JPS62209598A - Word voice recognition processing 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] [Summary] An improved processing method for speech recognition of multiple words. By comparing the speech pattern with the word speech pattern in the dictionary, the first word recognition section selects one or more candidate words, and the second word recognition section, which uses the dictionary of the word phonetic symbol network, sets the dictionary reference range to the candidate word. Execute recognition processing limited to words only. This method allows the second word recognition unit to perform recognition processing with a relatively high recognition rate at high speed by limiting the dictionary reference range.

[Industrial application field]

The present invention is a processing method that recognizes the sounds of a large number of words. A method of operating a document creation device and a device that uses a variety of commands by voice is about to be put into practical use. In order for such a device to become practical, it must have an economical configuration that can recognize a relatively wide variety of word sounds at a relatively high speed in real time.

[Conventional technology]

Two known methods are in practical use as word speech recognition methods.

In the first method, information on the phonetic patterns of all required words (
spectral information of word sounds at appropriate time intervals, etc.)
A word dictionary is created in which words are paired with the words expressed by the sounds.

In the speech recognition process, the input speech pattern is compared with the information in the word dictionary, the distance indicating the degree of agreement between the two patterns is determined for each registered word, and the word with the smallest distance is taken as the recognition result.

In comparing both patterns, a matching method using dynamic programming, which is a method for correcting the expansion and contraction of the time axis of speech utterance, is known as a method for improving this method.

In the second method, a word dictionary is created from the phonetic symbols of words, using a phonetic symbol network created using phonetic knowledge as dictionary information.

In speech recognition processing, the input speech pattern is compared with the information in the word dictionary, and the distance between the input speech pattern and the speech pattern corresponding to each phonetic symbol in the phonetic symbol string passing through any branch of the phonetic symbol network is calculated. Take the smallest one as the resulting word.

As is well known, the phonetic symbol network is, for example, when "Tomorrow ga" is a word, its pronunciation is expressed in romaji rASITAGA J, and then it is expressed in the actual pronunciation based on so-called phonetic knowledge. The aspect is conceptually represented by a network having the following configuration, for example.

As a result, in the above example, for the word "Tomorrow ga", 4
It is possible to systematically process different kinds of phonetic symbol strings as comparison targets.

In this example, the phonetic knowledge used to construct the phonetic symbol network is that in spoken language, SI becomes SHI, and the voiceless S
The 1 between H and T may be devoiced, a silent break may be inserted before the T, and the G may be nasally pronounced and pronounced as NG. .

This comparison process for recognition is a relatively complicated process that requires so-called artificial intelligence techniques, so the island drive method, LHFT-T
The o-RIGHT method, the matching method using dynamic programming, and the like are known.

[Problem that the invention seeks to solve]

In the first method, large-scale integrated circuits have been developed for speech recognition of several hundred types of words, and high-speed processing is possible. Therefore, when there are many types of recognized words, this circuit can be used in parallel. There is a possibility of equivalent high-speed processing.

However, since the first method relies only on comparison with registered individual word sound patterns, from these sound patterns, the sound of utterances that have changed due to devoicing, etc., will be recognized unless it is registered separately. Furthermore, when there are many types of words, there are many different words with very similar pronunciations, but such words are easy to misrecognize, so a high recognition rate cannot be expected.

It is well known that the second method uses phonetic knowledge to create a dictionary that covers the variations of utterances, and also uses phonetic knowledge in the matching process with the input speech pattern, which can be expected to achieve a high recognition rate. .

However, since the processing is complex, as the number of words to be recognized increases and the amount of dictionary information to be referenced increases,
Processing time increases rapidly.

For the reasons listed above, there has conventionally been a problem in that there is no suitable method for recognizing a relatively wide variety of word speech in real time at a high recognition rate.

[Means for solving problems]

FIG. 1 is a block diagram showing the configuration of the present invention.

In the figure, engineering is the first word recognition unit, 2 is the first word dictionary, and 3 is the first word recognition unit.
4 is a second word recognition unit, 4 is a second word dictionary, and 5 is a word dictionary restriction unit.

[Operation] The first word dictionary 2 is a dictionary in which the speech patterns of each word are registered, and the first word recognition unit 1 uses the first word dictionary 2 to determine the above-mentioned speech pattern input from the signal line 6. 1st
Execute speech recognition processing using the method described below.

However, instead of determining one word as a recognition result, a plurality of predetermined number of candidate words are selected as pre-processing for the second word recognition section 3 and notified to the word dictionary restriction section 5.

The second word dictionary 4 is a dictionary in which phonetic symbol network information of each word is registered, and the second word recognition unit 3 uses the second word dictionary 4 to determine the second Speech recognition processing is executed using the method described above, and the determined word is output as a recognition result of the input speech.

At this time, the dictionary information referred to by the second word recognition section 3 is only the dictionary information of the candidate word specified by the word restriction section 5 and previously received from the first word recognition section 1 by the word dictionary restriction section 5. shall be.

With the above method, by performing rough recognition processing for selecting multiple candidate words in the first word recognition unit capable of high-speed processing, the reference range of the dictionary used in the high recognition rate processing of the second word recognition unit is Because we narrow it down to a relatively small number of words,
It becomes possible to process high recognition rate processing at relatively high speed.

〔Example〕

FIG. 2 is a block diagram showing the configuration of a speech recognition device according to an embodiment of the present invention.

The first word recognition unit 11 uses the first word dictionary 12 to perform speech recognition processing according to the first method on the input speech pattern from the signal line 16.

The first word recognition unit 11 is configured, for example, by parallel processing of the aforementioned large-scale integrated circuit capable of high-speed processing.

The second word recognition unit 13 uses the second word dictionary 14 to perform speech recognition processing according to the second method on the input speech pattern from the signal line 16.

For this purpose, the second word dictionary 14 is configured as a dictionary in which the above-mentioned phonetic symbol network information for each required word is registered.

The word dictionary synthesis unit F uses the phonetic symbol network information of the second word dictionary 14 and the phonetic pattern information 18 of each phonetic symbol.
The first word dictionary 12 is generated by synthesizing the speech patterns corresponding to, for example, all the phonetic symbol strings of each word and storing them together with the word names.

By doing this, there is no need to construct the first word dictionary 12 from actual speech, and a word dictionary accommodating a large number of words can be constructed in a short period of time through completely automatic processing.

The first word recognition unit 11 processes the voice pattern input from the signal line 16, selects a plurality of predetermined number of candidate words, and notifies the word dictionary restriction unit 15.

The second word recognition unit 13 receives the input speech pattern from the signal line 16 in parallel with the first word recognition unit 11, and executes recognition processing using the second word dictionary 14 and the speech pattern information @i18.

The dictionary information to be referred to at this time is limited by the word dictionary restriction section 15 to only the dictionary information of candidate words that the word dictionary restriction section 15 has previously received from the first word recognition section 11 .

〔Effect of the invention〕

As is clear from the above description, the present invention has a remarkable industrial effect in that a method for performing high-speed processing with a high recognition rate can be realized economically when recognizing a large number of word sounds. There is.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing the configuration of the present invention, and FIG. 2 is a block diagram showing the configuration of an embodiment of the present invention. In the figure, 1.11 is the first word recognition unit, 2.12 is the first word dictionary, 3.13 is the second word recognition unit, 4.14 is the second word dictionary, 5.15 is the word dictionary restriction unit, 6. 16 is a signal line, 17 is a word dictionary synthesis unit, 18 is voice pattern information. Figure 1 is a block diagram showing the configuration of the present invention. Figure 2 is a block diagram of the configuration of an embodiment of the present invention.

Claims (1)

[Claims] When determining a word represented by an input audio signal from an input audio pattern of the input audio signal, a first word dictionary (2) that uses the audio pattern of the desired word as dictionary information; a phonetic symbol for the required word; A first word recognition unit ( 1) means (5) for limiting the reference range of the second word dictionary (4) to only the dictionary information of the candidate word; and speech recognition processing of the input speech pattern using the limited dictionary information. A word speech recognition processing method characterized by comprising a second word recognition unit (3) that executes.