JPS5855995A - Voice recognition 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 The present invention converts input speech into a plurality of candidate sequences based on information obtained by a speech recognition system, particularly a speech analysis device, and converts a word dictionary and at least a grammar connection matrix into a plurality of candidate sequences. The present invention relates to a speech recognition system that improves the recognition rate by indexing, converting the candidate string into a meaningful single eyelid string, and extracting the word string as a syntax-checked word string.

Conventionally, in speech AM systems, syllables of input speech are checked against a feature dictionary, and categories with a high degree of similarity are output as recognition results.

However, it is generally difficult to unambiguously determine the category of each syllable, and correct recognition results may not always be obtained.

The present invention aims to solve the above points,
A plurality of candidate strings are extracted, and for example, a word string stringed with a single M that has meaning in Japanese is extracted.
tlt! The purpose is to improve the IlJ rate. Therefore, the audio g# system of the present invention analyzes the coughing audio input signal based on the audio input signal and outputs the input audio! In a speech recognition system having a speech analysis device, the speech analysis device extracts a plurality of candidate sound carvings for each syllable from the input speech in the order of pronunciation, and attaches similarity information corresponding to each candidate syllable. a candidate string generation device configured to select and combine one of the candidate syllables for each syllable, and a candidate string analysis device; A word string buffer for storing word strings is provided, and the candidate string analysis device indexes a word dictionary having at least grammatical information, frequency information, and word information based on the input candidate strings. , is configured to convert the input candidate string into a meaningful word string while indexing the grammatical connection matrix for the extracted words based on the grammatical information, and converts the input speech into a word string. The system is characterized in that it outputs the word strings and gives priority to each word string. This will be explained below with reference to the drawings.

FIG. 1 is an overall configuration diagram of an embodiment of the present invention, FIG. 2 is an embodiment of the configuration of the candidate string creation device shown in FIG. 1, and FIG. 3 is an embodiment of the candidate string analysis device shown in FIG. 1. Show the configuration.

In FIG. 1, 1 is a microphone, 2 is a voice analysis device which has a conventionally known structure, for example, to perform frequency analysis to extract a feature parameter and check it with a feature dictionary, and 3 is a voice analysis device 2. A syllable candidate buffer 4 stores candidate monosyllables for each monosyllable that has been analyzed, and 4 is a candidate string creation device that stores candidate monosyllables stored in the syllable candidate buffer 3 for each monosyllable of input speech. 5 is a candidate string creation device that selects the candidate string and creates a candidate string; 5 is a candidate string analysis device that extracts a meaningful word string based on the content of the candidate string; , 7 is a word string] (a buffer for storing the extracted word string, 8 is an output display device control unit, and 9 is a display.

As the speech analysis device 2 is conventionally known, a detailed explanation will be omitted, but assuming that the input speech is composed of, for example, seven syllables, the speech analysis device 2 analyzes candidate word phrases for each syllable and generates similarity information. It is made to be attached and extracted. In the case shown,
Regarding the first syllable, candidate monosyllables "te", "ke", "he"
, "technical" force I is extracted, and the similarity information for each candidate monosyllable is "1" r2J, r! It is extracted as 5JIr9J. Regarding the second syllable t, the candidate monosyllables "i", "pi", "ki", "hi" are extracted and the respective similarity information "o", r 3 J +, r 4 J, r
5J is extracted.

The candidate sequence creation device 4 will be explained in detail with reference to FIG.
A combination is created by selecting one candidate single sound for each syllable/phrase of the input speech, and the candidate strings are arranged in descending order of the total value of the above similarity information: "Teiyausarel", "Keyausarel", "Teiyapusarel". The candidate sequence analysis device 6 will be described in detail with reference to FIG.
For each of the above candidate strings, we check the connections between meaningful words and store the results in a word string. In this case, priority is assigned to the obtained word strings based on the extraction order of the candidate strings and frequency information about words in a word dictionary, which will be described later.

The output display device control unit 8 causes the display 9 to display the word string in the form of letters and turns. At this time, the images are displayed in order according to the priority order.

FIG. 2 shows the structure of an embodiment of the candidate string creation device shown in FIG. 1. Reference numeral 5 in FIG. 1 corresponds to that in FIG. 1, and 10-1 to 10-7 are candidate monosyllable groups, respectively.

Reference numerals 11-1 to 11-7 represent monosyllable reading sections, 12 a candidate string generation register, and 13 a similarity addition section.

The monosyllable reading units 11 - 1 to 11 - 7 read one from each candidate monosyllable group and supply it to the candidate string generation register 12 . At this time, for example, each group l○-1 to 10-
Read each first monosyllable from 7 and take out "teiyausarel", then take out "keyausalel" with only group 1O-1 as the second monosyllable, then return only $10-1 to the first monosyllable. By making only 10-2 the second monosyllable and extracting "te♂yausarel", it is thought that the total value of similarity will be smaller. , and create all the unions one after another. The candidate sequences generated in the candidate sequence generation register 12 have similarity information added thereto by a similarity adder 13, and candidate sequences with smaller values are sequentially stored in the candidate sequence buffer 5.

FIG. 3 shows the configuration of an embodiment of the candidate sequence analysis device shown in FIG. Reference numerals 6 and 7 in the figure correspond to those in FIG. 1, G is a gate, 14 is an input register, 1δ is an analysis processing unit, 16 is a word dictionary, 17 is a grammar connection book search tool, and 20 is! 1ll
jIA check processing section 21 represents a word string checking section.

One candidate string is set in the input register 14 from the candidate string buffer 5 shown in FIG. 1, and processed by the analysis processing section 15. In the case shown in the figure, since the candidate string is "te", "i", "ya", etc. in order from the beginning, the word dictionary 16 is indexed via the dictionary search device 19, and the longest matching Words are sequentially read out from the word dictionary 16 based on the maximum evaluation method, which is a combination of the modulus and the frequency of occurrence of the word.

Note that the word dictionary 16 stores readings 22, word information 23, grammatical information 24, meaning information 25, frequency information 26, etc. as l#l.

In the case of illustrations, for example, "body", "bottom", "hand", etc.
*The staff is read out, and the grammatical information about each word is read out as well. Next, the word “body” and “bottom” correspond to “body” and “bottom”.
``ya'' is read out, and ``ya'' and ``i'' are read out in correspondence with ``hand''. Then, a grammatical connection check and a semantic connection check are performed via the connection check processing section 20.

What is illustrated in the analysis processing unit 15 shown in FIG. 3 represents the mode of connection between words, and "body" - "ya" indicates that connection is possible, "hand" - + ""No" indicates that connection is not possible.

The IJ box 17 can be thought of as a matrix in which grammatical information is arranged on the vertical and horizontal axes.
Words with grammatical information ``1'' are not grammatically connected to words with grammatical information ``53'' (logical ``O''), and <1) words with grammatical information ``3'' are grammatically connected to words with grammatical information ``53''. (Logic "1"), (lit) Grammar information "53"
A word with ``44'' can be connected to a word with grammatical information ``44'', indicating . Therefore, for example, from the dictionary 16, (1) the word "adjective" with grammatical information "3" is read out, (■) and then the word "adjective" with grammatical information "53" is read out.
” is read out, and then the single fj “ru” with (-) and grammatical information “44” is read out,
As a result of indexing the grammar connection matrix 17, "adjective" is obtained as one grammar connection possible. Note that the aspect of grammar checking using the grammar connection matrix 17 is disclosed in Japanese Patent Application No. 54-92270.

In addition, semantic conjunction fiff) + 3x 1B can be thought of as a matrix in which semantic information is arranged on the vertical and horizontal axes, in the same form as the grammatical connection matrix 1f, and there are correspondences between words that can be semantically connected. Logic "1" is written on the intersection point.

The semantic check using the semantic connection matrix is disclosed in Japanese Patent Application No. 102636/1982, and a detailed explanation thereof will be omitted.

As described above, the analysis processing unit 15 shown in FIG. Perform processing to form a certain word string. When the processing for one candidate string is completed, the game controller G is controlled to set the next candidate string in the input register 14, and the obtained meaningful word string is stored in the word string buffer 7.

At this time, the obtained word string is added according to its priority order. The configuration of the analysis processing section 15 is disclosed in Japanese Patent Application No. 1983-92271.

Candidate monosyllable groups 10-1 to 10-7 as shown in FIG.
If this is obtained, the resultant K, such as ``described'' or ``used in conjunction with'', etc., will be set in the word string BUTSUFF. .

As explained above, according to the present invention, the analysis result by the speech analysis device can be checked grammatically and semantically if necessary, and extracted as a meaningful word string. The recognition rate is greatly improved compared to the recognition rate obtained only by using this method.

[Brief explanation of drawings]

FIG. 1 is an overall configuration diagram of an embodiment of the present invention, FIG. 2 is a configuration diagram of an embodiment of the candidate string creation device shown in FIG. 1, and FIG. 3 is an embodiment of the candidate string analysis device shown in FIG. 1. Show the configuration. In the figure, 2 is a speech analysis device, 4 is a candidate string creation device, 6 is a candidate string analysis device, 7 is a word string buffer, 15 is an analysis processing unit, 16 is a word dictionary, 1f is a grammar connection matrix, and 18 is a meaning A connection matrix, 19 a dictionary search device, and 20 a connection check processing unit. Patent issuer Fujitsu Limited

Claims (1)

[Claims] 5. In a speech recognition system having a speech analysis device, the speech analysis device analyzes the input speech signal based on the speech input signal and recognizes the input mourning speech. The system is configured to extract a plurality of candidate syllables and provide similarity information corresponding to each candidate syllable, and create a candidate string by selecting and combining one of the candidate syllables for each syllable. A candidate string creation device and a word string buffer for storing word strings extracted by the candidate string analysis device and the cough candidate string analysis device are provided, and the candidate string analysis device is configured to perform processing based on the input candidate string. Based on this, while indexing a word dictionary having at least grammatical information, S* information, and word information, and indexing a grammatical connection matrix for the extracted words based on the grammatical information, the input candidate string is A speech recognition system configured to convert the input speech into a word string having a meaning, converting the input speech into a word string and outputting the same, and assigning a priority to each word string. .