JP3958908B2 - Transcription text automatic generation device, speech recognition device, and recording medium - Google Patents

Description

[0001]
BACKGROUND OF THE INVENTION
The present invention relates to a transcription text automatic generation device that automatically generates a transcription text used for learning an acoustic model, a speech recognition device that performs speech recognition using an acoustic model learned from speech data and corresponding transcription text. The present invention also relates to a recording medium on which a program for automatically generating a transcription text is recorded.
[0002]
[Prior art]
The acoustic model used in the speech recognition system is a statistical model for calculating how much the input speech seems to be a candidate phoneme (see NHK STRL R & D No. 57, 19999.8pp36). Learned from the voice data and its transcription text. The creation of this Kana-Kanji mixed transcription data and phonetic transcription text data relies on human hands. Conventionally, there is a limit to the amount of work performed by the person in charge of transcription, so it is difficult to obtain phonetic transcription text with delimiters inserted for each recognition unit, and an acoustic model is created without inserting delimiters. There were many cases.
[0003]
Moreover, even if transcription text data with a partial delimiter inserted is available, the same recognition unit used in the speech recognition system to be applied is not always used in the transcription text. Transcribed text has been used with inconsistencies between the data and the data handled by the recognition system.
[0004]
[Problems to be solved by the invention]
When constructing a speech recognition system, unifying the speech recognition unit handled by the speech recognition decoder, the speech recognition unit handled by the language model, and the speech recognition unit handled when creating the acoustic model is to ensure recognition accuracy. It is essential.
[0005]
In the conventional system, for the reason described above, a situation arises in which the phonetic transcription with no delimiter inserted at the optimum position must be written.
[0006]
Here, an example is a speech recognition system in which a morpheme that is generally used is a recognition unit, three consecutive phonemes in the morpheme are taken into consideration, and three phonemes across the morpheme are not taken into account.
[0007]
That is, in this system, the acoustic model used between morphemes considers the first two phonemes in the morpheme that do not depend on subsequent phonemes.
[0008]
In order to learn the above model, in the prior art, an acoustic model using all three continuous phonemes in the phoneme transcription text and an acoustic model using all two continuous phonemes must be created and used separately. Don't be.
[0009]
Moreover, sufficient accuracy cannot be ensured when the acoustic model of two continuous phonemes does not take into account the condition of two continuous phonemes appearing between morphemes.
[0010]
For this reason, in order to ensure the accuracy of the acoustic model, it is necessary to learn the part that spans between morphemes as a model that considers only two phonemes, and the other part needs a model that considers three phonemes. In order to create this model, phoneme transcription text with morpheme delimiters inserted is necessary, and a method that does not require much human intervention is desired for this text creation.
[0011]
In addition to the above examples, even in a system that considers three phonemes that span morphemes, it is necessary to have a transcription text in which a delimiter is inserted at a position where two consecutive phonemes such as short silence and breathing portion must be considered. Then, it is the same as the above-mentioned.
[0012]
In view of the above, the present invention provides a transcription text automatic generation apparatus capable of automatically generating a transcription text for learning an acoustic model, a speech recognition apparatus equipped with the transcription text recording apparatus, and a recording medium. There is.
[0013]
[Means for Solving the Problems]
In order to achieve such an object, the invention of claim 1 is an input means for inputting kanji-kana mixed transcription text data and phoneme-notation transcription text data describing a phoneme notation corresponding to the kanji-kana mixed text data. When,
A first information processing unit that divides the input kanji-kana mixed text data into predetermined recognition units, a character string and its reading are described, and a reading that allows a plurality of readings for one character string is described. A character string in the divided kanji-kana mixed text data is stored on the basis of the reading dictionary while being matched with the division position of the text data divided by the first information processing means and the storage means storing the dictionary. If there are multiple readings for the same character string, the optimal candidate among the multiple readings is determined based on the phonetic transcription text data input from the input means, and the reading 2nd information processing means for generating phoneme notation text data into which a delimiter has been inserted by inserting a delimiter into the division position of the character string converted into And wherein the door.
[0014]
According to a second aspect of the present invention, in the transcription text automatic generation device according to the first aspect, the predetermined recognition unit is a morpheme, and the first information processing means is a morpheme analysis unit.
[0015]
The invention according to claim 3 is the transcription text automatic generation device according to claim 1, wherein the second information processing means performs DP matching when a plurality of readings exist for the same character string to be converted. The optimum candidate is determined by processing, a reference pattern with a delimiter added in the character string converted into the reading for the DP matching processing is created, and the phoneme notation transcription text data has a predetermined length. A character string is extracted to create a test pattern, and the reference pattern and the test pattern are subjected to DP matching processing.
[0016]
According to a fourth aspect of the present invention, there is provided input means for inputting phonetic transcription text data describing kanji mixed kana mixed text data and phoneme notation corresponding to the kanji mixed text data, and the input kanji mixed text A first information processing unit that divides data into predetermined recognition units; a storage unit that stores a reading dictionary in which a character string and a reading thereof are described and a plurality of readings are permitted for one character string; Based on the reading dictionary, the character string in the divided kanji-kana mixed text data is converted into a reading based on the reading dictionary while matching the dividing position of the text data divided by the first information processing means, and the same character If there are multiple readings for a sequence, the best candidate among the multiple readings is written in the phoneme notation written from the input means. In addition, it is determined based on the text data, and a delimiter is inserted by inserting a delimiter into the divided position of the phonetic transcription transcription data input from the input means in correspondence with the division position of the character string converted into the reading. And second information processing means for generating finished phoneme notation text data.
[0017]
According to a fifth aspect of the present invention, in the transcription text automatic generation device according to the fourth aspect, the second information processing means generates a reference pattern composed of a character string converted into the reading for the DP matching processing. Creating a test pattern by extracting a character string of a predetermined length from the phonetic transcription text data, DP matching processing between the reference pattern and the test pattern, on the test pattern, A delimiter symbol is inserted at the delimiter position indicated by the DP matching processing result.
[0018]
The invention of claim 6 is the transcription text automatic generation apparatus according to claim 1 or claim 4, wherein an acoustic model storage means for storing an acoustic model, voice data input means for inputting learning voice data, and Decoding means for decoding learning speech data input from the speech data input means using delimiter-inserted phoneme notation text data generated by the second information processing means, and passing based on the decoding result And further comprising a third information processing means for detecting a delimiter having a time of zero and creating transcription text data in which the detected delimiter is deleted from the phoneme-notated text data with the delimiter inserted therein. To do.
[0019]
The invention of claim 8 includes an input means for inputting kanji-kana mixed text data, a first information processing means for dividing the input kanji-kana mixed text data into predetermined recognition units, a character string, and The storage means storing the reading dictionary in which the reading is described and the text data divided by the first information processing means are converted into readings based on the reading dictionary and determined by the first information processing means. And a second information processing means for inserting a delimiter at the divided position.
[0020]
According to a ninth aspect of the present invention, there is provided a transcription text automatic generation device according to any one of the first, fourth, and eighth embodiments, a voice data input unit that inputs speech data for learning, and the automatic generation device. An acoustic model creating means for creating an acoustic model based on the generated transcription text and learning speech data input from the speech data input means, and performing speech recognition using the created acoustic model It is characterized by.
[0021]
The invention according to claim 10 is a program that is executed by an automatic transcription text generator having a storage means that stores a reading dictionary in which a character string and its reading are described and a plurality of readings are allowed for one character string. In the recording medium recorded with the above-mentioned program, the program inputs the phonetic transcription text data describing the phonetic transcription corresponding to the kanji-kana mixed text data and the phonetic notation corresponding to the kanji-kana mixed text data, and the input A first information processing step for dividing kanji-kana mixed text data for each predetermined recognition unit, and a matching position of the text data divided by the first information processing step, while matching with the division position of the text data, The character string in the divided kanji mixed kana text data is converted into a reading and the same When there are a plurality of readings for a character string, an optimal candidate among the plurality of readings is determined based on the phonetic transcription text data input from the input step, and the character string converted into the reading is determined. And a second information processing step of generating delimiter-inserted phoneme-notation text data by inserting a delimiter at the division position.
[0022]
The invention according to claim 11 is a program to be executed by a transcription text automatic generation device having a storage means storing a reading dictionary in which a character string and its reading are described and a plurality of readings are allowed for one character string. In the recording medium recorded with the above-mentioned program, the program inputs the phonetic transcription text data describing the phonetic transcription corresponding to the kanji-kana mixed text data and the phonetic notation corresponding to the kanji-kana mixed text data, and the input A first information processing step for dividing kanji-kana mixed text data for each predetermined recognition unit, and a matching position of the text data divided by the first information processing step, while matching with the division position of the text data, The character string in the divided kanji mixed kana text data is converted into a reading and the same When there are a plurality of readings for the character string, the optimal candidate among the plurality of readings is determined based on the phonetic transcription text data input from the input step, and the character string converted into the readings A second information processing step of generating a phoneme notation text data into which a delimiter has been inserted by inserting a delimiter at the divisional position of the phonetic transcription text data input from the input step in correspondence with It is characterized by that.
[0023]
DETAILED DESCRIPTION OF THE INVENTION
Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.
[0024]
The recognition unit of the embodiment described below is a morpheme, and the phoneme model to be used is a model considering three consecutive phonemes in the morpheme, and the model used between morphemes is a two-phoneme that does not depend on the subsequent or preceding phoneme. This is a model that takes into account. In addition, as a model corresponding to the delimiter, the SP notation meaning short silence that can be jumped is used.
[0025]
FIG. 1 shows a functional configuration of the speech recognition system according to the first embodiment of the present invention.
[0026]
The configuration of FIG. 1 shows the configuration of the acoustic model creation unit of the speech recognition system. The structure of the other parts of the speech recognition system is the same as that of the prior art. For example, NHK R & D No. The speech recognition system described in 57 19999.8 pp36 can be used.
[0027]
In FIG. 1, reference numeral 101 denotes kanji-mixed transcription data describing a kana-kanji character string corresponding to learning speech data. Reference numeral 102 denotes a morpheme reading dictionary used to automatically generate delimiter-inserted phoneme notation text data. A morpheme character string notation and one or more corresponding readings (phoneme notation) are set as one data set. A plurality of data sets are described in the reading dictionary 102.
[0028]
103 is phonetic notation transcription text data, which describes phoneme notation corresponding to the description contents of the kana mixed transcription text data. Reference numeral 104 denotes learning voice data used for learning the acoustic model, which is learning voice data obtained when a character string written in the kanji mixed-text data is uttered.
[0029]
Reference numeral 105 denotes a morpheme analysis unit that analyzes a character string described in text data mixed with kanji and kana, and divides the character string into morpheme units. The morphological analysis unit 105 is well known, and in this embodiment, a CPU that executes the Japanese analysis program “tea bowl” version 1.5 is used as the morphological analysis unit 105.
[0030]
Reference numeral 106 denotes kanji mixed text data (hereinafter referred to as divided text data) divided for each morpheme by the morpheme analysis unit 105. Reference numeral 107 denotes a pattern matching unit that automatically generates delimiter-inserted phoneme notation text data 108 using the divided text data 106, the reading dictionary 102, and the phoneme notation transcription text data 103. The pattern matching unit 107 can use a CPU that executes the processes shown in FIGS.
[0031]
Reference numeral 108 denotes phoneme notation text data in which delimiters are inserted in units of morphemes (hereinafter referred to as phoneme notation text data with delimiters inserted). Reference numeral 109 denotes an acoustic model learning unit that creates the acoustic model 110 using the phoneme notation text data 108 with the separators inserted and the learning speech data 104. As the acoustic model learning unit 109, a CPU that is equipped with and executes an HTK (Hidden Markov Model Tool, http://www.entropic.com/http/html) can be used.
[0032]
FIG. 2 shows details of the pattern matching unit 107 of FIG. In addition, the same code | symbol is attached | subjected to the location similar to FIG. 1, and detailed description is abbreviate | omitted.
[0033]
In FIG. 2, reference numeral 201 denotes a reference pattern creation unit, which creates a reference pattern 202.
[0034]
Reference numeral 202 denotes a reference pattern, which is data in which one or more readings described in the reading dictionary 102 are arranged in correspondence with the division position of the morpheme shown in the divided text data 106. A delimiter is inserted at the division position.
[0035]
Here's how to create a reference pattern.
[0036]
A character string from the divided text data to the first delimiter, in this case, “Now” (excluding punctuation marks) is extracted, and a data set having a notation corresponding to “Now” is read and searched in the dictionary 102. A reading “soredewa” corresponding to the notation is taken out from the data set obtained as a result of the search and inserted into the reference pattern. Next, the morpheme expression “today” at the second division position is extracted from the divided text data 106. When a data set having the same notation as “today” is obtained by the reading dictionary 102, corresponding readings are obtained from the data set, and in this case, “kyo” and “koNnich” are obtained.
[0037]
Therefore, after adding a delimiter after the last character of the character string of the reference pattern, the two notations obtained second are added. At this time, in order to indicate that there are a plurality of reading candidates for one notation, two sets of character strings are enclosed with {}, and a delimiter between the two sets of character strings is represented with a “,” symbol.
[0038]
In this way, the reading corresponding to the character string notation for each morpheme described in the divided text data 106 is acquired based on the reading dictionary 102, and a reference pattern is created by adding a delimiter. The reference pattern 202 can be created by taking the notation indicating the reading from the reading dictionary for all the character strings of the notation described in the divided text data and arranging them with a delimiter.
[0039]
A test pattern 203 is extracted from the phoneme transcription text data in units of a predetermined length for DP matching.
[0040]
204 is a reference pattern 202 and a test pattern according to a DP (Dynamic Programming) algorithm, for example, a two-stage DP or a building level algorithm (“Sound recognition basics (bottom) NTT Advanced Technology, pp 188-241)”. 203 is an information processing unit that compares the data with 203 and determines an optimum candidate from among a plurality of notation candidates for one morpheme character string in the reference pattern 202.
[0041]
A DP result processing unit 205 inserts a delimiter SP into the determined combination character string so as to match the position of the delimiter inserted in the reference pattern.
[0042]
In this example, as shown in FIG. 3, a combination of character strings indicated by reference numeral 301 by DP matching between the reference pattern 202 and the test pattern 203, that is, a combination of character strings “soredewa → kyo → no → nyu: su → desu”. The distance accumulation result (score A) for the combination of the character string “soredewa → koNnichi → no → nyu: su → desu” is obtained. In this example, since the previous combination has a low distance accumulation result, the previous combination is determined by the best path selection unit 302 in the information processing unit 204 as the best path (character string combination method). Thereafter, the delimiter replacing unit 303 in the DP result processing unit 205 replaces the delimiter in the DP result 301 with the SP delimiter.
[0043]
In the example of FIG. 3, the divided text data is converted into the first phoneme transcription text data based on the reading dictionary 102, and compared to the phoneme transcription text data 103, the same character string is obtained from a plurality of phoneme transcriptions. This is an example of determining the optimal candidate. In this example, when the reading described in the reading dictionary does not match the actual notation (when nyu: su (see reference pattern 202 and niu: su (see test pattern 203) in FIG. 3), a delimiter has been inserted. The phoneme-notated text data 108 has a drawback in that partially incorrect notations are mixed.
[0044]
Therefore, FIG. 4 shows a form in which this defect is improved. In FIG. 4, the DP result (best path) 401 output from the best path selection unit 302 and the test pattern (phoneme notation sequentially extracted from the phoneme notation transcription data that most faithfully represents the reading of the text data mixed with kanji and kana characters. (Character string) 203 is input to the delimiter insertion position calculation unit 402, and the delimiter corresponding position in the test pattern 203 obtained by tracing the delimiter symbol position (division position) of the DP result from which the best path 401 is obtained. Is calculated.
[0045]
A delimiter insertion unit (403) inserts a desired delimiter SP at the calculated delimiter position of the test pattern 203, and outputs delimiter inserted phoneme description text data 108.
[0046]
This is nothing more than inserting delimiters into the phonetic transcription text data so as to correspond to the division positions for each morpheme obtained by morphological analysis of the kanji-kana mixed text. For this reason, even if the kanji transcription text contains mixed kanji characters that are not registered in the dictionary, the kanji readings are correctly reflected in the phoneme notation text data 108 with the separators inserted.
[0047]
The transcription text automatic generation processing and acoustic model learning processing executed with the above functional configuration will be described.
[0048]
The kanji mixed kanji transcription text data 101 is input to the morpheme analysis unit 105 used for language model creation, and the morpheme analysis unit 105 outputs divided text data 106 divided for each morpheme. The pattern matching unit 107 acquires a reading character string (phoneme notation) corresponding to the notation indicated by the divided text data 106 from the reading dictionary 102, and rearranges the divided text data 106 into the reading character string. At this time, a reference pattern 202 (see FIG. 2) is created with a delimiter added to match the division position of the morpheme.
[0049]
By DP matching between the reference pattern 202 and the phonetic transcription text data, a character string most similar to the test pattern character string is selected from among a plurality of reading candidates included in the reference pattern. Further, the delimiter character in the reference pattern is converted into an SP character, and the delimiter-inserted phoneme symbol text data 108 is automatically generated.
[0050]
Based on the learning speech data 104 and the phoneme symbol text data 108 into which the delimiter has been inserted, the acoustic model learning unit 109 performs learning in the same manner as before and creates the acoustic model 110. Speech recognition is performed using the created acoustic model 110.
[0051]
Further, a process for generating phoneme notation text data with a delimiter inserted when speech recognition is performed using a model that takes into consideration three consecutive phonemes that span between words will be described with reference to FIG.
[0052]
In FIG. 5, an acoustic model 110 is an acoustic model generated by the apparatus of FIG. The phoneme notation text data 108 with the morpheme separator inserted is text data generated by the apparatus of FIG.
[0053]
Reference numeral 501 denotes a decoder that performs decoding based on the above data.
[0054]
Reference numeral 502 denotes a jump path SP detection / deletion unit that deletes an SP having a passage time of 0 (zero) in the decoding of the decoder 501, that is, an SP in the phoneme notation text data 108 into which the morpheme delimiter has been inserted. .
[0055]
The output of the jump path SP detection / deletion unit 502 is phoneme description text data 503 with a delimiter inserted for each SP in consideration of consecutive three phonemes.
[0056]
In such a configuration, the learning speech data 104 is decoded by the decoder 501 using the created acoustic model 110 and the phoneme notation text data 108 in which the insertion symbol is inserted for each morpheme used when learning the acoustic model 110. Decode with. The jump path SP detection / deletion unit detects an SP having a transit time of 0 (zero) based on the decoding result, that is, an SP that has passed through the jump path, deletes the SP from the phoneme notation text data 108, and delimits each SP. Inserted phoneme notation text data 503 is generated.
[0057]
The above processing can be defined by a software program. The software program may be executed by the CPU.
[0058]
FIG. 6 shows a specific example of a transcription text automatic generation apparatus having the functional configuration of FIG. 1 and a speech recognition apparatus having the same.
[0059]
In FIG. 1, a CPU 1000, a system memory 1010, an input device 1020, a display 1030, an input / output interface (I / O) 104, and a hard disk 1060 are connected via a bus.
[0060]
The CPU 1000 loads the automatic generation program stored in the hard disk 1060 into the system memory 1010 and then executes the automatic generation program to generate the text data 108 with the delimiter inserted. At this time, the CPU 1000 functions as a transcription text data generation device.
[0061]
When performing voice recognition using a voice recognition program (not shown) stored in the hard disk 1060 (same as the conventional one), the CPU 1000 functions as a voice recognition device.
[0062]
The system memory 1010 temporarily stores a program executed by the CPU 1000, input data for information processing executed in accordance with the program, and an information processing result. The input device 1020 has a keyboard and a mouse, designates a program to be executed, and gives operation instructions for characters and character processing.
[0063]
The display 1030 displays instructions input from the input device 1020 and also displays voice recognition results and the like. The I / O 1040 converts an analog audio signal input from the microphone 1050 into a digital audio signal and delivers it to the CPU 1000. The audio signal input from the microphone 1040 is stored in the hard disk 1060 as learning audio data 1066 (reference numeral 104 in FIG. 1).
[0064]
The hard disk 1050 includes an automatic generation program 1061, a kanji-kana mixed text 1062 (reference numeral 101 in FIG. 1), a phonetic transcription text 1063 (reference numeral 103 in FIG. 1), a reading dictionary 1064 (reference numeral 102 in FIG. 1), an acoustic model. 1065 and the learning voice data 1066 are stored. In addition, the hard disk 1060 stores an operating system and a voice recognition program for controlling the operation of the entire system.
[0065]
The text data 1062 and 1063 and the reading dictionary 1064 may be created using a document processing program such as a word processor or an editor, and may be input (saved) to the hard disk 1060, or data created by another information processing apparatus may be used. The data may be transferred and input to the hard disk 1060 in advance by online communication or offline communication via a floppy disk.
[0066]
The automatically generated program is installed in the hard disk 1060 from a portable recording medium such as a floppy disk or CDROM under the control of the CPU 1000.
[0067]
The contents of the automatic generation program 1061 are shown in FIG. The automatic generation program 1061 is described in a program language that can be executed by the CPU 1000, but for convenience of explanation, the function generation is expressed in FIG. A person skilled in the art would easily create an automatic generation program based on the description.
[0068]
In FIG. 7, step S <b> 10 is a process of reading text data mixed with kanji and kana from the hard disk 1060 to the system memory 1020.
[0069]
Step S20 is a process of performing morphological analysis on the kanji-kana mixed text data stored in the system memory 1020 (corresponding to the morphological analysis unit 105 in FIG. 1).
[0070]
Step S30 is a process of writing the divided text data (106 in FIG. 1) obtained by the morphological analysis into the system memory 1020.
[0071]
Step S 40 is processing for reading the reading dictionary 1064 into the system memory 1020. This process can be omitted when the reading dictionary 1064 has a large capacity, and the described data can be read directly from the hard disk 1060 during the pattern matching process described later.
[0072]
In step S50, pattern matching processing is performed based on the reading dictionary, phonetic transcription text data, and divided text data stored in the system memory 1020, to generate delimiter inserted phoneme text data. The CPU 1000 for executing this processing corresponds to the pattern matching unit 108 in FIG. As a method for inserting the delimiter SP, any one of the methods shown in FIG. 3 or FIG. 4 is used by the user.
[0073]
Step S60 is a process of storing the generated descriptive symbol-inserted phoneme notation text data in the hard disk 1060 for storage.
[0074]
Step S70 is an end determination. When there are a plurality of sets of text data 1062 and 1063, the processing of steps S10 to S60 is repeated for all the texts, and the phoneme-notated text data with the separators inserted is automatically generated. .
[0075]
Step S80 is a process of reading the learning voice data from the hard disk 1060 to the system memory 1020. Step S90 is a process of creating an acoustic model using the phoneme notation text data with delimiters inserted stored in the hard disk 1060 and the learning speech data stored in the system memory 1020. The CPU 1000 for executing this process functions as the acoustic model learning unit 109 in FIG.
[0076]
In this way, the CPU 1000 uses the created acoustic model to recognize a voice signal input from the microphone 1050 according to a voice recognition program. This process is similar to the prior art and will not require detailed description.
[0077]
In addition to the embodiments described above, the following embodiments can be implemented.
1) In the above embodiment, the phoneme notation text data with the separator inserted is automatically generated as the pre-processing of the speech recognition process. However, the phoneme notation text data with the separator inserted is automatically generated by the system of FIG. It may be delivered to another voice recognition device.
[0078]
2) The system of FIG. 6 is an example using a general-purpose computer such as a personal computer, but the hardware is not limited to a personal computer, and a digital processor may be used, or various information processing devices such as a workstation or a server are used. can do.
[0079]
3) In the above-described embodiment, the processing for calculating the break position and the replacement process for the break symbol have not been described in detail, but a known information processing method may be used. For example, as for the calculation position of the delimiter position, a function for detecting the number of words / number of characters from the beginning to an arbitrary word / character is prepared in the operating system (for example, Microsoft Windows 95, 98). Therefore, using this function, the number of words and characters from the beginning can be separated into symbol positions.
Further, since the replacement process is often used in the document editing process, detailed information processing contents will not require explanation.
[0080]
4) The above-described embodiment uses a dictionary that allows a plurality of readings in consideration of practical use. However, in an environment where the speech content that is the target of speech recognition is limited, the reading of words mixed with kana-kanji in advance is performed. Only one set needs to be registered.
[0081]
5) The following processing can be performed as a method combining the processing methods of FIGS. 3 and 4 described above. That is, a kana-kanji mixed word is read and converted into a reading using a dictionary. If there is no kana-kanji mixed word in the phonetic dictionary even after searching the phonetic dictionary, the phoneme notation is created using the phoneme notation corresponding to the division position in the phoneme notation text data. To do.
[0082]
【The invention's effect】
As described above, according to the present invention, the phoneme notation transcription text data without the delimiter symbol does not know the position of the delimiter, but when using the kanji or kana mixed text data, the morpheme analysis is performed. The inventor of the present application notices that it is possible to detect a kana-mixed character string (word) break for each morpheme, and reads the kana-mixed text divided at the break position using a dictionary (phoneme-written character string). (Corresponding to the invention of claim 8). This also makes it possible to automatically generate a transcription with a phoneme notation inserted with a delimiter automatically instead of manually.
[0083]
According to the first aspect of the present invention, an optimal reading is obtained by comparing the phonetic transcription with no delimiter inserted and a divided text converted into a reading and having a plurality of readings for the same character string. Obtainable.
[0084]
According to the invention of claim 4, in accordance with the position indicated by the divided phoneme notation text changed to reading, the phonetic notation transcription that represents the most faithful reading with respect to the text data mixed with kana-kanji and has no delimiter inserted. Since the delimiter is inserted into the text data (phoneme transcription text data input from the input means), it is possible to automatically generate delimited symbol-inserted phoneme notation text data with high reading accuracy.
[0085]
In this way, the transcription text used to create the acoustic model can be automatically generated, so the generation time of the transcription text can be greatly reduced, and the user's operations related to the creation of the acoustic model can be reduced. Labor can also be greatly reduced.
[Brief description of the drawings]
FIG. 1 is a block diagram showing a functional configuration of an embodiment of the present invention.
FIG. 2 is a block diagram illustrating a functional configuration of a pattern matching unit according to an embodiment of the present invention.
FIG. 3 is a block diagram showing the contents of DP matching processing according to the embodiment of the present invention.
FIG. 4 is a block diagram showing total delimiter symbol processing contents according to the embodiment of the present invention.
FIG. 5 is a block diagram illustrating a delimiter insertion method using a decoder.
FIG. 6 is a block diagram showing a specific system configuration according to the embodiment of the present invention.
FIG. 7 is a flowchart showing the processing contents of an automatically generated program.
[Explanation of symbols]
101 Transcript text data mixed with kanji
102 Reading dictionary
103 Phonetic transcription text data
104 Voice data for learning
105 Morphology analysis unit
106 Divided text data
107 Pattern matching section
108 Phonetic text data with delimiters
109 Acoustic model learning unit
110 Acoustic model

Claims (10)

Input means for inputting kanji-kana mixed transcription text data and phoneme-notation transcription text data describing the phonetic notation corresponding to the kanji-kana mixed text data;
First information processing means for dividing the input kanji-kana mixed text data into predetermined recognition units;
A storage means that stores a reading dictionary in which a character string and its reading are described and a plurality of readings are allowed for one character string;
Based on the reading dictionary, the character string in the divided kanji-kana mixed text data is converted into a reading based on the reading dictionary while matching the dividing position of the text data divided by the first information processing means, and the same character When there are a plurality of readings for a sequence, the optimal candidate in the plurality of readings is determined based on the phonetic transcription text data input from the input means, and the character string converted into the reading A transcription text automatic generation apparatus comprising: a second information processing unit that inserts a delimiter symbol at a division position and generates delimiter-inserted phoneme-notation text data. 2. The transcription text automatic generation apparatus according to claim 1, wherein the predetermined recognition unit is a morpheme, and the first information processing means is a morpheme analysis unit. 2. The transcription text automatic generation apparatus according to claim 1, wherein the second information processing unit determines an optimum candidate by DP matching processing when there are a plurality of readings for the same character string to be converted. , Create a reference pattern with a delimiter in the character string converted into the reading for the DP matching process, and extract a character string of a predetermined length from the phonetic transcription text data A transcription text automatic generation apparatus characterized by creating a test pattern and subjecting the reference pattern and the test pattern to DP matching processing. Input means for inputting kanji-kana mixed transcription text data and phoneme-notation transcription text data describing the phonetic notation corresponding to the kanji-kana mixed text data;
First information processing means for dividing the input kanji-kana mixed text data into predetermined recognition units;
A storage means that stores a reading dictionary in which a character string and its reading are described and a plurality of readings are allowed for one character string;
Based on the reading dictionary, the character string in the divided kanji-kana mixed text data is converted into a reading based on the reading dictionary while matching the dividing position of the text data divided by the first information processing means, and the same character When there are a plurality of readings for a column, the optimum candidate among the plurality of readings is determined based on the phonetic transcription text data input from the input means, and the character string converted into the readings is determined. Second information processing means for generating phoneme notation text data into which a delimiter symbol has been inserted by inserting a delimiter symbol into the division position of the phonetic transcription transcription data input from the input means in correspondence with the division position; Transcript text automatic generation device characterized by that. 5. The transcription text automatic generation apparatus according to claim 4, wherein the second information processing unit creates a reference pattern including a character string converted into the reading for the DP matching processing, and the phoneme A test pattern is created by extracting a character string of a predetermined length from the written transcription text data, and the reference pattern and the test pattern are subjected to DP matching processing and indicated by the DP matching processing result on the test pattern. A transcription text automatic generation apparatus characterized by inserting a delimiter at a delimiter position. 5. The transcription text automatic generation apparatus according to claim 1 or 4, wherein an acoustic model storage unit storing an acoustic model, a voice data input unit for inputting learning voice data, and the second information processing unit. Decoding means for decoding learning speech data input from the speech data input means using the generated delimiter-inserted phoneme notation text data, and a delimiter whose transit time is zero based on the decoding result And a third information processing means for creating transcription text data in which the detected delimiter is deleted from the phoneme-notated text data into which the delimiter has been inserted. . An input means for inputting text data transcribed with kanji and kana,
First information processing means for dividing the input kanji-kana mixed text data into predetermined recognition units;
Storage means for storing a character dictionary and a reading dictionary in which the reading is written;
Second information processing that converts the text data divided by the first information processing means into reading based on the reading dictionary and inserts a delimiter at the division position determined by the first information processing means. A transcription text automatic generation device characterized by comprising means. The transcription text automatic generation device according to any one of claims 1, 4, and 8;
Voice data input means for inputting voice data for learning;
Acoustic model creation means for creating an acoustic model based on the transcription text generated by the automatic generation device and the learning speech data input from the speech data input means,
A speech recognition apparatus that performs speech recognition using the created acoustic model. In a recording medium in which a character string and its reading are described, and a program to be executed by a transcription text automatic generation device having a storage means storing a reading dictionary in which a plurality of readings are allowed for one character string is recorded, The program is
An input step for inputting kanji-kana mixed transcription text data and phoneme-notation transcription text data describing a phoneme notation corresponding to the kanji-kana mixed text data;
A first information processing step of dividing the input kanji-kana mixed text data into predetermined recognition units;
Based on the reading dictionary, the character string in the divided kanji-kana mixed text data is converted into a reading based on the reading dictionary while matching the dividing position of the text data divided in the first information processing step, and the same character When there are a plurality of readings for a sequence, an optimal candidate in the plurality of readings is determined based on the phonetic transcription text data input from the input step, and the character string converted into the reading A recording medium comprising: a second information processing step of generating delimiter-inserted phoneme notation text data by inserting a delimiter at a division position. In a recording medium in which a character string and its reading are described, and a program to be executed by a transcription text automatic generation device having a storage means storing a reading dictionary in which a plurality of readings are allowed for one character string is recorded, The program is
An input step for inputting kanji-kana mixed transcription text data and phoneme-notation transcription text data describing a phoneme notation corresponding to the kanji-kana mixed text data;
A first information processing step of dividing the input kanji-kana mixed text data into predetermined recognition units;
Based on the reading dictionary, the character string in the divided kanji-kana mixed text data is converted into a reading based on the reading dictionary while matching the dividing position of the text data divided in the first information processing step, and the same character If there are multiple readings for the sequence, the optimal candidate among the multiple readings is determined based on the phonetic transcription text data input from the input step, and the character string converted into the reading A second information processing step of generating a phoneme notation text data into which a delimiter has been inserted by inserting a delimiter at a division position of the phonetic transcription transcription data input from the input step in correspondence with the division position; A recording medium characterized by the above.

Images