JPS59160195A - 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] Technical field The present invention relates to a device for identifying audio input.

Traditionally, speech recognition devices analyze input speech to extract features, then compare them with the features corresponding to each word that have already been registered in the dictionary, and select the one with the greatest similarity or distance. The smallest one is output as the recognition result.

In that case, it is desirable to perform speech recognition at high speed, but it would take too much time to match all the words registered in the dictionary. For this purpose, the following methods have been proposed. (1) The dictionary is classified into groups based on vowels that are abundant in words, and the input speech is also analyzed for vowels and matched only in the corresponding dictionary groups. (2) Before inputting, the operator inputs the category of the input voice using a switch to limit the range of dictionaries to be compared. (3) Record the length of time at the beginning of each word registered in the dictionary, and do not match words whose length is significantly different from the input speech length.

However, method (1) requires vowel analysis of input speech, which requires extra time, and method (2) requires operator effort, and operational errors may lead to errors in the results. Furthermore, method (3) has the disadvantage that all dictionaries must be read regardless of whether or not verification is necessary.

OBJECTS The present invention aims to eliminate the above-mentioned drawbacks of the prior art and provide a speech recognition device that can recognize input speech at high speed.

Configuration The configuration of the present invention will be described below based on examples.

FIG. 1 is a block diagram for explaining an embodiment of the speech recognition device according to the present invention. In the figure, 1 is a microphone, 2 is a reset switch, and 3 is only the section of the voice from the signal input from the microphone 1. 4 is a register to record this voice, 5 is a register to record a number or symbol, 6 is the contents of register 4 and register 5
7 is a dictionary section, 8 is a feature extraction section, 9 is a matching section that matches the features recorded in the dictionary section 7 and the features from the feature extraction section 8, 10 is a similarity of matching results A determining unit 11 is an output unit outputting a determination result, and after registering some words in the dictionary unit in advance, the recognition process begins. First, all the contents of the register 4 are set to O using the reset switch 2, and then the voice to be recognized is inputted from the microphone 1. This voice is recorded in the register 4 after a voice section is cut out by the voice section detecting section 3. On the other hand, numbers are written in register 5, such as 1, 2, 3, etc. 1 and the contents of the first address of register 4, 2 and the second address, and 3 and the third address, and so on. The contents of register 4 and register 5 are correlated and multiplied. The results are f (1
), 2 f (2), 3 f (3), ・=n f(n
), (n+L)f(n+LL...Nf(N). However, the input audio is f(,t), N pieces of data can be recorded in the register, and the audio length is 1 to n. In other words, since f(t)=O when t>n, (n
+1) 0 is lined up in f(n+1) to Nf(N).

This is done in the opposite direction, from N f (N) to (N-1) f(
N-1), . . . , the first value that is not O appears is nf(n), which allows the voice length n to be easily determined. By arranging words in the dictionary in ascending order of shortest length and comparing words starting from the words with the phonetic length n, the number of calculations required to obtain the correct answer will be reduced and faster processing can be expected. However, in the illustrated embodiment, the input audio is configured to enter the register 4 and then reach the feature extractor 8, but conversely, the input signal is first entered into the feature extractor 8 and then stored in the register 4. It is more common to do so. Also, although multiplier 6 is used here, it may also be a divider,
In that case, just divide the contents of register 4 by the contents of register 5. Furthermore, in the above description, register 5
The numbers inside are the word file numbers or file names in the dictionary, but they may also be addresses in the dictionary.

FIG. 2 is a configuration diagram for explaining another embodiment of the present invention. In the figure, parts that have the same effect as the embodiment shown in FIG. Reference numbers are included.

This embodiment differs from the embodiment shown in FIG. 1 in that a natural number generator 12 is provided in place of the register 5. In the embodiment shown in FIG. 2, after the dictionary is created, the contents of the register 4 are first set to 0 by the reset switch 2, and then the voice is emitted into the microphone 1. Then, the voice section detecting section 3 cuts out only the voice signal part, converts it into a feature quantity at the feature extracting section 8, and temporarily stores it in the register 4. On the other hand, a fairly large natural number is generated in the natural number generator 12, first, the product is multiplied with the contents of the last part of the register 4, and if the result is 0, -1 is added or 1 is subtracted from the natural number, Next, register 4
Multiply by the contents of the second register from the end. If the result is 0, -1 is added to the above natural number again, and the same operation is repeated until the result is no longer 0. When a non-zero result is obtained for the first time, the natural number at that time corresponds to the input speech length. From this, similar to the embodiment shown in FIG. 1, it is possible to select a feature having approximately the same length as the input speech from the features of the words in the dictionary, depending on the number. At this time, if only feature amounts of approximately ±30% of the input speech length are compared, the objects to be matched are limited, and it is possible to aim for faster recognition. Note that it goes without saying that the same effect can be obtained even if the voice section cutout and feature extraction section are arranged in reverse. Further, the natural number generator need not be a very large number and may be 0 or 1; in this case, it is effective to add 1 or another number to the arithmetic unit. Furthermore, as mentioned above, the operation may be division.

As is clear from the explanation in Section 2, according to the present invention, a speech recognition device with high recognition speed can be provided.

[Brief explanation of drawings]

FIGS. 1 and 2 are configuration diagrams for explaining embodiments of a speech recognition device according to the present invention, respectively. 1...Microphone, 2...Reset switch, 3...
Speech section detection unit, 4, 5... Register, 6... Multiplier (divider), 7... Dictionary unit, 8... Feature extraction unit, 9
... Verification section, 10... Judgment section, 11... Output section, 1
2...Natural number generator.

Claims (2)

[Claims] (1) A register for temporarily recording input speech in a speech recognition device that has a dictionary section that accommodates speech characteristics and identifies the input speech by comparing and collating the features of the input speech with the features of the dictionary section. , a reset switch for the register, a register in which numbers or symbols are recorded in advance, and an arithmetic unit that calculates the product or quotient of the contents of the register and the register in which the input voice is recorded, and a dictionary is created based on the result of the arithmetic operation. A speech recognition device characterized by referencing. (2) A speech recognition device that compares and collates the characteristics of the input speech with a dictionary section that stores the speech features and identifies the input speech based on the result, including a register that temporarily records the input speech and a means for generating a natural number. , means for increasing or decreasing the number, and means for taking the product or quotient of the contents of the register and the natural number, and referring to a dictionary based on the result.