JPS6266299A - 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 speech recognition device.

Prior Art The present applicant has developed a BTS system in which speech generated word by word is binarized to obtain a feature pattern, and the feature pattern obtained by this binarization processing is linearly matched with a dictionary pattern for recognition. P (Binary T S P) was proposed.

However, in this BTSP method, vowels are given a large weight in recognition, and differences between consonants do not appear much, which tends to cause misrecognition of, for example, ``doku'' and ``billion''.

The present invention was made in view of the above-mentioned circumstances.
In particular, this was done with the aim of improving the recognition accuracy of consonant parts in a speech recognition device.

SUMMARY OF THE INVENTION In order to achieve the above object, the present invention includes a voice collection section,
It has a feature converter, a voice section detecting section, and a standard pattern storage section.The feature converter converts the voice obtained by the voice collecting section into a feature, and extracts only the portion related to the voice section. is extracted and its similarity (distance) to a pre-registered standard pattern is determined.

In a speech recognition device that outputs the highest (lowest) similarity (distance) as a recognition result, the power of the voice is determined, and the similarity is calculated by assigning weights that have an inverse (positive) relationship with the increase or decrease in power. It is characterized by calculation.

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

FIG. 1 is an electrical block diagram for explaining an example of a speech recognition device to which the present invention is applied. is a pattern storage section, 5 is a changeover switch, 6 is a matching section, 7 is a result output section, 8 is a DP matching section, 9 is an interframe distance section, 10 is a frame power multiplication section, and the matching section 6 is the same as shown in FIG. It is characterized by the fact that it looks like this. Therefore, the present invention is as follows.

It was developed by focusing on the fact that consonants have lower power than vowels, and it is comprised of a voice collection section, a feature value conversion section, a voice section detection section, and a standard pattern storage section. The voice obtained in the section is converted into a feature by feature conversion, only the parts related to the voice section are extracted, and the similarity (distance) with the standard pattern registered in advance is calculated.
In a speech recognition device that outputs the highest (lowest) distance (distance) as the recognition result, the power of the speech is determined, and the degree of similarity is calculated by assigning weights that have an inverse (positive) relationship with the increase or decrease in power. This is how it was done.

As described above, in the present invention, the matching unit 6 of the general speech recognition device shown in FIG. 1 is configured as shown in FIG. Extract features. The voice section may be extracted from the period from when the voice power exceeds a certain value to when it drops, and the feature extracting section may perform frequency analysis using a group of band-pass filters. When creating a standard pattern, the output of this bandpass filter is 12 to 16 bits every 10 msec.
Sample and store it. During recognition, the system compares the extracted unknown cover patterns with registered standard patterns, and outputs the most similar pattern as the recognition result. For matching, a known method such as a method using dynamic programming based on interframe distance (DP matching) may be used. The number of bandpass filters is n, and the input cover pattern is: If i and the standard pattern are bi, the interframe distance is.

When performing DP matching, the distance between frames is determined and the distance is proportional to the frame power, and the smaller the power, the greater the weight.

This makes it possible to recognize parts with low power.

Here, distance is used, but when using similarity, it is sufficient to divide the part to which frame power is applied. Also, if the power is multiplied, the distance of the part where the output is 0 becomes O, so to correct the problem, it is sufficient to subtract the power from the value α, which is larger than the maximum power, and divide the frame width distance by this.

FIG. 3 is a main part configuration diagram showing an example of the above-mentioned case. In the figure, 11 is a power calculation section, 12 is a part of α-power, and 13 is a division section. As shown in the figure, the subtraction unit 12 subtracts the power from α, which is larger than the maximum power, and the division unit 13 divides the inter-frame distance by this subtracted value as described above. It is.

As is clear from the above description, according to the present invention, a large weight is given to consonant parts, making it possible to perform highly accurate speech recognition.

[Brief explanation of the drawing]

FIG. 1 is a diagram showing an example of a speech recognition device to which the present invention is applied, FIG. 2 is a detailed diagram of the matching section shown in FIG. 1, and FIG.
The figure is a main part configuration diagram showing another embodiment of the present invention. DESCRIPTION OF SYMBOLS 1... Microphone, 2... Voice section detection section, 3... Feature amount extraction section, 4... Pattern storage section, 5... Changeover switch, 6... Verification section, 7... Result output section, 8
... DP matching section, 9... inter-frame distance section,
DESCRIPTION OF SYMBOLS 10... Frame power multiplication part, 11... Power calculation part, 12... α-power part, 13... Division part.

Claims (1)

[Claims] It has a voice collection section, a feature amount conversion section, a voice section detection section, and a standard pattern storage section.The voice obtained by the voice collection section is converted into a feature amount by the feature amount conversion section. A speech recognition device extracts only the part related to the section, calculates the similarity (distance) to a pre-registered standard pattern, and outputs the one with the highest (lowest) similarity (distance) as the recognition result. Seeking power, increasing and decreasing power and vice versa (
A speech recognition device characterized in that similarity calculation is performed by applying weights such that a relationship of (positive) is established.