the dictionary speaker vowel spectrum holding unit 5, the vowel difference spectrum calculation unit 3
calculates the difference in channel classification for each vowel as shown in FIG. Based on this difference, the word dictionary correction section 6 reads and corrects the word dictionary 4 to create a corrected word dictionary 7.

■ After the corrected word dictionary 7 is created in this way, if an unknown input is input from the input speaker Ad and the microphone 8, the recognition unit 9 matches this unknown input with the corrected word dictionary 7 and performs the matching. The results will be output as recognition results.

Even in the case of a time-compressed word dictionary, the time ratio of each vowel in each time segment before compression is memorized.
If these are used as weights and the weighted sum of the spectra of the differences between vowels is used as the corrected sound, speaker adaptation can be easily performed. An embodiment in this case will be described with reference to FIGS. 4 and 5.

For example, as shown in FIG. 4, it is assumed that the word IAKAII is divided into eight parts in the time direction and represented by the sum of spectra weighted by power for each interval. At this time, there is no problem since the first section, second section, etc. are only IAI sounds, but since the seventh section contains IAI and III sounds, it cannot be simply corrected. Therefore, it is sufficient to calculate the sum of powers separately for each of the five vowels included in each section, use them as weights to multiply the difference spectrum of the five vowels, and add the results to correct the division using the obtained spectrum.

That is, in the case of Figure 4, as shown in Figure 5, the weight of each category and vowel (expressed as positive, medium, and small for convenience)
Create a table showing specific numerical values (in which specific numerical values are entered) and store it in a word dictionary, for example,
When modifying the classification, reduce the weight of IAI and
It is sufficient to set the weight as medium and correct it using the spectral difference obtained in FIG.

If information on the power sum of each vowel in each time interval of a word in the dictionary is stored in this way, correction can be easily made in such a case.

(6) Effects of the Invention According to the present invention, a word dictionary can be easily adapted to the input speaker so that vowel information, which has a large weight in speech recognition, is adapted to the input speaker. Using the word dictionary you created, Ili! ! It is possible to perform speech recognition with high recognition rate.

[Brief explanation of the drawing]

1 to 3 are configuration diagrams of one embodiment of the present invention, and FIGS. 4 and 5 are explanatory diagrams of other embodiments of the present invention. In the figure, l is a microphone, 2 is a vowel spectrum collection unit, 3 is a vowel difference spectrum calculation unit, 4#i word dictionary, Sa
A dictionary speaker vowel spectrum holding section, a 6Fi word dictionary correction section, a 7Fi modified single candy dictionary, 8 a 1-iclophon, and 9 a recognition section. Patent Applicant: Fujitsu Ltd. Representative Patent Attorney Akie Yamatani? 2 meals: 'j'4 diagrams S round

Claims (2)

[Claims] (1) A unit for correcting the single-modal dictionary using the Kusame dictionary, the vowel information holding means of the core word dictionary creator, the vowel information created in advance from the voice of the usage number, and the vowel information of the single-modal dictionary creator. Ameji i# - A voice using vowel information, which is equipped with a correction means and adapts the word dictionary to the user's voice by correcting the vowel information of the unimodal dictionary according to the user's voice. Recognition method. (2) Each vowel data included in each time-compressed section is stored for each word, and based on this, the vowel information is corrected by 1. Vocal g weave method using the vowel information listed.