JPS5814199A - Voice recognizer - 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 The present invention relates to a speech recognition device capable of recognizing the speech of one person.

In general, this type of speech recognition device is divided into two types: devices for specific speakers, which recognize only the voices of registered speakers, and devices for non-specific speakers, which recognize the voices of many people. Separated.

However, there is currently no speaker-independent device that can completely handle the voices of any person, and it is necessary to register in advance a number of specific voices that represent individual differences in voice, and use these registered voices and input voices. The speech recognition device for specific majority speakers is used as a substitute for the device for unspecified speakers, and there is no particular difference between these recognition devices for specific and unspecified speakers.

FIG. 1 shows a speech recognition device in which the voices of a specified number of speakers are registered for unspecified speakers. In the same figure, 11
) is a microphone that converts sound into electrical signals, (2)
is the eight spectral values xi (i=1+2
It is a group of 8 parallel bandpass filters that extracts
00 (approximately IHz) divided into eight parts. (3
) is an audio pattern creation circuit that samples the eight spectral values, 27t, obtained from the bandpass filter group (2) and calculates 1 for the existing range of the input audio.
6 samples are collected. (4) is an input cover turn memory, and the above band-pass filter group (input audio consisting of 16 samples of 8 spectra □ values xi from the field) (turn Cx
1j ) + (izt 1 2 , -,
s ) ° e l , 2 , I+
.

16) can be stored. (5) is a reference pattern memory in which reference speaker patterns [yLj)rILrL are registered in advance for a plurality of N words of M speakers.

(m-1, 2, -, M n-], 2.-.., N
) is stored. +61 is a distance calculation circuit, which uses the reference speaker pattern Cy of the above-mentioned turn memory (5).
ij]mn and the input voice pattern in the cover pattern memory (4) is calculated. (7) is a recognition processing unit, which calculates the distance @ d (m, r
The input speech at this time is recognized as the first recognized word. Such a speech recognition device uses the matrix pattern CMi in the distance calculation circuit (6).
The distance D (m, n) between Cyij”: J
If you increase M in order to accommodate the voices of a larger number of people or to improve the recognition rate, a large amount of calculation processing will be required, making it difficult to recognize voices in real time. There was a fear that this would happen.

The present invention has been made in view of the above circumstances, and it is an object of the present invention to provide a speech recognition device that simplifies calculation processing for recognition.

FIG. 2 shows a speech recognition device of the present invention. In the same figure,
(1) to [4] indicate a microphone-input cover pattern memory similar to the conventional device shown in FIG.

(8) is a standard speech pattern memory, in which S-quasi standard speech patterns C9'j are stored in advance for each of N recognized words.
'In, (ts=1.2.-A') is stored.

(9) is a distance calculating circuit which calculates the distance between each standard pattern [No. 7] in the standard pattern memory (8) and the input voice pattern in the input cover turn memo 9 (4).

(1 (I is the distance obtained from the distance calculation circuit (8)'
l ) le g) = (dllll, d(2), -, d(M)
This is a distance vector memory in which trX is stored. 01) is a reference distance vector memory, in which each speech butter 'C'/') ], -", (m-1+2 m" Good luck 1
M n-1, 2, ..., N) and the standard voice pattern [y series] of the standard pattern memory (7), (n+wl
Distance vector [+) mn-(dV'l?,
d (舊几..., d shame.

(m=1, 2,..., M n-1, 2,...N)
are stored in a matrix with mn columns. 3 is the error calculation circuit described above, and the reference distance vector memory 01
The error δ(door, rL) −0) −rDm n between each reference distance vector IDmn of 1 and the distance vector of the distance vector memory GO is calculated. α3 is a recognition processing unit, and by detecting the time when the error δ(m, n) obtained from the error calculation circuit 02 is the minimum, the voice input to the microphone +11 at this time is recognized as the nth It is determined that it is the th recognized word.

The speech recognition device having such a configuration uses its distance calculation circuit (9) to input the input speech pattern [X°] into N pieces of standard pattern memory (8), one for each word. The standard voice pattern Cy°] is the input voice pattern [X
distance d (N dimension (7) expressed in yzl)
to ffill'7) le1) = (dllll, d
(2+, ・, d(/V))c conversion. Then, in the reference distance pect 7 memory a11, each speech pattern [yij)mn of N words of 1M people is stored in advance by using N standard speeches t<y-y(ij)n to account for individual differences in speech. The reference distance vector l converted into an N-dimensional vector indicating the value [D
771 n-t d'TrV, dl? .

..., dM> are stored in a matrix manner, and depending on the error calculation circuit decoy, these reference distance vectors ptn
The error δ (m, n) between n and the distance vector corresponding to the input voice *that is, the degree of similarity including the component of individual differences in voice is calculated, and the recognition processing unit (2) calculates the minimum δ (
m, n) is detected.

In this way, the matrix-like voice pattern Cx1j).

Cyij] rnn as standard voice pattern (y i j)
A distance vector ρ indicating the value of individual differences in speech depending on n.

Since these p and lp yx a are matched after converting to rDmn, the above audio pattern [X
A recognition process equivalent to directly matching s゛), (yij)mn is performed. In the arithmetic processing during this recognition, since the distance d lnl between the matrix patterns in row t and column ` is calculated N times in the distance calculation circuit (9), subtraction is performed a total of N times.

Since the error calculation circuit 0 calculates the error between the distance vectors consisting of 1M components KNN times, subtractions are performed KN'' times, for a total of 1M subtractions.

Compared to the conventional example shown in FIG. 1 in which the total number of subtractions is ijM, the number of subtractions is approximately reduced by (°-N)MW. That is, as in the example, i
-3, j-16, it is clear that the number of subtractions can be reduced if the number of recognized words N is 128 or less.

As is clear from the above description, the present invention converts an input speech pattern obtained from a pattern creation circuit into a distance vector with a standard speech pattern in a standard pattern memory using a distance calculation circuit, and It is calculated by multiple voice butter error calculation circuits of specific majority speakers, and this error is recognized as the input voice at this time corresponding to the distance vector in the reference distance memory at the time when the error is the smallest. Therefore, compared to the conventional device that directly matches the matrix-shaped reference audio pattern and the input audio re-pattern, the amount of calculation processing can be greatly reduced, and the memory capacity can also be reduced. Therefore, by increasing the number of registered speakers while measuring real-time responses, it is possible to realize a speech recognition device that can handle unspecified speakers with a high recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing the configuration of a conventional speech recognition device.
FIG. 2 is a block diagram showing the speech recognition device of the present invention, where (2) is a group of band-pass filters, (3) is a pattern creation circuit, and (4) is an input cover pattern memory 16)+91 is a distance calculation circuit. , 171 (13 is the recognition processing unit, (8) is the standard Hatern memory, αO is the distance vector memory, al
l is a reference distance vector memory, and d is an error calculation circuit. are shown respectively. Figure 1

Claims (1)

[Claims] (1) A pattern creation circuit that creates a voice characteristic pattern based on the voice signal of input voice, a standard pattern memory that stores standard voice patterns for a plurality of voices in advance, and a Regarding multiple voice patterns of specific majority speakers. A reference distance memory that stores distance vectors consisting of a plurality of distances between each of the voice patterns and the plurality of standard voice patterns in a matrix manner so as to correspond to a plurality of voices of a specific plurality of speakers, and the pattern creation circuit. a distance calculation circuit that calculates the distance between the audio badan obtained from the standard pattern memory and the plurality of standard patterns in the standard pattern memory, and a distance vector consisting of the plurality of distances obtained by the distance calculation circuit and each distance in the reference distance memory. a distance error calculation circuit that calculates an error with the distance vector, detects a distance vector in the reference distance memory that minimizes the error obtained from the distance error calculation circuit, and calculates the audio corresponding to this distance vector. A speech recognition device characterized by recognizing input speech at times.