JPH079598B2 - Method for correcting standard parameters in voice recognition device - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device, and more specifically to a method of correcting standard parameters in a device of a method of separately recognizing consonants and vowels.

In voice recognition, a predetermined voice is input in advance to extract a characteristic parameter, this is registered as a standard parameter for a plurality of voices, and when an unknown voice input is input, the characteristic parameter is compared with the standard parameter. This is a technique for searching for the closest standard parameter and specifying that a voice corresponding to this is input.

If there are many false recognitions, it is necessary to correct the registered standard parameters as if they were inappropriate.

As a method of such correction or re-registration, when voice input is not performed and the voice is not correctly recognized, the voice is input again, and when the voice is not correctly recognized by this, the voice parameter at the time of the second input is used as a standard. There is a method of registering as a parameter.

On the other hand, recently, a method for recognizing a larger number of vocabulary words, or a method of performing recognition in a syllable unit or a phoneme unit in order to eliminate the limitation of the vocabulary is being developed.

[Problems to be solved by the invention]

In the case of recognition in syllable units, it is particularly difficult to recognize consonant parts, and the parameters obtained each time the utterance varies, so it is not possible to absorb the variations in each utterance by modifying the standard parameters as described above. . As one of the methods for solving this, Japanese Patent Laid-Open No. 58-31398.
There is a method of Japanese Patent No.
Inappropriate parameters obtained by incorrect segmentation (separation of consonants and vowels) could adversely affect recognition.

[Means for solving problems]

The present invention has been made to solve such a problem, and when both the consonant and the vowel of the voice input for correction are erroneously determined, it is determined that the voice input is inappropriate. It is an object of the present invention to provide a standard parameter correction method that avoids the influence of noise and the like without performing correction.

A standard parameter correction method in a voice recognition apparatus according to the present invention is a consonant standard parameter memory and a vowel standard parameter memory that store consonants of an input voice and standard parameters for vowel determination, respectively, and extracted from input voice. The consonant feature parameter and the parameter stored in the consonant standard parameter memory are compared to perform consonant identification, and the vowel feature parameter extracted from the input voice and the vowel standard parameter memory are stored. A vowel discrimination unit that compares vowels with certain parameters to determine the vowels, and a standard that corrects the contents of the consonant standard parameter memory or the vowel standard parameter memory using the consonant feature parameters or vowel feature parameters extracted from the input speech. A parameter correction unit, which inputs the syllable to input the standard parameter When correcting the contents of the consonant standard parameter memory or the vowel standard parameter memory by the positive part, the determination result in the consonant identification unit is different from the consonant of the input syllable, and the determination result in the vowel identification unit is the vowel of the input voice. In the case of a difference, it is characterized in that the correction by the standard parameter correction unit is prohibited.

[Operation] When noise is input at the same time during voice input during correction,
Alternatively, if the segmentation is poor, the consonant identification unit and the vowel identification unit cannot determine the child of the input syllable and the vowel. It goes without saying that it is inconvenient to register the input syllable in such a case as a standard parameter, and the present invention avoids it.

〔Example〕

Hereinafter, the present invention will be described in detail with reference to the drawings showing an embodiment thereof.

FIG. 1 is an overall block diagram of a voice recognition device according to the present invention.

The syllable input from the microphone 1 is subjected to processing such as high-frequency emphasis by the preprocessing unit 2 and then input to the parameter extraction unit 3 where the characteristic parameters are extracted. As the characteristic parameter, the frequency spectrum obtained by FFT, L
PC cepstrum or power information, number of zero crossings, and autocorrelation coefficient are used.

The special feature parameter is input to the phonological unit determination unit 4, where the phonological property is determined. The segmentation unit 7 determines the consonant section and the vowel section using the determination result, the spectrum change, and the duration. The result is input to the consonant parameter creating unit 5 and the vowel parameter creating unit 6. The characteristic parameters are also input to the consonant parameter creating unit 5 and the vowel parameter creating unit 6, where the consonant parameter and the vowel parameter are created according to the determination result of the segmentation unit 7. When this device is operating in the standard parameter registration mode, the above consonant parameters and vowel parameters are input to the consonant standard parameter memory 11 and the vowel standard parameter memory 12, respectively. Will be registered in. The consonant standard parameters and vowel standard parameters for such syllables are registered in advance for a plurality of required syllables.

When performing voice recognition, the operation unit 15 is switched from the registration mode to the recognition mode, and the parameters created by the consonant parameter creating unit 5 and the vowel parameter creating unit 6 are used in the same manner as before to recognize unrecognized consonant parameters and vowels. The parameters are input to the consonant identification section 8 and the vowel identification section 9, respectively. The consonant identification unit 8 and the vowel identification unit 9 successively read out a plurality of consonant standard parameters and vowel standard parameters respectively stored in the consonant standard parameter memory 11 and the vowel standard parameter memory 12 in advance, and read these unrecognized consonant parameters. And the vowel parameters, respectively, to determine the most similar consonant standard parameter and vowel standard parameter, and give the result to the output unit 13 as if the corresponding consonant and vowel were input. The output unit synthesizes a consonant and a vowel as a syllable and outputs this to an appropriate external device 14.

Next, the configuration related to the correction mode will be described with reference to the flowchart shown in FIG. When the correction mode is instructed by the operation unit 15, the standard parameter correction unit 10 synthesizes the consonant standard parameter memory 11 and the vowel standard parameter memory 12 with the syllable determined by the combination of registered contents (for example, KA) and the display unit
Call out to 16. When the operator inputs this syllable from the microphone 1, consonant parameters and vowel parameters are created as before, and are input to the consonant identification section 8 and the vowel identification section 9, respectively.

The consonant identification unit 8 and the vowel identification unit 9 sequentially store the contents of the consonant standard parameter memory 11 and the vowel standard parameter memory 12 in the same manner as in the recognition mode.
Also, the consonant and vowel of the input syllable are determined by comparing each parameter input from the vowel parameter creating unit 6.

The standard parameter correction unit 10 reads the result of this determination. If the consonant determination result is not K and the vowel determination result is not A, the standard parameters are corrected, that is, the contents registered in the consonant standard parameter memory 11 for K are corrected, and the vowel standard parameter memory 12 for A is corrected. The contents registered in will not be modified.

On the other hand, if both consonants and vowels are correctly judged as K and A, the standard parameters of K and A are corrected. That is, the characteristic parameters obtained from the newly input syllable are written in the consonant standard parameter memory 11 and the vowel standard parameter memory 12 as standard parameters. When only one of the consonant and the vowel is correctly judged, only the standard parameter for the correct judgment may be corrected as shown in the flowchart of FIG. 2, or a message requesting re-input is displayed on the display unit 16. You may let me.

〔effect〕

According to the present invention as described above, the parameter at that time is registered as the standard parameter when the noise is high when the standard parameter is to be corrected, or when there is a problem in the segmentation result due to the vocalization condition. There is no risk of this, and the recognition accuracy is improved.

[Brief description of drawings]

FIG. 1 is a block diagram of a voice recognition device according to the present invention, and FIG.
The figure is a flow chart showing the contents of the method of the present invention. 5 ... consonant parameter creation unit, 6 ... vowel parameter creation unit, 7 ... segmentation unit, 8 ... consonant identification unit, 9 ... vowel identification unit, 10 ... standard parameter correction unit,
11 …… Consonant standard parameter memory, 12 …… Vowel standard parameter memory

Claims (1)

[Claims] 1. A consonant standard parameter memory and a vowel standard parameter memory, which store parameters that are standards for determining a consonant and a vowel of an input voice, a consonant characteristic parameter extracted from an input voice, and the consonant standard. A consonant identification unit that compares a parameter stored in a parameter memory to determine a consonant, and a vowel characteristic parameter extracted from an input voice and a parameter stored in the vowel parameter memory are compared. A vowel discrimination unit for performing vowel determination, and a standard parameter correction unit for correcting the contents of the consonant standard parameter memory or the vowel standard parameter memory using the consonant feature parameter or vowel feature parameter extracted from the input voice, Input a syllable and use the standard parameter correction unit to store a consonant standard parameter memory or a vowel mark. When the content of the quasi-parameter memory is modified, the determination result in the consonant identification unit is different from the consonant of the input syllable, and the determination result in the vowel identification unit is different from the vowel of the input voice, the standard parameter A method for correcting a standard parameter in a voice recognition device, characterized in that the correction by a correction unit is prohibited.