JPS61240298A - 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] String 1uL1 The present invention relates to a speech recognition device.

Speech recognition devices include those that recognize words and those that recognize monosyllables, and monosyllabic speech recognition is an essential technology for inputting Japanese sentences such as word processors. However, monosyllables are very similar, and since there are only five types of vowels, which make up the majority of monosyllables when uttered, it is necessary to distinguish them only by consonants. Among them / p /
Distinction between t / t / e/ and /, /b/, /d/, /
It is difficult to distinguish between g/ and furthermore between /m/l /n/+ /g/, resulting in a low recognition rate.

l-Itsuke The present invention was made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to provide a speech recognition device that can distinguish and recognize speech that is difficult to distinguish, such as monosyllables.

Toyo-Ichiri The present invention provides a "rounding and voice collecting section" that achieves the above object,
It has an extraction section that extracts an audio signal and a time length detection section that determines the time length of the extracted signal, and the user is informed in advance of the time length of the audio to be distinguished, and the signal passes through the audio signal extraction section. distinguishing voices based on the time length of the signal, or a voice collecting section, an extraction section that extracts the voice signal, and a time length detection section that determines the time length of the extracted signal;
It has a dictionary section that registers standard patterns and a similarity judgment section that judges the degree of similarity.The utterance length during recognition is determined for speech that is particularly difficult to distinguish in pattern matching, and the utterance length is determined during recognition. If the recognition result is classified as the hard-to-distinguish speech, the recognition result is modified based on the input speech length. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is an electrical circuit diagram for explaining one embodiment of the present invention, and in the figure, 1 is a voice collecting section (microphone).

2 is a voice section detection section, 3 is a sampling section, 4 is a counter, 5 is a length classification section, and 6 is a result output section. The user is informed in advance of the time length of the voice to be distinguished, and recognition is made based on the time length of the signal that has passed through the voice signal extraction section.

For example, if a recognition device that recognizes three words is used to recognize the words "yesterday,""kiyo," and "ashida," the user is asked to recognize "kiyo" in advance at normal speaking speed. , when making the child recognize ``yesterday'', say it slower than usual so that the length is about twice as long as ``today'', and conversely say ``tomorrow'' faster so that it is about half the length. Then, if you input any of them through the microphone, it will detect which of the three words it is by detecting the interval of the audio and determining the length of the audio. can be recognized. In the example shown here, the length of the voice is measured by sampling the voice signal and using the number of samples. However, there are several known methods for detecting voice intervals. The determination may be made based on whether or not a certain threshold has been exceeded.If the length of the audio measured using this method is 1 second or more, it is considered "yesterday".

By setting "Kiyo J" for 0.5 seconds to 1 second and "Tomorrow" for 0.5 seconds or less, a speaker-independent word speech recognition device with a very simple configuration can be realized. However, with this method, if the number of words exceeds three, it becomes difficult to confuse the words.

FIG. 2 is an electric circuit diagram for explaining another embodiment of the present invention, in which 11 is a voice collection section (microphone), 12 is a voice section detection section, 13 is a frequency analysis section, and 14 is a sampling section. 15 is a counter, 16 is a register, 17 is a registration/
Recognition changeover switch, 18 is a dictionary section, 19 is a collation section,
20 is a judgment section.

21 is a result correction section, 22 is a result display section, and this embodiment includes a voice collection section, a voice section detection section, a time length detection section for determining the time length of a signal, and a dictionary section for registering standard patterns. and a similarity determination unit, and determines the utterance length at the time of recognition for voices that are particularly difficult to distinguish by pattern matching,
When the recognition result of the input pattern is classified as the difficult-to-distinguish speech, the recognition result is modified according to the length of the input speech. This recognition device recognizes monosyllables, making it difficult to distinguish between /p/ and /l/.

This was done by focusing on the fact that it is easy to misrecognize between the lines 1, ``pa'', ``ta'', and ``ka''. When speaking, the voice should be short, 25 msec or less, and the evening line should be uttered as usual, and the Ka line should be long and at least 40 msec long. /b/.

If / d / t / g / are likely to be confused, create a similar rule for these. Here /p/, /l/
.

Let's explain with an example of / to / only. First, switch 17 is set to the registration side to register the dictionary. This registration is no different from a normal speech recognition device.

After inputting a single syllable from a microphone and converting it into a feature quantity (in this case, a frequency spectrum) and sampling it, the number of samples, that is, the time length, is measured with a counter and stored in the register 16. The sampled feature quantity is collated. Resignation at the department
Match each pattern in the book and find the degree of similarity. As a result, the result with the maximum similarity is / p / e / t
Determine whether the syllable starts with / e / k /,
If not, output the result as is, and if so, if the audio length stored in the register is less than 25ms, it will be output as a line, if it is 25 to 40ms, it will be changed to a line, and if it is longer, it will be corrected to a line and output. .

Normally, the vowel part of a monosyllable can be recognized with almost 100% accuracy, so if the result before correction is / k e
/, and the voice length is only 20 msec, this modifies only the consonant and outputs it as /pe/.

Note that the present invention is not limited to the matching method described above, and any other method such as using DP matching may be used.

As is clear from the above description, according to the present invention, it is possible to realize a speech recognition device with fewer recognition errors.

[Brief explanation of drawings]

4. FIGS. 1 and 2 are electrical circuit diagrams for explaining the present invention in detail, respectively. 1... Voice collection unit, 2... Voice section detection unit, 3...
- Sampling unit, 4... Counter, 5... Length classification unit, 6... Result output unit, 11... Audio collection unit, 1
2... Voice section detection section, 13... Frequency analysis section, 1
4...Sampling section, 15...Counter, 16
...Register, 17...Selector switch, 18.
...Dictionary section, 19... Collation section, 20... Judgment section, 2
1...Result correction section, 22...Result display section.

Claims (2)

[Claims] (1), an audio collection unit, an extraction unit that extracts an audio signal,
and a time length detecting section that determines the time length of the extracted signal, the user is informed in advance of the time length of the audio to be distinguished, and the audio is detected based on the time length of the signal that has passed through the audio signal extracting section. A speech recognition device characterized by distinguishing between. (2) an audio collection unit and an extraction unit that extracts audio signals;
It has a time length detection unit that calculates the time length of the extracted signal, a dictionary unit that registers standard patterns, and a similarity determination unit that determines the degree of similarity. A speech recognition device characterized in that the length of speech during recognition is determined, and when the recognition result of the input pattern is classified as the difficult-to-distinguish speech, the recognition result is corrected according to the length of the input speech. .