JPH02212899A - Voice recognition system - 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] Chiyiko mutual The present invention relates to a voice recognition method.

With so-called speaker-specific speech recognition devices, in which users register their voices in advance and then use them, is there no difference in the number of utterances in which the recognition ability of the device depends on the quality of the standard patterns created through training?

It is important that the pattern is registered as a normal pattern.

For example, as shown in Japanese Unexamined Patent Publication No. 59-20114i, there is also a method of registering while listening to the voice uttered by the user. However, 1. For example, a long word such as ``Mstislav Rostlovovich J'' is called ``Mustis.''
If you register it as abbreviated as ``Rosudoro'' or ``Rosdro,'' you will forget what you registered it as and have to re-register it at a later date. For that purpose, JP-A-57-8
There is a method of recording audio as in the 6979 publication, but this requires a separate large memory and a recording/playback device. In order to solve these problems, it is sufficient to be able to reproduce the sound uttered at the time of registration from a dictionary for speech recognition.
0 times.

Information Theory and Its Applications Symposium, No. 19-21.
1987. pp. 475-480) had the disadvantage that sufficient audio information for reproduction could not be obtained.

Purpose The present invention was made in view of the above-mentioned circumstances.
In particular, voice recognition shows the utterances uttered during registration.

This was done with the aim of improving recognition accuracy by reminding users of the utterances they uttered at the time of registration.

In order to achieve the above-mentioned object, the present invention analyzes the frequency of audio, processes it into binarization to obtain a standard pattern, and further divides the frequency analysis results into types based on the distribution of high and low frequencies. In the speech recognition method, which is registered in conjunction with the standard pattern and recognized, it has a pulse sound source and a noise source, and modulates the output of the sound source with the characteristic frequency of the standard pattern, and furthermore, The present invention is characterized in that it selects one of the two sound sources according to day and night representing the frequency distribution, and includes a part that converts the modulated wave into an audible wave and outputs it.

Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a block diagram for explaining the details of the present invention. In the figure, 1 is an acoustic/electrical signal converter, 2 is a speech analysis section, 3 is a switch, 4 is a dictionary section, 5 is a recognition section, 6 is a reproduction section, 7 is an output, and 8 is an electric signal/acoustic converter, and a switch 3 allows selection of dictionary registration (side a) and recognition m (side b).

However, this diagram assumes a specific speaker method.

Needless to say, such a switch is not necessary in the speaker-independent system.

FIG. 2 is a detailed view of the acoustic/electrical signal converter 1 and the acoustic divider 2 shown in FIG. 1, as is well known.

A microphone is used as the acoustic/electrical signal converter 1, and the output thereof is amplified by a microphone amplifier A and filters F1 to F1. Enter. This filter is a bandpass filter with 15 filters arranged in a row, but their number and characteristics have no meaning.Normally 150-
1. It is sufficient that all OOOOOHz or only important parts are covered.After rectifying the output with rectification mR□~R□ and using each data to find the least square straight line in the least square straight line part 9, A digitization unit 10 binarizes the data.

Figure 3 is a diagram showing the process. After the filter output, the WI
If the streamed data is arranged from the lowest frequency to the lowest frequency, (a
) However, in (a), 15 points should be lined up correctly, but this is shown as a continuous line. This is minus the minimum own device ALL. By subtracting each value of L from each output value, a result like (b) is obtained. By binarizing the peak of the mountain on the positive side as ``1'' and the others as ``O'', we can obtain 2 values like (Q).
A value pattern is created. At this time, when the slope of the least squares straight line in (a) is negative, 1. When it is positive, O is added next to the 15 data in (c) to create 16 data sets. A standard pattern is created using the data thus obtained and registered in the dictionary section. The creation of the standard pattern is described in the above-mentioned literature, but each 2@ created by uttering it multiple times
There is a method that adds digitization patterns. However, the present invention is not limited to this. In the case of recognition, this 2+fi data is transferred to the recognition unit, and the most similar standard pattern is output as the recognition result by comparing it with the pattern in the dictionary. The method of matching in this case is not particularly limited and may be any method, but it is best to find the similarity from the degree of overlap between the input and dictionary patterns as shown in the above-mentioned literature. Are suitable.

FIG. 4 is a diagram showing the details of the playback section, in which from the dictionary section 4, 1
Data arranged in groups of six are sent at regular intervals. The shorter this interval is, the better the quality of the reproduced sound will be, but since the amount of data will increase, it is best to use the interval used for normal speech recognition, that is, about 5 to 10 rns. Switches 81-S treated as 16ch
15 is ON.

The OFF switch turns ON or OFF depending on whether the data of channels 1 to 15 is O or not. The switch S□ changes the sound source depending on whether the data of L6ch is 0 or not. That is, 16c
If the h-th data is 0, the noise source 13. Otherwise, the pulse sound source 12 is connected to the switches S1 to StS. However, for values above 1, the switch operation is divided into 0 and other values, but it is better to change this value depending on the average number of standard patterns.For example, if you create a standard pattern by adding three patterns, It is preferable to change the operation of the switches between 0 and 1 and 2 and 3. The data that has passed through switches 81 to S is sent to filter F! ~F1. is applied to In this case, it is desirable that the filter be the same as the filter analyzed in FIG. 2, and care must be taken to ensure that the data is input to the same filter as the filter used when it was analyzed. Assuming that the center frequency of the filter increases in ascending order of number, and the 1 analysis results are set from lch to 15ch from the lowest frequency,
The data analyzed by filter n becomes nch data, and is input to filter n when being reproduced. The sum of the outputs thus obtained is determined by the adder 11, and after being amplified by the amplifier A, becomes data for driving the speaker. The period of the pulse sound source 12 is preferably close to the human pitch period.

Approximately 200-300 Hz is appropriate.

With the above-described configuration, it is possible to reproduce the voice from the binarized data for voice recognition, and it becomes possible to hear what was uttered and how it was uttered.

As is clear from the explanation of "Beikai" 2, according to the present invention, the standard patterns in the dictionary are not only made audible and it is possible to recall the utterance at the time of registration, but also the standard patterns that are unnecessary are If there is a sound and it is registered, you will be able to hear it. As a result, the recognition accuracy of the device can be improved.

[Brief explanation of the drawing]

FIG. 1 is a configuration diagram for explaining the present invention in detail, and FIG.
The figure shows details of the acoustic/electrical signal converter 1 and the voice analysis section 2 shown in FIG. 1, FIG. 3 is a diagram for explaining an example of binarization, and FIG. FIG. 3 is a detailed diagram of the reproduction section 6 shown in the figure. 1...Acoustic/electrical signal converter, 2...Speech analysis section,
3... Switch, 4... Dictionary section, 5... Recognition section,
6... Reproducing section, 7... Output, 8... Electric signal/acoustic converter, 9... Least square linear section, 10... Binarization section, 11... Adding section. 12...Pulse sound source, 13...Noise source. Figure 1 Figure 2

Claims (1)

[Claims] 1. Analyze the frequency of the audio, process it into binarization to create a standard pattern, and also register the results of classifying the frequency analysis results into types based on the distribution of high and low frequencies along with the standard pattern. In a speech recognition method that uses a pulse sound source and a noise source, the output of the sound source is modulated at a characteristic frequency of the standard pattern, and the output of the sound source is modulated at a characteristic frequency of the standard pattern, and which of the two sound sources is determined according to data representing the frequency distribution at that time. 1. A speech recognition method, comprising a part for selecting one of the modulated waves and outputting the modulated wave as an audible wave.