JPH04128899A - Voice recognition device - Google Patents

Abstract

PURPOSE:To speedily and accurately recognize a correct word by taking accent into consideration for rendering data obtained from a voice. CONSTITUTION:An AD conversion part 31 converts the voice into a digital value and a feature extraction part 32 extracts the feature quantity of its voice waveform. A score calculation part 33 matches a template with the feature quantity of the input voice to calculate a matching score. A score order sorting part 34 classifies candidates according to their scores, a priority order sorting part 41 reclassifies and displays the candidates at a rendering display part 42, and a rendering determination 43 perform determination. Simultaneously, an access information extraction part 51 extracts the accent information of the input voice and collates it with accent information stored in an accent information matching part 52. A display control part 53 performs control so that candidates having large consistency of accent information are displayed preferentially on a candidate display part 6. A candidate selection part 7 outputs a word selected by a user.

Description

[Detailed Description of the Invention] [Summary] Regarding a speech recognition device that recognizes words expressed by speech from speech, the present invention is capable of quickly and accurately recognizing correct words by considering accents in reading data obtained from speech. A voice dictionary that stores voice features; a voice recognition unit that identifies human-generated voices and outputs candidates corresponding to the voice by referring to the voice dictionary; a pronunciation determining section that determines the pronunciation; an accent dictionary that stores features of the accents of voices; An accent identification section that refers to an accent dictionary and outputs a word that corresponds to this reading accent; a candidate display section that displays the words output by this accent identification section; and a candidate display section that displays words from among the words displayed in this candidate display section. and a candidate selection unit that selects a word representing the input voice.

[Industrial application field]

The present invention relates to a speech recognition device that recognizes words expressed by speech from speech.

Daigotama's voice recognition system is required to create documents by voice and search databases and dictionaries.

In the current Daigo speech recognition system, the candidate with the highest recognition score is unlikely to be the correct answer. Since there are many such words, even if the recognition score is high, the words often differ from the words actually uttered. For this reason, a method is used in which multiple candidates are displayed on a display and the correct answer is selected from among them.

[Conventional technology]

The principle of a conventionally used speech recognition system will be explained with reference to FIG.

In the figure, when audio is input, the score calculation unit 10
1, the input speech is matched with the template of each candidate word stored in the dictionary 104, the score of the candidate word is calculated, and the candidate display unit 102 displays candidates that are higher than a certain standard based on this calculation. are displayed in score order or dictionary order, and the candidate selection unit 103 outputs the recognition result as a word (C character string) representing the input voice based on the user's selection of the correct answer.

[Problem to be Solved by the Invention fl] As described above, the Daigotosei speech recognition system employs a method in which a plurality of candidates are displayed on the display and the user selects the correct answer from among them. Methods for displaying multiple candidates on the display include displaying them in dictionary order, and displaying candidates with a high probability of being correct by changing their brightness and color. However, even if the user selects pronunciation candidates using an interface that makes it easy for the user to select candidates, since there are many homonyms in Japanese, it is difficult to quickly select words that include the desired kanji. This is often difficult. Note that the word in this specification also means a clause that includes a particle.

The present invention has been made in view of the above-mentioned problems, and an object of the present invention is to provide a speech recognition device that can quickly and accurately obtain the correct word by taking into account accents in reading data obtained from speech. do.

[Means to solve the problem]

Figure 1 is a diagram showing the principle of the present invention.The present invention adds accent data to conventional pronunciation data to limit the range of candidate words. a speech dictionary 1 which has accumulated the features of the speech dictionary 1; a speech identification section 3 which identifies the input speech and outputs a candidate corresponding to the speech by referring to the speech dictionary 1; A pronunciation determining unit 4 that determines the pronunciation, an accent dictionary 2 that stores accent characteristics of voices, and a pronunciation determining unit 4 that identifies the accent characteristics of the input voice.
an accent identification unit 5 which inputs the pronunciation of the input voice determined by and outputs a word corresponding to the accent of this pronunciation by referring to the accent dictionary 2; and displays the word outputted by the accent identification unit 5. Candidate display section 6;
The device includes a candidate selection section 7 that selects a word representing the input voice from among the words displayed on the candidate display section 6. Further, the accent identifying section 5 causes the candidate display section 6 to display words ordered according to the degree of matching with the accent type of the input speech. Also,
The accent identification unit 5 extracts the fundamental frequency of the input voice and estimates the accent type. Further, accent data corresponding to the dialect is stored in the accent dictionary 2, so that the accent identifying section 5 refers to the accent data corresponding to the dialect of the input voice. Further, the standard accent data is stored in the accent dictionary 2, and the conversion rules between the standard accent and each dialect accent are stored in the accent identification unit 5, and the standard accent data is converted in accordance with the dialect of the input speech and is referred to. do it like this. Further, the accent identifying section 5 determines the dialect of the voice inputting person based on the selection result of the candidate selecting section 7 for the first voice of the same voice inputting person. In addition, the accent identification section 5 stores accent information for each dialect and the frequency of use of this accent information for each dialect in the accent dictionary 2. Consider the frequency of use when selecting words that correspond to the words. Further, when the word selected by the candidate selection section 7 is different from the word output from the accent identification section 5 based on the dialect-specific accent information, the word selected by the candidate selection section 7 is different from the word output from the accent identification section 5 based on the dialect-specific accent information. The frequency of use of the accent information for each dialect is corrected. Further, when the accent type for the word selected by the candidate selection unit 7 and stored in the accent dictionary 2 is different from the accent type recognized by the accent identification unit 5 for this word, the accent The accent type for the word stored in the dictionary 2 is changed to the accent type recognized by the accent identifying section 5.

[For production]

When a voice is input, the voice recognition unit 3 identifies this voice,
Refer to the voice dictionary 1 to find candidates that correspond to this voice,
The reading determination unit 4 determines the reading of this candidate. On the other hand, the accent identification unit 5 identifies the characteristics of the accent of the input voice, and based on the pronunciation determined by the pronunciation determination unit 4, refers to the accent information of the word corresponding to the pronunciation stored in the accent dictionary 2. Output the word corresponding to the accent. The candidate display section 6 displays the words output by the accent identification section 5, and the candidate selection section 7 selects the correct word for the input speech from among the displayed words. In this way, candidates are limited in consideration of the accent, making it possible to make a quick selection.

At this time, the candidate display unit 6 displays the words output by the accent identification unit 5 in an order according to the degree of matching with the accent type of the input voice, thereby limiting the selection range and allowing quick selection. . Further, when the accent identification unit 5 extracts the fundamental frequency of the input voice and performs frequency analysis or the like to grasp the accent characteristics of the input voice, it is possible to accurately grasp the accent characteristics.

Furthermore, since the input speech is often in a dialect, by storing accent data corresponding to the dialect in an accent dictionary and referring to the accent data corresponding to the dialect, it becomes possible to accurately grasp the characteristics of the accent. In addition, the accent dictionary 2 stores standard accent data, and the accent identification unit 5 stores conversion rules between the standard accent and each dialect accent, and converts the standard accent according to the dialect of input speech. Be able to understand the characteristics of correct accents depending on the dialect.

Further, by determining the dialect of the input voice based on the selection result of the candidate selection unit for the first voice of the same voice input person, it is possible to make a correct determination depending on the dialect from now on.

In addition, the accent dictionary 2 stores dialect-specific accent information for each word and the frequency of use of this dialect-specific accent information by the accent identification unit 5, and preferentially uses the most frequently used accent information. This improves the accuracy rate of words output by the accent identification section 5. At this time, if the word of the accent information used by the accent identification section 5 is different from the word selected by the candidate selection section 7, the accent information used by the accent identification section 5 is deemed to be incorrect, and the frequency of its use is reduced. Increase the probability that the correct one will be used.

Further, for a word selected by the candidate selection unit 7, if the accent type for the word stored in the accent dictionary 2 is different from the accent type recognized by the accent identification unit 5 for this word, the accent type for the word is stored in the accent dictionary 2. By changing the accent type to the accent type recognized by the accent identification section 5, the contents of the accent dictionary 2 can be updated correctly according to usage.

〔Example〕

Embodiments of the present invention will be described below with reference to the drawings.

FIG. 3 is a block diagram showing the first embodiment of the present invention.
This embodiment includes a template storage section 10 that stores feature amounts for a predetermined voice, an accent information storage section 20 that stores accent information for a predetermined voice, an AD conversion section 31 that digitally converts the voice, and the digital value A feature extraction unit 32 analyzes the waveform of the voice and extracts the feature quantity of the voice, matches this feature quantity with the template stored in the template storage unit 10, and calculates the matching score of the candidate corresponding to the matched template. a score calculation unit 33 that performs matching, a score order sorting unit 34 that sorts the candidates selected in the order of the matching scores, a priority order sorting unit 41 that sorts the candidates classified by the score order sorting unit 34 based on predetermined criteria; A pronunciation display section 42 that displays candidates up to a predetermined rank among those classified by the priority sorting section 41, a pronunciation confirmation section 43 that determines the correct pronunciation from among the pronunciations displayed on the pronunciation display section 42, and an AD an accent information extraction section 51 that extracts accent information of input speech from the digital value output from the conversion section 31; and an accent information storage section 20 for the pronunciation determined by the pronunciation determination section 43;
an accent information matching unit 52 that matches the accent information stored in the input voice with accent information extracted from the input speech; and a display control unit that controls preferential display of candidates with a large matching accent information based on this accent information matching. 53, candidates instructed by the display control unit 53, a candidate display unit 6 that displays the candidates with priority, and a candidate selection unit 7 that selects the correct word from the candidates displayed on the candidate display unit 6. .

Next, the operation of this embodiment will be explained.

When audio is input, an AD converter 31 converts the audio as an analog value into a digital value, and a feature extractor 32 analyzes the audio waveform represented by the digital value and extracts a feature. The score calculation unit 33 matches the templates stored in the template storage unit 10 with the feature amounts of the input speech, and calculates matching scores for candidates corresponding to each template. The score order sorting unit 34 sorts candidates having scores up to a certain rank based on this matching score in order of score. The priority sorting unit 41 further reclassifies the above-mentioned candidates according to a predetermined criterion such as dictionary order. This classification result is displayed on the pronunciation display section 42, and the pronunciation determining section 43 determines the correct pronunciation from the displayed candidates.

Meanwhile, in parallel with this, the accent information extraction unit 51
Accent information of the input voice is extracted from the digital output of the D converter 31. The accent information collation unit 52 collates the accent information stored in the accent information storage unit 20 for the pronunciation determined by the pronunciation determination unit 43 with the accent information of the input voice. The display control section 53 receives the collation information from the accent information collation section and performs control to preferentially display candidates with a high degree of matching of accent information.

The candidate display unit 6 displays candidates with priority display according to instructions from the display control unit 53. The candidate selection unit 7 outputs a candidate word (character string) selected by the user from among the candidates displayed on the candidate display unit 6.

Next, a second embodiment will be explained using FIG. 4, FIG. 5, and FIG. 6.

FIG. 4 is a block diagram showing the configuration of the second embodiment.

This embodiment extracts the fundamental frequency of the input voice, and determines the type of accent of the input voice from this fundamental frequency. Correspondingly, the related blocks in FIG. 3 are changed as follows. First, the accent information storage unit 20 in FIG. 3 is changed to the accent type storage unit 201 that stores accent type data corresponding to the word that indicates the meaning expressed by the voice.
The accent information extraction section 51 is made into a fundamental frequency extraction section 511 that extracts a fundamental frequency from an input voice and an accent type determination section 512 that determines the accent type of the voice from the extracted fundamental frequency. Correspondingly, the accent information matching unit 52 is changed to an accent type matching unit 521 that matches the accent type stored in the accent type storage unit 201 with the accent type of the input voice for the determined pronunciation. Other than the above, it is the same as FIG. 3.

Next, the accent type, which is a feature of this embodiment, will be specifically explained.

As an example, if the uttered sound is "kaeru", there are many homophones. Therefore, even after it is determined that the reading is r, it takes a certain amount of time to select the desired word.

Therefore, by using information on the accent type of the word ``r kaeru'', we can display preferentially those that are close to the accent type of the current pronunciation from among the many homophones, so that you can quickly select the word you need. Make it. Homophones of r frog 1 are categorized by accent type as follows.

Flat plate type (frog) → (111): Change. change. frog,
exchange. Change head height type (frog) → (100): Return. I'm going home.

Sir. The character 'flj' indicates a strongly accented part, and '0' indicates a weakly accented part. Since the words are divided into two categories based on accent type, you can definitely select them faster than if they were simply displayed.

The accent type determination method estimates the accent type of the utterance from the function of the fundamental frequency. As an example, the fundamental frequency patterns of the utterance of "r frog" for flat type and high head type are shown in Figures 5 and 6, respectively. As is clear from this figure, the fundamental frequency of the flat type is flat, and it decreases slightly toward the end of the word. On the other hand, in the head-height type, the frequency is high from the end of the word to "al," and the frequency drops dramatically just before the word "re." As in this example, the accent type can be estimated by finding no change in the fundamental frequency or finding places where it changes suddenly.

Next, a third embodiment will be explained using FIGS. 7 and 8.

The third embodiment is the same as the second embodiment by providing an accent type storage section corresponding to the dialect, so that an accent corresponding to the dialect of the speaker is used. FIG. 7 shows a block diagram of this embodiment. The accent type storage unit 201 in FIG. 4 is provided with N accent type storage units for each dialect, and an accent type selection unit 54 is provided to extract the accent type corresponding to the dialect. The rest is the same as the second embodiment.

Figure 8 is a diagram showing an example of accent types depending on the dialect. 'IJ and 'OJ in Figure 8 indicate high and low accents, respectively. In this way, depending on the dialect,
Since the kanji corresponding to the same reading often differ, this embodiment makes it possible to select an accent type appropriate to the dialect and obtain the correct word for the voice (that is, a word using the correct kanji).

Next, a fourth embodiment will be explained using FIG. 9 and FIG. 1O.

In this embodiment, whereas the third embodiment has an accent type storage unit 201 for each dialect, the accent type storage unit 201 stores a standard accent type, and the standard accent type and the accent type of each dialect are stored in the accent type storage unit 201. The standard accent type is converted according to the dialect using the conversion rules that exist between .

This conversion is performed by the accent type conversion unit 55 shown in FIG. The rest is the same as the second embodiment shown in FIG.

FIG. 10 is a diagram showing an example of this accent type conversion rule. In other words, when the voice is in the Keihan dialect, r sell j
The accent type of the Tokyo dialect (if this is taken as the standard) is ``
011 should be converted to rllll.

Next, a fifth embodiment will be explained using FIG. 11.

In this embodiment, the dialect of the same voice inputter is estimated from the selection result of the correct word for the first voice of the same voice inputter, and thereafter data in the accent type storage unit 201 corresponding to the dialect is used. .

In FIG. 11, the dialect estimating unit 56 converts the voice actually uttered by the speaker into the accent type determined by the accent type determining unit 512 and the candidate actually selected by the candidate selecting unit 7 (the meaning of the pronunciation). If the accent types of the words assigned the kanji representing
1 to estimate the correct dialect, and instruct the accent type selection unit 54 to switch to the estimated dialect from next time onwards.

An example of estimation performed by the dialect estimating unit 56 will be described below.

For example, when a speaker of the Keihan dialect utters ``r return'', the accent type determining unit 512 determines the accent type as rlllj based on the fundamental frequency analysis result. In this case, change J with the closest accent from the standard accent dictionary.
is preferentially displayed on the candidate display section 6 as the first candidate.

If the word selected by the actual candidate selection unit 7 is not ``change 1'' but r return 1, search the accent type dictionaries of each dialect and search for dictionaries that have words with matching accent types,
Use that dictionary from the next human-powered voice. 11 is the same as the third embodiment shown in FIG. 7 except for the above.

Next, a sixth embodiment will be explained using FIG. 12 and FIG. 13. In this embodiment, the dialect-specific accent types stored in the accent type storage unit 201 are also stored together with their frequency of use, and the accent type collation unit 521 uses the frequently used accent types. If the accent type selected by the candidate selection section 7 is different from the accent type of the human voice, an accent type frequency correction section 57 corrects the usage frequency of the dialect used for the input speech. The contents of the accent type storage section 201 and the accent type frequency correction section 57 are the same as the third embodiment shown in FIG.

FIG. 13 is a diagram showing an example of accent type usage frequencies for each dialect stored in the accent type storage unit 201.

Next, a seventh embodiment will be explained using FIG. 14.

In this embodiment, when the accent type of the word selected by the candidate selection unit 7 and the accent type of the input voice determined by the accent type determination unit 512 are different, the accent type registration unit 58
The accent type of the selected word stored in the memory is replaced with the accent type uttered this time, and the data is updated according to the actual usage situation.

14 is the same as the second embodiment shown in FIG. 4 except for the above.

〔Effect of the invention〕

As is clear from the above explanation, the present invention determines the pronunciation from the input speech, takes accent information into consideration in this pronunciation, displays candidates in priority order, and allows the user to select the correct candidate. can be selected quickly.

[Brief Description of the Drawings] Fig. 1 is a diagram showing the principle of the present invention, Fig. 2 is a diagram showing an example of the principle of the conventional method, Fig. 3 is a block diagram showing the first embodiment of the invention, Fig. 4 is a block diagram showing the second embodiment, FIG.
FIG. 6 is a diagram showing an example of the fundamental frequency pattern of the input voice, FIG. 7 is a block diagram showing the third embodiment, FIG. 8 is a diagram showing an example of the accent type depending on the dialect, and FIG. FIG. 10 is a block diagram showing an example of dialect-side accent rules. FIG. 11 is a block diagram showing the fifth embodiment. FIG. 12 is a block diagram showing the sixth embodiment.
The figure shows an example of the frequency of accent type usage in dialects.
FIG. 4 is a block diagram showing the seventh embodiment. In the figure, 1-speech dictionary, 2-accent dictionary, 3-speech identification section, 4-reading confirmation section, 5-accent identification section, 6-
・-Candidate display section, 7-Candidate selection section, 10-Template storage section, 20-Accent information storage section, 3l-AD
conversion unit, 32-feature extraction unit, 33-score needle military unit,
34-Score order sorting unit, 41-Priority sorting unit, 42-Reading display unit, 43-Reading confirmation unit, 51-
accent information extraction unit, 52-accent information collation unit, 53-display control unit, 5
4-Accent type selection unit, 55-Accent type conversion unit, 56-Dialect estimation unit, 5
7-accent type frequency correction unit, 58-accent type registration unit, 201--accent type storage unit, 511-fundamental frequency extraction unit, 512-accent type determination unit, 521--accent type matching unit. An example of the principle diagram of the conventional method Figure 2 Reading Kanji Tokyo dialect Tokyo dialect frequency Keihan dialect
Keihan dialect frequency change Return 100 70
111 20 Frog Change 011
40 100 60 Frequency of use of accent type by dialect Figure 13 Voice string @l Example Figure 3 Voice 2 Example Figure 4 Tokyo dialect Rereading Kaeru Kanji Return Changing accent type +00 Keihan dialect Kaeru Returning Kaeru Depending on the dialect to change Accent type Figure 8 Reading Kanji Tokyo dialect Keihan dialect Accent rules by dialect Figure 10

Claims (1)

[Scope of Claims] 1) A speech dictionary (1) that stores features of speech, and a speech identification unit that identifies input speech and refers to the speech dictionary (1) to output candidates corresponding to this speech. (3), a pronunciation determination unit (4) that determines the pronunciation of the candidate output from the speech identification unit (2), an accent dictionary (2) that stores characteristics of the accent of the voice, and an accent of the input voice. an accent identification unit (which inputs the pronunciation of the input voice determined by the pronunciation determination unit (4) and outputs a word corresponding to the accent of this pronunciation by referring to the accent dictionary (2); 5) and this accent recognition part (5
), and a candidate selection section (7) for selecting a word representing the input voice from among the words displayed on the candidate display section (6). A voice recognition device characterized by: 2) The accent identification unit (5) causes the candidate display unit (6) to display words ordered according to the degree of matching with the accent type of the input speech. Speech recognition device. 3) The speech recognition device according to claim 1 or 2, wherein the accent identification unit (5) extracts the fundamental frequency of the input speech and estimates the accent type. 4) Accent data according to the dialect of the input voice is stored in the accent dictionary (2), and the accent identification section (5) refers to the accent data according to the dialect of the input voice. 3. The speech recognition device according to any one of 3 to 3. 5) Accumulate standard accent data in the accent dictionary (2), store conversion rules between the standard accent and each dialect accent in the accent identification unit (5), and store the standard accent data in accordance with the dialect of input speech. The speech recognition device according to any one of claims 1 to 3, characterized in that the speech recognition device converts and references the speech recognition device. 6) A claim characterized in that the accent identification unit (5) determines the dialect of the voice input person from the selection result of the candidate selection unit (7) for the first voice of the same voice input person. The speech recognition device according to any one of Items 4 and 5. 7) The accent dictionary (2) stores dialect-specific accent information for each word and the frequency of use of this dialect-specific accent information by the accent identification unit (5). 5. The speech recognition device according to claim 1, wherein when selecting a word corresponding to the input speech, the speech recognition device also takes into consideration the frequency of use of the word. 8) The word selected by the candidate selection unit (7) is determined by the accent identification unit (7) based on the dialect-specific accent information.
5) If the output word differs from the output word, the frequency of use of the dialect-specific accent information used to determine the output word is corrected.
The voice recognition device described. 9) For the word selected by the candidate selection unit (7), the accent type for the word stored in the accent dictionary (2) is the same as the accent type recognized by the accent identification unit (5) for this word. If the accent type is different, the accent type for the word stored in the accent dictionary (2) is changed to the accent type recognized by the accent identification unit (5).
3. The speech recognition device according to any one of 3 to 3.