JPS6024595A - 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 (ii) Industrial Application Field The present invention is directed to a speech recognition device capable of recognizing speech!

1d Conventional speech recognition devices generate a speech pattern A consisting of a time series of feature parameters indicating the characteristics of the speech from a speech signal.
x, (a (11, @ fi+ , ...) are extracted for a plurality of predetermined voices, and an unknown pattern x = consisting of a time series of feature parameters of two unknown voices S is extracted. (χ11+, x'In,...)
For each time series A m (@ (tl, a (tl,...
), and the distance between these two patterns is −D = i A
-X l -Z l'a(il-x(111) was determined to be the voice of time series A with the minimum value as the unknown voice at this time.

However, although a time series of speech spectral values or a time series of autocorrelation coefficients are used as the above-mentioned voice characteristic parameters, these parameters may change to a greater or lesser degree depending on the speech production situation. ＠IJT 1. Because of this, there is a drawback that even if the voice is the same, there is a large difference between the patterns of the previously registered voice and the unknown input voice, leading to misrecognition.

OBJECT OF THE INVENTION The present invention reduces the occurrence of erroneous recognition and improves the recognition rate.
We are currently offering a voice recognition device.

B) Structure of the Invention The speech recognition device of the present invention shows the average value of the feature parameters of one registered speech stored in advance, the time series of the mean value parameter, the time series of its standard deviation, and the unknown unknown. Decorate the time series of the characteristic parameters of the input voice using statistical processing,
When the difference between the characteristic parameter and the average value parameter of this input voice is larger than the standard deviation times the number of feet, and when it is smaller: Converted into a time series of binary signals of ゝゝ1'' and 0'' that indicate the value state. Similarity! Then, the registered voice with the highest degree of similarity is determined to be the input voice at that time.

(E) Embodiment $1 Figure 1 shows an embodiment of the voice recognition snowmaking method of the invention. In the figure, (l) is a microphone that converts voice into an electrical voice signal, (2) is This is a parameter extraction circuit that extracts a frequency spectrum value, which is a characteristic parameter indicating the characteristics of the voice, from the voice signal obtained from the microphone (1). 4000Hri Y8
Each 8 of the divided frequency spectrum values f f ...f8
A speech pattern expressed in time series consisting of 1% 2% samples is obtained. That is, when the filter number is n and the sample number is t, the characteristic parameter is expressed as fn(tl), and the voice pattern F is as follows. (31 is a mode selection switch that switches between registration mode and recognition mode yIl, If it is connected to the Q side, it will be in the registration mode, and if it is connected to the P side, it will be in the recognition mode. (4) is the parameter extraction circuit (2) when in the registration mode when the mode selection switch (31 is connected to the q side). ) is a statistical processing circuit that receives voice patterns from the following inputs, and calculates each feature parameter fn based on a plurality of voice patterns obtained by inputting the same voice a small number of times in succession.
If ltl is a normal distribution YET as shown in Fig. 2, then the average value pattern and standard deviation pattern consisting of the average value parameter fn (tl) are calculated. The average value (turn
'ii-c-■, the average value stored in i-turn memory section (51), and the standard deviation of the turn actual value corresponding to this,
a standard deviation memory unit (52) for storing the saving, t, meeting, and meeting;
It consists of -10 (6) is connected to the mode selection switch 131tP side, and unknown input audio x6 obtained from the parameter extraction circuit 12) in the recognition mode
On the other hand, it is a buffer memory that temporarily stores the voice (turn (5) Each average value of the average value pattern memory section (51) x 1 stone...
Each parameter'; n (tl, 1n(t, l.

A subtracter (71) that subtracts ... and the memory circuit (5
) of each standard deviation pattern t in the standard deviation memory section (52)
, society,...Respective standard deviations: n, z,, society□□,...
A multiplier (72
) and the subtracted value xnit) from the subtractor (71) -
”;, n, tI, x n ltl −M nltl,
...Multiply the above! Multiply value Ktn0 from (72)
．． 5. Government□12. Compare with...' ” ””a
When n(tll-Ka n(kl, output 1'1ゝゝL
I X n1tl, n(tll > K Jln
(tl -a n1
When tl, it is assumed that xn(tl is similar to an(tll2) with a probability of 68.5% based on the normal distribution (2) shown in Figure 2, so 1" is given, In the opposite case, "0" is given, and the unknown pattern is converted into a matrix pattern represented by a binary signal δ of each registered voice (2 vs. C1 "11/%'0").

(S+ is a hole based on the matrix pattern △ of the binary signal obtained from the upper string comparison means (7), the total sum of its 16 components pXδij, that is, 1", the number of existence Z class C degree It is a processing circuit, and this degree of similarity for each registered voice A, B, C, D, and E is, for example.

11.2, 8.7, 3, the input audio at this time is A
It is determined that this was the case.

Then, in the registration mode in which the mode selection switch (31 is connected to In the recognition mode connected to switch (3) P,
An unknown voice is input, and its voice pattern It is converted into a binary signal S pattern in which the fluctuation component of the sound is removed.At this time, if the value of K for determining the permissible fluctuation of the audio is set to about 015 to 2, the 2fT signal pattern This optimally shows the similarity between the unknown speech pattern and the registered speech pattern, and this allows highly reliable pattern execution!

In the above explanation, a voice pattern consisting of a time series of frequency spectrum values is used as the L♀ series of characteristic parameters, and in addition, various characteristic parameters such as the Percoll coefficient, which is an autocorrelation coefficient, are used. The use of series is also possible.

(f) Effects of the Invention As is clear from the above description, the speech recognition device of the present invention shows the average value of the feature parameters of one registered speech stored in advance, and the time series of the mean value parameter and its standard deviation. 7. Perform statistical processing on the time series of the feature parameters of the unknown input voice.
The system converts the binary signals into a time series to determine the degree of similarity, and determines the registered voice when the degree of similarity is maximum as the input voice at that time, so each feature parameter depending on the voice utterance situation is determined. It is possible to remove the variation of T, and it is possible to derive the optimal degree of similarity T, and a wide range of recognition rates can be expected.

[Brief explanation of the drawing]

The first factor is a block diagram showing an embodiment of the speech recognition device of the present invention, Figure @2 is a normal distribution diagram 1, txt is a microphone, 12; is a parameter extraction circuit, and 14) is a statistical processing circuit % (51 is the memory circuit, (7) is the comparison means, (8
) is a recognition processing circuit! shown respectively.

Claims (1)

[Claims] Ri audio! / A voice input means for converting into an electric signal, and a parameter extraction means for extracting characteristic parameters indicating the characteristics of the voice from the electric signal of the input W voice. a storage means for storing a time series of the mean value parameter indicating the mean value of the feature parameter of the voice and a time series of the standard deviation corresponding to the time series of the mean value parameter, and a storage means for storing the time series of the standard deviation corresponding to the time series of the mean value parameter, In response to the time series of the feature parameters, the time series of the average value parameter and the time series of its standard deviation for each input voice are read out from the storage means, and the error value between the feature parameter of the input voice and the average value parameter is determined. The above error value is multiplied by a constant and compared with the standard deviation, and based on this comparison result, the above error value is multiplied by a constant and the binary state when it is larger and smaller than the standard deviation is displayed.T"1", 10" a comparison means for converting the time series of characteristic parameters of the unknown human voice into a time series of binary signals by outputting a binary signal; a recognition processing means that calculates the degree of similarity of each registered voice to the unknown voice and determines that the registered voice with the highest degree of similarity is the unknown voice input at this time; .