JPS63261400A - Voice recognition method - 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] [Summary] The present invention is a speech recognition method that compares registered words and input words to determine speech recognition candidates. In order to solve this difficult problem, a compressed pattern is generated by dividing the input speech by the average number N of syllables determined in advance for the group of words to be speech recognized, and this compressed pattern is compared with the compressed word group registered in advance. By determining the speech recognition candidates, the number of speech recognition candidates can be reduced with a small amount of calculation.

[Industrial application field]

In the present invention, the average number of syllables N of a word group to be speech recognized is
The present invention relates to a speech recognition method configured to divide input speech into compressed patterns, generate compression patterns, and compare them with registered compression patterns to determine speech recognition candidates.

[Problems to be Solved by the Prior Art and the Invention] As a method for reducing the amount of calculation and effectively reducing the number of speech recognition candidates, there is a method that compresses and matches patterns on the time axis. In this method, audio data is divided into sections, an average pattern within this section is determined, and the average patterns are compared. One way to implement this method is, for example, to create a compression pattern for each syllable section as a section on the time axis. In this case, 1! for the word to be voice recognized in advance! The compression pattern provided can be easily generated by generating a compression pattern for each syllable. However, human speech is generated continuously and it is difficult to divide it into syllables.
There is a problem in that it is difficult to generate a compressed pattern compressed into syllable units [Means for solving the problem] In order to solve the above problem, the present invention aims to solve the above-mentioned problem. Divide the word template created for this word using the average syllable fiN,
The compression pattern 6 is created in advance to express the characteristics of the voice for each divided section, and the input voice is divided by the average number of syllables N, and a compression pattern representing the characteristics is created for each divided section. A pattern creation section 8 and a matching section 9 that compares the compressed pattern created by the compressed pattern creation section 8 with the compressed pattern 6 to calculate the degree of matching are provided. Words with a high degree of accuracy are determined as speech recognition candidates.

FIG. 1 shows a basic configuration diagram of the present invention. Syllable registration part 1 in the diagram
This registers a monosyllable as a syllable template 2 based on discrete monosyllable speech.

The word template generation unit 3 generates a word template by combining monosyllable templates 2 corresponding to each syllable of a word key-input in kana notation.

The average number of syllables calculation unit 4 calculates the average number of syllables N of a word group manually written in kana notation.

The compressed pattern creation unit 5 divides the word template generated by the word template generation unit 3 by the average number of syllables N, and creates a compression pattern 6 for the divided sections.

The speech analysis section 7 analyzes the characteristic points of the input speech to be speech recognized.

The compression pattern creation section 8 divides the input speech into segments using the average number N of syllables calculated by the average syllable number calculation section 4. JL,
A compression pattern for this divided section is created.

The matching unit 9 compares the compressed pattern created by the compressed pattern creating unit 8 with the compressed pattern 6 that has been compressed and registered in advance to calculate the degree of matching.

The determining unit 10 determines the one with a high degree of matching calculated by the matching unit 9 as a speech recognition candidate.

[Effect]

Next, the operation will be explained.

In FIG. 1, a syllable template 2 is registered by inputting discrete monosyllabic speech into a speech registration section 1 manually. The word template generation unit 3 generates a word template by combining syllable templates 2 corresponding to words to be voice recognized manually in kana notation. The average number of syllables calculation unit 4 calculates the average number of syllables N of words manually written in kana notation.

The compression pattern creation unit 5 divides the word template generated by the word template generation unit 3 by the average number of syllables N notified from the average syllable number calculation unit 4, and creates compression patterns 6 for each of the divided sections. . As a result, compression pattern 6 '36M of the word to be speech recognized is created.

Next, when the human voice is notified to the compressed pattern creation unit 8 via the voice analysis unit 7, the input voice is divided by the average number N of syllables calculated by the average number of syllables calculation unit 4, and the divided sections are A compression pattern is created for each. The matching unit 9 matches the created compressed pattern with the compressed pattern 6 created in advance to calculate the degree of matching. The determining unit 10 determines that the one with a high degree of matching notified from the matching unit 9 is a speech recognition candidate and outputs it.

As described above, the average number N of syllables of the word to be recognized is calculated, the human speech is divided by this average number N of syllables, compressed patterns are created, and the degree of matching is determined by comparing them with pre-registered compressed patterns. By determining the highest value as a speech recognition candidate, it is possible to reduce the number of speech recognition candidates with a small amount of calculation.

〔Example〕

Next, the structure and operation of one embodiment of the present invention will be specifically explained using FIGS. 1 to 4.

FIG. 2(A) shows an example of a word template corresponding to the kana notation "Toukiyo" generated by combining the first homosyllabic template 2. This word template example "Tokyo" consists of four syllables as separated by dotted lines. If the average number of syllables N of all the words registered as speech recognition targets is calculated as, for example, 3, this word template "Tokyo" is divided equally into sections 1 to 3 as shown by the solid line in the figure. be done.

Figure 2 (B) shows section 1 divided into three parts shown in Figure 2 (A).
An example of a compression pattern generated for section 3 is shown below. This compression pattern is, for example, as shown in FIG.
The accumulated spectrum intensities between each section 1 to section 3 in FIG. 2(a) are calculated for, for example, 20 channels in the frequency direction.

Next, the procedure for generating compressed pattern 6 will be explained in detail using FIG.

FIG. 3(a) shows the procedure for registering a syllable template. ■ in the figure indicates a state in which voice input for registration is performed. This is done by inputting discrete monosyllabic sounds to the 11th syllable registration unit 1, for example, inputting sounds divided into syllables such as "to", "tsu", "kyo°," and "tsu". It means.

■ in the figure indicates a state in which a syllable template is created. This means that a speech tenboot is created from the discrete monosyllabic speech input to the speech registration section 1 in FIG.

■ in the figure indicates the state to be registered. This shows the state in which the voice template created in section 3 in the figure is registered in a file.

■ in the figure indicates a state in which it is determined whether or not all syllables have ended; if YES, the process ends.

In the case of No, repeat the steps below (■) in the figure.

Through the above procedure, all syllable templates (for example, 101 types) of the word to be voice recognized are registered in the file as syllable template 2.

FIG. 3(b) shows a procedure for registering a compression pattern of a word to be voice recognized. 0 in the figure indicates a state in which a word in kana notation for registration is entered manually. This means that the 11th word template generation section 3 and the average syllable count calculation section 4 are notified of a word written in kana, for example, "Tokyo" by keystroke.

0 in the figure indicates a state in which the average value of the number of syllables is calculated. This means that the average number of syllables calculation unit 4 in FIG. 1 calculates the average number N of the number of syllables of the word in kana written manually.

In the figure, @ indicates a state in which a word template is generated. This means that the 11th word template generation unit 3 generates a word template, for example, the word template shown in FIG. 2 (a) by combining the syllable templates 2 corresponding to the key-generated word in kana notation, for example, "Tokyo". It means.

In the figure, [phase] indicates the state in which the compression pattern is created.

This means that the compressed pattern creation section 5 in FIG. As shown in (a), it is divided into three equal parts from section 1 to section 3, and each divided section 1 to section 3 is divided into three equal parts.
This means creating a compression pattern as shown in FIG. 2 (b), respectively.

■ in the figure indicates the state to be registered. This means that the compressed pattern 6 created by @ in the figure is registered in the file.

Through the above procedure, all compressed patterns 6 of words to be speech recognized are created and registered in a file.

Next, the procedure for determining recognition candidates for human speech will be explained in detail using FIG.

In FIG. 4, ■ indicates a state in which analysis is performed. This means analyzing the characteristics of the input voice.

0 in the figure indicates a state in which a compressed pattern is created by dividing. The compression pattern creation unit 8 in Figure 1 divides the human voice into equal parts by the average number of syllables N notified from the average syllable calculation unit 4, and each divided section is shown in Figure 2 (B). This means creating a compression pattern like this.

■ in the figure indicates a state to be compared. This is done by comparing the compressed pattern created by the 11th matching unit 9 at 0 in the figure with the compressed pattern 6 registered in advance for voice recognition in Figure 3 (b), and then formulating the following formula. This means that the degree of matching S is calculated using

Here, fi(j) is the compression pattern for speech recognition, f+'(j) is the compression pattern 6 created and registered in advance for speech recognition, i is the average number of syllables N=3, and j is the compression pattern shown in Figure 2. As shown in (c), it represents the number of divided channels (20) in the frequency direction of the accumulated spectrum intensity between section 1 and section 3.

In the figure, [phase] indicates the sorting state. this is,
In the figure, 0 indicates the highest degree of matching calculated using 2 (1) (S
This means to rearrange them in ascending order (from smallest to largest).

In the figure, [phase] indicates the state to be determined. This means that among the items rearranged in descending order of degree of matching according to [phase] in the figure, for example, the items from the beginning to a predetermined number are determined as speech recognition candidates.

As described above, continuous input speech is divided on the time axis using the pre-calculated average number of syllables N, and the compression patterns generated for each section are compared with the pre-registered compression patterns to find the one with a high degree of matching. By determining things as speech recognition candidates, it solves the difficulty of dividing continuous human-generated speech into syllable units, easily generates compressed patterns, performs matching, and determines speech recognition candidates. becomes possible.

〔Effect of the invention〕

As explained above, according to the present invention, a compression pattern is generated by dividing input speech by the average number N of syllables determined in advance for a group of words to be speech recognized, and a compression pattern of words in which this compression pattern is registered in advance. Since the system adopts a configuration in which speech recognition candidates are determined by comparing with the syllables, it is possible to solve the difficulty of dividing continuous input speech into syllables and to easily determine speech recognition candidates with a small amount of calculation.

[Brief explanation of drawings]

FIG. 1 is a diagram illustrating the principle of the present invention, FIG. 2 is an explanatory diagram for creating a compression pattern, FIG. 3 is a flowchart for creating a compression pattern, and FIG. 4 is a flowchart for determining recognition candidates for input speech. In the figure, 2 represents a syllable template, 3 represents a word template generation unit, 4 represents an average syllable number calculation unit, 5.8 represents a compression pattern creation unit, 6 represents a compression pattern, 9 represents a collation unit, and 10 represents a determination unit.

Claims (1)

[Claims] In a speech recognition method in which speech recognition candidates are determined by extracting and compressing the features of input speech and comparing them with pre-registered compression patterns, a group of words written in kana to be speech recognized is provided. Divide the word template created for this word using the average number of syllables N, and use the compression pattern (6) created in advance to represent the characteristics of the voice for each divided section, and the average number of syllables for the input voice. Divide by number N,
A compressed pattern creating unit (8) creates a compressed pattern representing the characteristics for each divided section, and the compressed pattern created by this compressed pattern creating unit (8) is compared with the compressed pattern (6). The matching unit that calculates the degree of matching (
9), and is configured to determine words with a high degree of matching calculated by the matching unit (9) as speech recognition candidates.