JPS59195299A - Sepecific speaker's 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] (a) Technical Field of the Invention The present invention registers input speech of a specific speaker in a speech dictionary, compares the input speech to be recognized with the registered speech dictionary, and selects the most similar input speech. The present invention relates to a specific speaker speech recognition device.

(b) Background of the technology In recent years, speech recognition devices have been used to convert the input speech of a limited number of registered speakers into appropriate input speech parameters and register them in a registered speech dictionary. , a speaker speech recognition device that identifies the input speech of a specific speaker by comparing it with a registered speech dictionary, or converts the input speech of a specific speaker's words into appropriate input speech parameters to generate registered speech. With the advancement of voice recognition technology, there are advances in voice recognition technology, such as specific speaker voice recognition devices that register the input voice uttered by a specific speaker who wants to be recognized with a registered voice dictionary and display the recognition results as characters. It has expanded into the field of dialogue.

(C1 Prior art and problems) In the conventional specific speaker speech recognition method, the word speech is registered in the registered speech dictionary with just one utterance by the specific speaker, and the word is compared with the input speech to be recognized and the registered speech dictionary. The structure of this type of specific speaker speech recognition system will be explained. Fig. 1 shows a block diagram of the circuit configuration of a conventional specific speaker speech recognition device. First, the speech of a specific speaker is recognized. In order to register, a single word voice uttered by a specific speaker is input through the microphone 1, converted into input voice parameters of digital information by the voice feature extraction unit 2, and stored in the input voice variation buffer 3.Input The input voice variation stored in the voice parameter buffer 3 is registered in the registered voice parameter dictionary section 5 through the registration processing section 4 via the line i.
When a specific speaker utters a plurality of words, the plurality of word sounds uttered by the specific speaker are registered as a plurality of registered speech parameters. Next, in the speech recognition processing procedure, the word speech of the specific speaker to be recognized is input through the microphone I, and the word speech of the specific speaker to be recognized is inputted into the input speech variation make buffer 3 in the same manner as at the time of registration.
Stores input audio parameters. This input voice parameter and a plurality of registered voice parameters in the registered voice parameter dictionary section 5 are sequentially collated by a voice parameter collation section 6, and a collation ratio #ltg of the plurality of collation results is sent to a recognition result selection section 7. . When the comparison between this input voice parameter and a plurality of registered voice parameters is completed, the recognition result selection unit 7 selects the registered voice with the smallest matching distance g, and outputs the registered voice as the recognition result to the recognition result output terminal 8. do. The above is the conventional specific speaker speech recognition processing procedure. In this method, if 100 words are to be registered, but if a particular speaker's voice becomes hoarse due to their physical condition that day, the 100 words may not be completely registered, and some words may be registered incorrectly. As a result, recognition errors occur during recognition, resulting in a reduction in recognition rate.

(cll) OBJECT OF THE INVENTION The present invention aims to solve this conventional drawback that lowers the recognition rate.

tel Structure of the Invention Therefore, the present invention provides a specific speaker speech recognition device that compares input speech parameters of a specific speaker with a plurality of registered speech dictionary parameters registered in advance.
a plurality of input voice parameter buffers, an input voice parameter selection control unit that selects the plurality of input voice parameter buffers, and a registration processing control unit that processes the input voice parameters, when registering the voice of the specific speaker. , having the specific speaker utter one word voice three or more times, storing each input voice parameter uttered three or more times in the plurality of input voice parameter buffers; From the stored individual input voice parameters, the input voice variation selection control section selects two input voice parameters that are close to each other, and the registration processing control section performs a process of synthesizing and averaging each A. It is characterized by registering.

if) Embodiment of the Invention An embodiment of the present invention will be described below. FIG. 2 is a block diagram of the circuit configuration of the speaker-specific speech recognition apparatus according to the present invention, and the same objects are indicated by the same symbols as in FIG. 1. 9 is a registration recognition switching unit; 10 is an input audio parameter hasher group; 1
1 is an input audio variation selection control section, 12 is an audio parameter collation section, and 13 is a registration processing control section. The registration recognition switching unit 9 is equipped with a changeover switch S for registering input voice and recognizing input voice, and the contact a side of the changeover switch S is connected to the input voice parameter selection control unit 11, and the contact side is connected to the input voice parameter selection control unit 11, and the contact side is for registration. The contact C side C is connected to the audio parameter dictionary section 5 and the audio parameter matching section 12 .

When registering a voice, the switch is switched to the contact a side, and when recognizing a voice, the switch is switched to the contact side.

In this circuit configuration, when registering the voice of a specific speaker, the changeover switch S of the registration recognition switching unit 9 is switched to the a side and the same word voice is uttered three times, so that the input voice variation ~ Tahasofa group 1o has three input audio variable make buffers PBI, PB2 . Equipped with PB3. The first voice of the specific speaker is converted into input voice parameters P1 of digital information by the voice feature extraction unit 2, and the input voice parameters in the input voice parameter buffer group 10 ~ Tapasofa F
Stored in BI. Next, the second and third voices are sequentially inputted with voice parameters P2 in the same manner as above.
P3 and the respective input audio parameter buffers PB2. Stored in PH1. The input audio parameters Pi, P2 . P3 is input to the audio parameter matching unit 1 via the audio parameter selection control unit 11.
Enter 2. Under the control of the input audio parameter selection control unit 11, the input audio parameter P1 is input to the audio parameter matching unit 12 via the registration recognition switching unit 9 via the line d, and the input audio parameter P2 is input to the audio parameter matching unit 12 via the line e. Ru. Then, the audio parameter matching unit 12 checks both input audio parameters P1. The matching process P2 is performed, and the matching distance g12 of the obtained matching result is sent to the input audio parameter selection control section 11 via the line f. Next, in the same way, PI and P3 verification processing and P2. Perform P3 verification processing. In this way, the input audio parameter P1
The three matching distances g are sent to the input audio parameter selection control section 11, such as the matching path Mh13 of P3 and the matching path 11tg23 of the input audio variations P2 and P3. Input audio parameters PL, P2. P3 and matching distance g12. 'gl
3. The collation distance positional relationship of g23 is shown in FIG. These three matching distances g12°gl3.
g23, the input audio parameter selection control unit 11 selects the one with the closest matching distance to the input audio parameters Pi, P2.
．． Two input audio parameters are selected from P3, and the input audio parameters are registered from the input audio parameter buffer group 10 in which the selected two input audio parameters are stored.
The registration processing control unit 13 synthesizes and averages these two input voice parameters by z, and registers them in the registered voice parameter dictionary unit 5, so that one word voice of a specific speaker is , will be registered as a registered audio variation.

Repeat the above steps for as many words as you want to register,
It is registered as a plurality of registered audio parameters. Next, in the speech recognition processing procedure, the changeover switch S of the registration recognition switching section 9 is switched to the b side, the word speech of the specific speaker to be recognized is inputted from the microphone 1, and the input speech parameter buffer group is set as in the case of registration. The input audio parameters are stored in the input audio parameter buffer FBI No. 10 as input audio parameters P1. This Kugata voice parameter P1 is sent to the voice parameter matching unit 12 via the line e via the Kugata voice parameter selection control unit 11, and one of the multiple registered voice parameters registered in the registered voice parameter dictionary unit 5 is sent. The registered voice parameters are sent from the registered voice parameter dictionary section 5 via the line to the voice parameter matching section 12 via the registration recognition switching section 9.
I'll send it to you. The received input voice parameter P1 and the registered voice parameter are collated by the voice parameter collation unit 12, and the collation distance g of the collation result is sent to the recognition result selection unit 7. Subsequently, other registered voice parameters registered in the registered voice parameter dictionary section 5 are sequentially collated with Kugata voice parameters by the voice parameter collation section 12, and a plurality of collation distances g are checked by the recognition result selection section 7. send to When the comparison between this input voice parameter and a plurality of registered voice parameters is completed, the recognition result selection unit 7 selects the most matching path Ml.
A registered voice with a small tg is selected, and the registered voice as a recognition result is output to the recognition result output terminal 8. The above is the speech recognition processing procedure for a specific speaker according to the present invention. In this example, the explanation was given by repeating one word voice of a specific speaker three times.
In the case of repeating the process five times, it is sufficient to provide five input voice parameters. The recognition rate improves as the number of times increases, but an appropriate number of times can be determined by taking into consideration the handling time at the time of registration and the device cost of 1~. With this method, even if a particular speaker's voice becomes hoarse due to their physical condition that day, the voice will be uttered three or more times during voice registration, and the two input voice parameters of the better voice will be synthesized and averaged, so registration errors will occur. This eliminates this problem, and improves the effectiveness of learning that a specific speaker must practice in order to register his or her own voice, which has the advantage of improving the recognition rate during recognition.

(gl As described in detail of the invention, when a specific speech according to the present invention is uttered three or more times during voice registration, two Kugata voice parameters are synthesized two by two among the input voice parameters of 3 (fli + or more). By equipping the specific speaker speech recognition device with a method that performs averaging processing and registers it in the registered speech parameter dictionary, registration errors can be eliminated and the learning effect of specific speakers practicing to register their own speech can be improved. This has the effect of improving the recognition rate during recognition.

[Brief explanation of the drawing]

FIG. 1 is a circuit configuration block diagram of a conventional speaker specific speech recognition device, FIG. 2 is a circuit configuration block diagram of a specific speaker speech recognition device according to the present invention, and FIG. 3 is a matching distance between input speech parameters. It is a relationship diagram of a position. In the drawings, 1 is a microphone, 2 is a voice feature extraction section, 5 is a registered voice parameter dictionary section, 7 is a recognition result selection section, 9 is a registered recognition switching section, 10 is an input voice parameter group of 8, and 11 is an input voice parameter Reference numeral 12 indicates a selection control section, reference numeral 12 indicates an audio parameter matching section, and reference numeral 13 indicates a registration processing control section.

Claims (1)

[Claims] In a specific speaker speech recognition device that matches input speech parameters of a specific speaker with a plurality of pre-registered registered speech dictionary parameters, a plurality of input speech parameter buffers and the plurality of input speech parameter buffers are provided. an input voice parameter selection control unit that selects an input voice parameter, and a registration processing control unit that processes the input voice parameter, and when registering the voice of the specific speaker, causes the specific speaker to utter one word voice three or more times; Each input voice parameter uttered three or more times is stored in the plurality of input voice parameter buffers, and the individual input voice parameters stored in the plurality of input voice parameter buffers are stored in the input voice parameter selection control section. A specific speaker speech recognition device characterized in that two input speech parameters that are close to each other are selected, and the registration processing control section performs a process of combining and averaging them by % and registers the same.