JPH0261700A - Speech recognition device - Google Patents

Abstract

PURPOSE:To prepare correct phoneme candidates by adding an another candidate to candidates by phonemes which are selected by a speech recognition part while referring a rule in a rule dictionary storage means. CONSTITUTION:A standard pattern storage means 3 is stored with standard patterns which are generated by phonemes and a speech is inputted by an input means 1; and a speech recognition part 2 compares an input speech with standard patterns and calculates the extent of matching with the standard patterns to select candidates by phonemes. Then a phoneme processing part 4 while referring to the rule dictionary storage means 5 where recognition rules derived from past speech recognition experience adds a candidate to the phoneme candidate by the phonemes which are selected by the speech recognition part 2. Thus, the phoneme candidates are increased by using a rule dictionary 5, so even if an error occurs in a speech recognizing process, a correct phoneme candidate can be prepared.

Description

DETAILED DESCRIPTION OF THE INVENTION [Field of Industrial Application] The present invention relates to a speech recognition device used in speech word processors and the like.

[Prior Art] When recognizing each phoneme in a speech recognition device, it is difficult to simply connect the first candidates of the recognition results for each phoneme.
It is unlikely that words can be recognized correctly. because,
For example, even if the recognition rate for each phoneme is 95%, the recognition rate for a syllable calculated from it is (0,95) 2 = 0.
90, and if 4 syllables were generated, the overall recognition rate would be ((0,95)")' = 0.66, which would be quite low.

Therefore, conventional methods have been used to increase the accuracy rate by outputting multiple candidates for each phoneme instead of just one.

[Problem to be solved by the invention] However, if multiple candidates for each phoneme are output in this way, the number of combinations of each candidate will increase to an enormous number, and the subsequent kana-kanji conversion. When used for processing such as, there is a problem that the processing time is actually too long.

Furthermore, in the case of continuous speech, even if multiple candidates are given for each phoneme due to effects such as articulatory combination, there may not be a correct answer among them, and each phoneme itself may not be correctly extracted. Therefore, there is also the problem that accurate kana-kanji conversion cannot be done as is.

An object of the present invention is to provide a speech recognition device that solves the problems of conventional speech recognition devices.

[Means for Solving the Problems] The present invention according to claim 1 provides an input means for inputting speech, a standard pattern storage means for storing a standard pattern created for each phoneme, and a method for inputting speech input from the input means. and a standard pattern stored in the standard pattern storage means, a speech recognition unit that calculates the degree of matching with the standard pattern, and selects candidates for each phoneme, and a recognition rule derived from past speech recognition experience. A device comprising: a collected rule dictionary storage means; and a phoneme processing section that adds other candidates to the candidates for each phoneme selected by the speech recognition section while referring to the rules in the rule dictionary storage means. It is.

The present invention according to claim 2 provides an input means for inputting a voice, a standard pattern storage means for storing a standard pattern created for each phoneme, and a method for storing the voice input from the input means and the standard pattern storage means. A speech recognition unit that compares the selected standard pattern with the standard pattern, calculates the degree of matching with the standard pattern, and selects candidates for each phoneme, and combines the candidates selected by the speech recognition unit, A clustering unit that compares Mao words in a word dictionary prepared in advance and outputs the corresponding word if there is a corresponding word in the word dictionary, and a matching degree for at least each phoneme calculated by a speech recognition unit. and a word DP matching unit that selects a more appropriate word from the candidate words selected by the clustering unit based on the above.

[Operation] The present invention according to claim 1 stores a standard pattern created in advance for each phoneme in a standard pattern storage means, inputs a voice through an input means, and combines the input voice and the standard pattern with a voice recognition unit. The phoneme processor selects candidates for each phoneme by calculating the degree of matching with the standard pattern, and uses the phoneme processing unit to refer to a rule dictionary storage means that collects recognition rules derived from past speech recognition experience. The present invention according to claim 2, wherein the candidates for each phoneme selected by the speech recognition section further include other candidates, the standard pattern storage means stores a standard pattern created in advance for each phoneme, and the input means The voice recognition unit compares the input voice with the standard pattern to calculate the degree of matching with the standard pattern, selects candidates for each phoneme, and the clustering unit selects candidates for each phoneme. The candidates are combined, and the combined various words are compared with words in a word dictionary prepared in advance. If they match, the word is output, and the word DP matching unit outputs the word and the speech recognition unit Based on the calculated degree of matching for each phoneme, a more appropriate word is selected from the candidate words selected by the clustering unit.

[Examples] The present invention will be described below based on drawings showing examples thereof.

FIG. 1 is a block diagram showing an embodiment of a speech recognition device according to the present invention.

The standard pattern storage means 3 is a means such as a ROM (read-only memory) that stores standard pattern waveforms of various phonemes.

The speech recognition unit 2 compares and matches the speech input from the microphone of the input means 1 with the standard pattern of the standard pattern storage means 3, calculates the degree of matching (distance) with the standard pattern for each phoneme, This is a means of selecting phoneme candidates with a high degree of matching.

For example, the candidates are up to 4th place for consonants, and up to 1st or 2nd place for vowels. Furthermore, for vowels, a flag indicating ambiguity is also attached to the candidates.

The rule dictionary storage means 5 is a ROM that stores rules obtained from past analysis experience in speech recognition.

The phoneme processing unit 4 is a means for adding other candidates to the candidates for each phoneme selected by the speech recognition unit 2, as necessary, while referring to various rules in the rule dictionary storage unit 5.

The clustering section 6 combines the candidates for each phoneme selected by the phoneme processing section 4 with each other. Furthermore, it is checked whether each of the combined phoneme candidate strings (words) corresponds to a word in the word dictionary 11 (stored in the word dictionary storage means 7) extracted from the Kana-Kanji conversion dictionary, which will be described later. If there is a word to be used, this means outputs that word as a phoneme candidate string.

The word-to-word DP unit 8 determines the state of the ambiguity flag attached to the vowel candidate and the matching degree (distance) of the standard pattern.
This means calculates the degree of matching of the candidate words output from the clustering unit using the above. Then, words determined to have a high degree of matching are passed to the next kana-kanji converter 9.

The kana-kanji conversion unit 9 uses the kana-kanji conversion dictionary to perform kana-kanji conversion on the kana sent from the word-to-word DP unit 8, and calculates the validity (cost) of the output result for the input voice. This is a means of outputting those that are judged to have high quality.

Next, the operation of the embodiment 1.L will be explained by taking as an example the case where the phrase "Kobano" is input by voice.

FIG. 2 shows what is output from the speech recognition section 2 when the phrase "Kobano" is input by voice from the input means 1. The “Kou” part of “Kobano” is actually r
Since the ko-J and rOJ vowels are pronounced with a long pronunciation, the continuous vowel flag "4" is attached. That is, the meaning of the ambiguity flag "4" is that there is a possibility of continuous vowels, as shown in FIG. Also, r in the “no” part
The nJ consonant is sometimes misrecognized as "n", so
In this case, the outside of the l syllable of “no” is 2 as in [zero Ngo].
It is recognized in syllables. However, this [zero N
] is marked with an ambiguity flag "2". Its meaning is that there is a possibility of addition.

The meaning of the notation of phoneme candidates shown in FIG. 2 is as follows.

That is, the phoneme candidates basically conform to the notation of each consonant and vowel when Japanese is displayed in Roman letters. however, "
"N" is indicated by a capital letter "N", and "A" and "A" are indicated by a capital letter "N".
``木'' is written in the place corresponding to the consonant part of ``n''.

The phoneme processing unit 4 receives candidates for each phoneme as shown in FIG. 2 and processes them based on the rule dictionary.

FIG. 4 shows the results. Based on past speech analysis experience, rOJ sounds with continuous vowel flags have ro
*u" and "0 0 0" are added, and considering the rJJ] sound with the addition flag and the next vowel candidate, for the part that is judged to be two syllables in speech recognition, r m o ], rnOJ are added as candidates outside the l syllable. In addition, when consonants are uttered continuously,
Since it is easy to change due to the influence of the preceding and following vowels, and there may be cases where the correct answer is not among the four candidates, consonant candidates are also added as shown in the figure.

Comparing the phoneme candidates in FIG. 2 output by the speech recognition unit 2 and the phoneme candidates in FIG. 4 output by the phoneme processing unit 4, the candidates in FIG. I cannot get the correct kana string for the input "Kobano". However, depending on the combination of the candidates in Figure 4, ``Kobano'' may be obtained.

FIG. 5 shows that the clustering unit 6 creates all possible combinations of phoneme candidates output from the phoneme processing unit 4 and searches for those included in the word dictionary. In the figure, the numbers written to the left of the extracted words indicate the way they are combined, and the words are extracted from the same syllable to the same syllable of the result obtained by the speech recognition unit 2. It is meant to show that something is true. In this embodiment, since voice input is performed for each phrase, the number of words in the input voice is not constant, and it is not possible to limit whether words start from the same syllable within a phrase. For this reason, the clustering unit 6 extracts words without limiting the number of words in a clause, the number of characters in each word, and the position of each word.

The numbers shown on the left side are used in the word-to-word DP section 8 and the kana-kanji conversion section 9.

FIG. 6 illustrates the DPP matching in the word-to-word DP section 8, taking as an example the word "Koba" among the various words shown in FIG. 5 output as candidates from the clustering section 6. This is what I did.

“Koba” is the 1st and 2nd syllable of the recognition result (1 in Figure 5).
Since this is one m-word candidate that has come out of the part corresponding to (see -2), the matching degree of this part is used as the matching degree used in DPP matching. That is, the degree of matching (distance) between each phoneme and each standard pattern passed from the speech recognition unit 2 is entered into each blank column of the DP matching child table shown in FIG. 6, and DPP matching is performed. Similar DPP
Matching, No. , another two-syllable word candidate "kobo",
``Kona'', ``Koma'', etc. will also be explained.

Then, the obtained word candidate with the highest degree of matching is taken as a representative for that 1-2 syllable. In this case, it would be "Koba".

In this way, 11 syllables, 1-2 syllables, 13 syllables,
... Select up to 5 representative lucid word candidates for each of 2-2 syllables, 2-3 syllables, etc.

Note that if the vowel has an ambiguity flag, the DP
In some cases, the above-mentioned matching degree is not entered in each blank column of the child table for matching, and instead, a special value is entered. - Figures 7 and 8 are flowcharts showing the flow of the special processing.

FIG. 7 is a flowchart for determining the barking status of the ambiguous flag positions of the dropout flag, addition flag, and continuous vowel flag.

For example, the processing of the omission flag rlJ will be explained as an example.

In the first place, the omission flag "1" is added in the following cases.

As shown in FIG. 9(a), it is assumed that, for example, a voice waveform when uttering "fusoku" is input from the microphone 1.

In the speech recognition section 2, as shown in FIG. 9(b), first a speech section (A) is detected, then a vowel section (B) is detected, and the vowel is recognized. Assume that for the input "Fusoku", "U' of "Fu" could not be detected as a vowel.

Next, the remaining section of the vowel section (B) is set as a consonant (C),
Perform consonant recognition. In this way, vowel and consonant candidates are listed as shown in FIG. 9(C) (see FIG. 2). By the way, in the consonant section (C), the section marked with ☆ is too long as one consonant.

Therefore, since there may be a vowel in between, the omission flag "1" is turned on. Change the consonant in the dropped section to C1 vowel■
Indicated by Furthermore, since the consonants in this section are "h" in the first half and "s" in the second half, both of these may appear as consonant candidates.

In this way, regarding the speech recognition result to which the omission flag "1" has been added, the inter-word DP section 8 determines that the omission flag "l" is on (step Sl) as shown in FIG. ``The possibility that there is no character at the flag position'' is checked (step S2).

That is, when creating a kana character string to be a candidate, processing is performed both when there is a character at the character position of the omission flag and when there is no character. Therefore, the candidates for "Fusoku" are two-letter and three-letter ones. When the number of candidates is two characters, this means that the processing is performed with the determination that there is no character at the character position of the omission flag, so the determination is YES and 1 is assigned to chk (step S3).

"The consonant of the next letter is for the consonant in this position."
In (step S4), there is a possibility that the correct recognition result 'h' of the first character of "Fusoku" appears in the next character of the omission flag. In such a case, the consonant of the first letter of the word candidate obtained by dictionary lookup is ranked high among the speech recognition results of the consonant of the second letter. At this time, this judgment is YES
S (step S4), and 2 is assigned to ch k (step S5).

"The position of the flag may be a consonant" (step S6), since the consonant is a silent section and there is a fairly long time interval between vowels, a dropout flag may be attached (
Consonants have their own flags, but they may be omitted and judged). This fact is also registered in the rule dictionary, so "
Words that contain "tsu" are also suggested. The number of consonants after the small “tsu” is limited (1 (, SS t, p)
Therefore, if these consonants are candidates for the next character with the omission flag, and the character corresponding to the flag position is a small "tsu", the determination is YES (step S6).
3 is assigned to chk (step S7).

In this way, the positional status of the additional flag and continuous vowel flag is also determined, and the result is substituted into chk.

FIG. 8 shows how each column of the DP matching table is corrected using chk, which is the result of determining the state of the flag position in this manner.

In the same figure, since the dropping flag rlJ is on (step St), next in steps 82 to S5, c
The contents of hk are checked, and the DP matching table is corrected accordingly (steps 86 to S
9).

In step S6, child and mother mean the consonant and vowel of the word from the dictionary, and C9■ is the consonant and vowel at the position of the omission flag. C and V are recognition results without flags, C is a consonant, and ■ is a vowel. 1.010
．． 5 is the matching degree value entered by this process. A single value of 1.0 or 0.5 is entered.

In step S7, C-Child 1 is for Child 1 among the recognition results of the character next to the omission flag. C-child 2 is
Since there is no recognition result regarding the consonant dropout position for child 2 in the consonant score for the dropout flag, the number (lσ
(prepare both consonants and vowels).

In step S8, the child and mother are small "tsu".

In step S9, (c) and (v) are DP
Consonants and vowels in positions that are not directly involved in matching, (
c) -Child is for the child in the result of the next character's recognition. The thick line indicates the end of the matrix during actual DP matching (when the word candidate is for the middle of the utterance).

In step SIO, (C) and (V) are DP
Consonants and vowels with dropout flags in positions that are not directly involved in matching, (C)-Child are for children in the score of consonants for dropout flags.

The thick line is the starting point during actual DP matching (when the word candidate is from the middle of the utterance).

In step Sll, "V attached" means a vowel candidate with an additional flag on, and 1.3☆ means that after putting 1.3 into the DP matching table, correction is performed based on the recognition results of the consonants before and after this. be.

In step S13, c t and v l are recognition results without flags.

In step S14, ``2'' is a vowel candidate for the continuous vowel flag on, and vrnb is a value for the vowel from among the numerical values prepared for continuous vowels.

Child 2 and mother 9 are marked with "2" to distinguish them from the older child and mother.

In this way, the inter-word DP unit 8 performs DP matching processing, and each syllable part (1st syllable, l syllable mouth 42
Word groups with a high degree of matching are passed to the kana-kanji transfer unit 9 at the syllable, the first syllable to the third syllable, the second syllable, the second to third syllable, and so on.

The kana-kanji conversion unit 9 performs kana-kanji conversion using a kana-kanji conversion dictionary, and furthermore, the clustering unit 6 assembles phrases using numbers attached to word candidates, and the word-to-word DP matching unit 8 performs matching. This system uses language knowledge and linguistic knowledge to rank and output phrase candidates.

FIG. 10 shows the results.

Note that the rule dictionary 5 is created taking into account ambiguity flags during vowel recognition, but if there are no flags of this kind in the speech recognition results, a rule dictionary may be created based on the tendency of recognition errors. can.

Further, in calculating the degree of matching in the word-to-word DP matching section 8, it is needless to say that other calculation methods can be used instead of the method described above.

[Effects of the Invention] As is clear from the above description, the present invention according to claim 1 uses a rule dictionary to reinforce phoneme candidates.
Even if an error occurs during the speech recognition process, correct phoneme candidates can be prepared.

Further, the present invention according to claim 2 provides a clustering unit, an inter-word D
Since the P-matching unit 8 can narrow down the number of candidates, it has the advantage that, for example, the burden of subsequent kana-kanji conversion processing can be reduced.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of the speech recognition device according to the present invention, FIG. 2 is a block diagram showing an example of the output of the speech recognition section of the embodiment, and FIG. 3 is a block diagram showing an example of the output of the speech recognition unit of the embodiment. FIG. 4 is a configuration diagram showing the processing results of the phoneme processing unit in the same embodiment. FIG. 5 is a configuration diagram showing the processing results by the clustering unit in the same embodiment. , FIG. 6 is a block diagram showing the state of processing in the inter-word DP matching section in the same embodiment, and FIGS. 7 and 8 show processing using ambiguity flags in the inter-word DP matching section in the same embodiment. Flowchart, No. 9
Figures (a), (b), and (c) are block diagrams for explaining the omission flag in the same embodiment, and Fig. 10 is a block diagram showing the processing results of the kana-kanji converter of the same embodiment. 1... Input means 3... Standard pattern storage means 2... Speech recognition section 4... Phoneme processing section 5... Rule dictionary storage means 6... Clustering section 8... Word DP matching section Applicant: Brother Industries, Ltd.

Claims (2)

[Claims] (1) An input means for inputting speech, a standard pattern storage means for storing a standard pattern created for each phoneme, and a speech input from the input means and the standard pattern stored in the standard pattern storage means. A speech recognition unit that compares and calculates the degree of matching with a standard pattern and selects candidates for each phoneme, a rule dictionary storage unit that collects recognition rules derived from past speech recognition experience, and a rule dictionary storage unit that collects recognition rules derived from past speech recognition experience. A speech recognition device comprising: a phoneme processing unit that adds other candidates to the candidates for each phoneme selected by the speech recognition unit while referring to rules. (2) An input means for inputting speech, a standard pattern storage means for storing a standard pattern created for each phoneme, and a method for inputting the speech input from the input means and the standard pattern stored in the standard pattern storage means. A speech recognition unit that compares, calculates the degree of matching with a standard pattern, and selects candidates for each phoneme, and combines the candidates selected by the speech recognition unit, and combines various words with pre-prepared words. a clustering unit that compares the words with the words in the dictionary and outputs the corresponding word if there is a corresponding word in the word dictionary; A speech recognition device comprising: a word DP matching unit that selects a more appropriate word from the selected candidate words.