JPS6211898A - 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] [Field of Application of the Invention] The present invention recognizes speech uttered in units of words and phrases,
The present invention relates to a method for selecting a correct sequence from among a plurality of candidate sequences using a dictionary that stores syllable patterns such as doublets and triplets that may appear in a language, and relates to improving the performance of the method.

[Background of the invention]

The conventional system is Information Center IH-6 (1983-3).

As described in Proceedings of the Acoustical Society of Japan 2-7-4 (October 1982), plural syllables obtained by speech recognition are The correct sequence was selected from candidate sequences. However, it is difficult to collect all the consecutive syllable patterns that can appear in Japanese, and the performance with newly coined words has not been considered.

[Purpose of the invention]

An object of the present invention is to provide a syllable recognition device that can obtain high recognition performance even for newly coined words.

[Summary of the invention]

In order to achieve the above object, the present invention inputs a correct answer sequence when the recognition result of input speech is incorrect, inputs the input sequence, and decomposes the input sequence into continuous syllable patterns such as doubles and triplets. Then, the connected syllable pattern is compared with the connected syllable pattern already stored in the dictionary, and if the connected syllable pattern does not exist in the dictionary, the connected syllable pattern is additionally registered in the dictionary.

[Embodiments of the invention]

An embodiment of the present invention will be described below with reference to the drawings. 1st
The figure is a block diagram of an example of a speech recognition device for explaining the present invention in detail. The input word/phrase audio is
The acoustic recognition unit 1 performs a comparison with the phoneme/syllable standard pattern 3, and forms and outputs a lattice of multiple candidates for each syllable.

For the embodiment of the acoustic recognition unit 1, a known technique (for example, see Yajima et al., ``Recognition Method for Japanese Speech Input Device'', IEICE University, 1602 (Sho 6O-3)) can be used.

FIG. 2 is an example of a candidate lattice when the acoustic recognition unit 1 recognizes the phrase "Kangae Tail J."

The candidate lattice is composed of candidate syllables and acoustic similarities. The acoustic similarity indicates that the larger the value, the more acoustically similar the input syllable is. Each syllable is arranged in descending order of acoustic similarity from each content. This candidate lattice is sent to the candidate sequence selection section 2.

There are 6,400 candidate phrase sequences generated by combining syllables from this lattice, and FIG. 3(a) shows the top 10 in terms of sum of acoustic similarities as sequences.

At this stage, the correct answer series "Kangae Tail" is the 6th. The candidate sequence selection unit 2 narrows down the candidate sequences from these multiple sequences using the syllable dictionary 4. As shown in FIG. 3(b), this continuous syllable dictionary 4 stores two continuous syllables that can appear in Japanese, such as "kan" and "ete". These may be stored by position, such as beginning, middle, and end of the phrase. The number of consecutive syllables is not limited to two consecutive syllables, but may be n consecutive syllables (n≧2). Continuous syllable patterns such as "tepi" and "gape" which cannot appear in Japanese are not stored in the concatenated syllable dictionary 4. First, each candidate sequence in FIG. 3(a) is decomposed into continuous syllables.

For example, "Kangaetepir" means "Kan", "Nga",
It is broken down into ``gani'', ``ete'', ``tepi'', and ``pil''. Next, these six types of decomposed continuous syllables are compared with each item in the continuous syllable dictionary 4, and if all six types are present in the continuous syllable dictionary 4, the candidate transfer is performed together with the acoustic similarity sum. On the other hand, if even one of the six types does not exist in the continuous syllable dictionary, the candidate sequence is not stored in the candidate sequence output memory 2'. Through the above operations, the three sequences shown in FIG. 3(c) are stored as recognition candidates together with the acoustic similarity sum in the candidate sequence output memory 2', and the contents of the candidate sequence output memory 2' are displayed on the display device 5. be done.

By the way, when the input speech contains a newly coined word, the number of consecutive syllables stored in the consecutive syllable dictionary 4 may not be sufficient, and the output from the candidate sequence selection section 2 may not include the correct answer sequence. Figure 4(a) shows the word processor
This is an example of a candidate series created from a candidate lattice that is the output of the acoustic recognition unit 1 when recognizing the following. In addition, Fig. 4 (b
) is an example of the continuous syllable dictionary 4. When this dictionary is used, the output of the candidate sequence selection unit 2 is as shown in FIG. 4(c), which does not include the correct sequence "word processor HA". The reason for this is that the continuous syllable "dobu" does not exist in the continuous syllable dictionary 4, and unless "r dobu" is registered, "word processor ha" will not be recognized. In order to solve this problem, when the output series of the candidate series selection section 2 is displayed on the display device 5, if the correct series is not included in the displayed series, the correct series is selected using a means such as the keyboard 7. Enter. The continuous syllable learning registration unit 6 breaks down the input correct answer sequence into continuous syllables. In the previous example, "word" and "-do".

"Dobu", "Pro", "Rose", "Sera", "Tssa",
It is broken down into eight types of disyllables: ``Saba.'' Next, the eight types of input continuous syllables mentioned above are loaded one by one into the input continuous syllable buffer memory 6', and the items in the continuous syllable dictionary 4 are loaded]
Load them one by one into the dictionary concatenated syllable buffer memory 6",
The contents of the input continuous syllable buffer memory 6' and the dictionary continuous syllable buffer memory 6' are compared and verified. If the two contents do not match, the next item in the continuous syllable dictionary 4 is loaded into the dictionary continuous syllable buffer memory 6' and compared with the input continuous syllable buffer memory 6'. If the two contents match, the next input syllable is loaded into the input syllable buffer memory 6', and the items in the syllable dictionary 4 are loaded one by one into the dictionary syllable buffer memory 6'. Compare and check both buffer memories.

The above procedure is repeated, and input continuous syllables that do not match all items in the continuous syllable dictionary 4 are additionally registered as new items in the continuous syllable dictionary 4. In the above example, "Gutter" is additionally registered as a new item.

〔Effect of the invention〕

As explained above, according to the present invention, only sequences that can exist as Japanese are selected from acoustic recognition candidate character strings using a concatenated syllable dictionary. By learning the items, correct candidate sequences can be output.

[Brief explanation of the drawing]

Fig. 1 is an overall configuration diagram of an embodiment of the present invention, Fig. 2 is an example of the output candidate lattice of the acoustic recognition unit when the phrase ``Kandaeteil'' is input, and Fig. 3(a) is the phrase ``Kangaeteil''.
Among the series of syllable combinations obtained from the candidate lattice when inputting
b) is an example of a continuous syllable dictionary, FIG.
a), (b), and (c) are the results of FIGS. 3 (a) and (b) when the phrase [word processor HA] is input, respectively. This corresponds to (c). 1... Acoustic recognition section, 2... Candidate sequence selection section, 2'.
... memory for candidate series output, 3... phoneme/syllable standard button, 4... continuous syllable dictionary, 5... display device, 6... continuous syllable learning registration unit, 6'... input sequence Buffer memory for syllables, 6 "... Puppy memory for dictionary connected syllables, 7... Keyboard. ■Figure 2 v Figure 33 (arrow) (b) No. 412]" (phantom ゛ (C)

Claims (1)

[Claims] A dictionary that recognizes speech uttered in units of words and phrases in units of phonemes or syllables, and stores the correct sequence from multiple recognition candidate sequences, as well as consecutive syllable patterns such as doublets and triplets that may appear in the language. If the input voice is not correctly converted into a character code in a speech recognition device that selects and converts it into a character code or similar code, the correct answer sequence is input, and the input sequence is converted into a double, triple, etc. If the continuous syllable pattern does not exist in the dictionary, if the continuous syllable pattern does not exist in the dictionary,
A speech recognition device characterized in that the continuous syllable pattern is additionally registered in a dictionary.