JP2014134640A - Transcription device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a transcription device and a program capable of improving accuracy and performance of voice recognition.SOLUTION: A model storage unit stores an acoustic model and a language model. A voice recognition unit performs recognition processing based on feature quantity extracted from input voice data, while referring to the acoustic model and the language model. A correction unit enables correction of a recognition result text to be done based on an operation of a user, and creates a correction result text as a result of the correction. A pronunciation imparting unit imparts a phoneme string corresponding to the correction result text. An alignment unit associates the imparted phoneme string and the input voice data in time, and aligns them. An adaptation unit adapts the acoustic model and the language model based on the aligned phoneme string, the input voice data and the correction result text.

Description

  The present invention relates to a transcription apparatus that wakes up characters based on speech and outputs the resulting text and a program thereof.

  It is a common practice to record speech on some medium, such as the speech of a conference participant or a speech at a press conference, and transcribe it as a speaker and record it as text. Such an operation for converting a voice into a character is called a transcription or a transcription.

  In order to perform transcription manually, that is, a method in which a transcription operator inputs the utterance content from a keyboard or the like while listening to the recorded voice, requires a great deal of labor. Therefore, efforts are being made to reduce the labor required by the human hands using voice recognition technology.

  Patent Documents 1 and 2 disclose a configuration in which text of a speech recognition result obtained by speech recognition processing is edited and written to a storage device.

Japanese Patent No. 4020083 Japanese Patent No. 3859612

  However, the techniques disclosed in the prior art documents listed above cannot be said to have sufficient speech recognition accuracy. If an error is included in the recognition result, there is a problem that it is necessary to manually correct the part, and the work efficiency does not increase.

  The present invention has been made based on the above problem recognition, and provides a transcription apparatus and a program thereof that can improve the accuracy and performance of speech recognition. In particular, the present invention provides a transcription apparatus and a program thereof that can adaptively improve the accuracy and performance of speech recognition even during operation of the apparatus.

  [1] In order to solve the above-described problem, a transcription apparatus according to an aspect of the present invention includes a model storage unit that stores an acoustic model and a language model, and an acoustic model and a language model that are stored in the model storage unit. While referring to the speech recognition unit that performs recognition processing based on the feature amount extracted from the input speech data and outputs the recognition result text, and enables the correction of the recognition result text based on the operation by the user, and results of the correction A correction unit for creating a correction result text to be obtained, a pronunciation giving unit for giving a phoneme string corresponding to the pronunciation of the correction result text, the phoneme sequence given by the pronunciation giving unit, and the input voice data An alignment unit that associates and aligns, a correspondence relationship between the phoneme sequence aligned by the alignment unit and the input voice data, and the input The model storage unit is updated by adapting or creating an acoustic model based on the feature amount of the speech data, and the model storage unit is updated by adapting or creating a language model based on the correction result text. And an adapting unit.

  According to this configuration, the pronunciation sequence is automatically assigned based on the recognition result text and the correction result text obtained by correcting the recognition result text. In addition, alignment in the time direction is performed between the assigned pronunciation sequence and the input voice data. Therefore, it is possible to adapt the model based on the recognition result text and the correction result text for both the acoustic model and the language model. Therefore, the accuracy of voice recognition can be improved.

  [2] Further, according to one aspect of the present invention, in the above transcription apparatus, when the correction unit does not correct the recognition result text based on an operation by a user, the correction result text is corrected as it is. The result is output as text.

  With this configuration, it is possible to adapt the model even when all or part of the recognition result text is not corrected.

  [3] Further, according to one aspect of the present invention, in the above transcription apparatus, a recognition process by the voice recognition unit, a process in which the correction unit creates the correction result text, and the pronunciation giving unit in the phoneme sequence A series of processes including a process for assigning the phoneme sequence and the input speech data, and a process for the adaptive unit to update the model storage unit are repeated a plurality of times. A control unit for controlling.

  As a result, the transcription process based on certain input voice data can be repeated a plurality of times. By using such a transcription device, the user can gradually create a transcription result and adapt the model to improve the accuracy of the speech recognition result.

  [4] Further, according to one embodiment of the present invention, a computer stores a model storage unit that stores an acoustic model and a language model, and inputs audio data while referring to the acoustic model and the language model stored in the model storage unit. A speech recognition unit that performs recognition processing based on the extracted feature values and outputs a recognition result text, a correction that enables correction of the recognition result text based on a user operation, and creates a correction result text obtained as a result of the correction Means, a pronunciation giving means for giving a phoneme string corresponding to the pronunciation of the correction result text, an alignment means for aligning the phoneme string given by the pronunciation giving means and the input speech data in association with each other in time, the alignment A correspondence relationship between the phoneme string arranged by the means and the input voice data, and a feature amount of the input voice data Adapting or creating an acoustic model based on the above, updating the model storage unit, and adapting or creating a language model based on the correction result text to update the model storage unit, It is a program to make it function.

  According to the present invention, since the model is updated based on the voice recognition result and the correction result, the voice recognition accuracy can be improved. Therefore, the work efficiency of transcription can be improved. Further, according to the present invention, the accuracy and performance of voice recognition can be adaptively improved even during operation of the apparatus.

It is a block diagram which shows schematic function structure of the transcription apparatus by embodiment of this invention. It is the schematic which shows the structure of the input audio | voice data which the input audio | voice storage part by the same embodiment memorize | stores. It is the schematic which shows the data structure of the feature-value extracted by the feature-value extraction part by the embodiment. It is the schematic which shows the structure of the data which the recognition result memory | storage part by the same embodiment memorize | stores. It is the schematic which shows the structure of the correction result data which the correction result output part by the embodiment outputs. It is a flowchart which shows the whole process sequence of the transcription apparatus by the same embodiment. It is the schematic which shows the example of a screen of the user interface which the correction part by the embodiment provides.

Hereinafter, embodiments of the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram showing a schematic functional configuration of the transcription apparatus according to the present embodiment. As shown in the figure, the transcription apparatus 1 includes a voice acquisition unit 11, an input voice storage unit 12, a feature amount extraction unit 14, a model storage unit 15, a voice recognition unit 16, a recognition result storage unit 18, The correction unit 20 includes a correction result output unit 22, a sound generation giving unit 24, an alignment unit 26, an adaptation unit 28, and a control unit 40.

The voice acquisition unit 11 acquires voice data from the outside and writes it in the input voice storage unit 12.
The input voice storage unit 12 stores voice data to be subjected to a transcription process.
The feature quantity extraction unit 14 reads the input voice data stored in the input voice storage unit 12 and extracts the acoustic feature quantity. The feature quantity extraction unit 14 stores the extracted acoustic feature quantity data in association with the input voice data.

The model storage unit 15 includes an acoustic model storage unit 151 and a language model storage unit 152 therein. The acoustic model represents a relationship between an acoustic feature amount extracted from speech and a phoneme symbol. The language model represents a statistical tendency regarding the appearance of language elements (such as characters and words). In the present embodiment, the acoustic model storage unit 151 stores an HMM (Hidden Markov Model) acoustic model. The language model storage unit 152 stores an n-gram language model.
The speech recognition unit 16 executes input speech data recognition processing while referring to the acoustic model and the language model stored in the model storage unit 15 using the acoustic feature amount extracted by the feature amount extraction unit 14. . The voice recognition unit 16 writes text data as a recognition result in the recognition result storage unit 18.
The recognition result storage unit 18 stores the text data of the recognition processing result output from the voice recognition unit 16.

  The correction unit 20 has a function of editing and correcting the text data of the recognition result stored in the recognition result storage unit 18, and enables the correction of the recognition result text based on a user operation. And the correction part 20 produces the correction result text obtained as a result of correction. Specifically, the correction unit 20 provides a user interface. As a result, the user can perform an operation of correcting the erroneous part of the recognition result text data. At this time, the correcting unit 20 displays the position of the text data of the recognition result and the position of the input voice data in association with each other according to time. Further, the correction unit 20 displays a waveform corresponding to the input voice data on the screen. In addition, the correction unit 20 has a function of starting the input voice data from an arbitrary position designated by the user and reproducing the voice.

  The correction result output unit 22 outputs the text data of the recognition processing result and the text data reflecting the correction result when corrected. The output destination by the correction result output unit 22 is an information storage medium, an external device, or the like. The correction result output unit 22 also outputs the above text data reflecting the correction result to the pronunciation providing unit 24. The text data output by the correction result output unit 22 is data of a transcription result (hereinafter sometimes referred to as “transcription text”). If there is no correction portion for the recognition result, the text data output by the correction result output unit 22 is the same as the text data of the recognition processing result.

The pronunciation giving unit 24 gives a pronunciation series to the transcription text. The pronunciation series is typically data of a string of symbols representing phonemes.
The aligning unit 26 associates the position (time) of the pronunciation sequence provided by the pronunciation providing unit 24 with the position (time) of the input voice data (alignment, alignment). At this time, the feature amount extraction unit 14 passes the input voice data and the corresponding acoustic feature amount data string (shown in FIG. 3) to the alignment unit 26.

The adaptation unit 28 includes the transcription text output from the correction result output unit 22 or the recognition result text read from the recognition result storage unit 18, the phoneme string aligned by the alignment unit 26 corresponding to the text, and the input The model (acoustic model and / or language model) stored in the model storage unit 15 is adapted using the voice data and the corresponding acoustic feature data string (shown in FIG. 3). To do. Specifically, the adaptation unit 28 performs acoustical sound based on the correspondence between the phoneme sequence aligned by the alignment unit 26 and the input speech data (the relationship associated by relative time) and the feature amount of the input speech data. The model storage unit 15 is updated by adapting or creating the model. Further, the adaptation unit 28 adapts or creates a language model based on the correction result text and updates the model storage unit 15. There are a case where the adaptation unit 28 updates an existing model already registered in the model storage unit 15 and a case where a new model is created and registered in the model storage unit 15. The model adapted here can be referred to in the recognition process after the next time.
The control unit 40 controls the operation of each unit included in the transcription apparatus 1. The control unit 40 also controls a user interface (output to a computer display screen, character input from a keyboard, instruction input using a pointing device such as a mouse).

  In addition, each said part is implement | achieved, for example using an electronic circuit etc. Each storage unit for storing data is realized using a semiconductor memory, a magnetic disk device, or the like.

  FIG. 2 is a schematic diagram showing a configuration of input voice data stored in the input voice storage unit 12. As an example, data stored in the input voice storage unit 12 is stored in an object-oriented database. As shown in the figure, this data is table data having items of utterance number, relative time, and speech waveform data. Each row in this table corresponds to an utterance. Here, an utterance is a segment of a word spoken by a person. The breaks at this time are the beginning and end points, punctuation marks in the middle, or pause. The utterance number is a serial number that allows each utterance to be uniquely identified. The relative time indicates the relative position of the utterance in the entire input voice data, and is expressed in the format of “HH: MM: SS.hh” (hour, minute, second, one hundredth of a second). The The voice waveform data is data representing a voice waveform, and is specifically a string of amplitude values sampled at a predetermined sampling frequency.

  FIG. 3 is a schematic diagram illustrating a data structure of a result of feature amount extraction by the feature amount extraction unit 14. As an example, the feature amount extraction result data is stored in an object-oriented database. As shown in the figure, this data is data of a table having items of utterance number, relative time, speech waveform data, and feature amount data. Each row in this table corresponds to an utterance. The items of the utterance number, relative time, and voice waveform data are as described in FIG. What is stored in the feature quantity data item is acoustic feature quantity data extracted from the speech waveform data. For example, frequency cepstrum coefficients (MFCC, Mel-Frequency Cepstrum Coefficients) are used as the acoustic feature amount. Specifically, the feature amount data item stores time-series data of frequency cepstrum coefficients. Note that the length of the window when performing Fourier transform for feature quantity extraction is, for example, 10 milliseconds.

  FIG. 4 is a schematic diagram illustrating a data configuration of the recognition result storage unit 18. The voice recognition unit 16 writes this data. As an example, the recognition result data is stored in an object-oriented database. As shown in the figure, this data is table data having items of utterance number, relative time, speech waveform data, feature amount data, and recognition result (text). Each row in this table corresponds to an utterance. The items of the utterance number, relative time, voice waveform data, and feature amount data are as described in FIGS. The item of the recognition result stores text as a result of the speech recognition unit 16 recognizing the utterance. Note that relative time data is associated with each word or character included in the text of the recognition result. Thereby, the word or character included in the text of the recognition result is associated with the corresponding time data in the speech waveform data or the corresponding time data in the feature data.

  FIG. 5 is a schematic diagram illustrating a configuration of correction result data output by the correction result output unit 22. As an example, the correction result data is stored in an object-oriented database. As shown in the figure, this data is table data having items of utterance number, relative time, speech waveform data, feature data, recognition result (text), and correction result (text). Each row in this table corresponds to an utterance. The items of utterance number, relative time, speech waveform data, feature amount data, and recognition result (text) are as described in FIGS. The correction result item stores data of the result corrected by the correction process in the correction unit 20. As described above, the correction result is data obtained by correcting the recognition result by the user. This correction means that a part of the recognition result is deleted on a word or character basis, a part of the recognition result is replaced with another word or character, or a new one that was not included in the recognition result. An operation such as inserting a word or a character. The pair of the recognition result text and the correction result text itself has information on what correction operation has been performed. Additional data indicating what correction operation has been performed may be held together.

  FIG. 6 is a flowchart illustrating a processing procedure performed by the transcription apparatus 1. Hereinafter, the flow of processing will be described along this flowchart. The input voice data acquired by the voice acquisition unit 11 is already stored in the input voice storage unit 12. In the model storage unit 15, a plurality of acoustic models and language models are already stored. Since the acoustic model depends on the speaker and the language model depends on the topic, a plurality of acoustic models are prepared.

First, in step S201, the control unit 40 selects input voice data to be transcribed based on an instruction from the user.
Next, in step S202, the control unit 40 selects an acoustic model corresponding to the speaker to be transcribed and a language model corresponding to the topic based on an instruction from the user. At this time, if the optimal acoustic model or the optimal language model does not exist in the model storage unit 15, the acoustic model closest to the corresponding speaker or the language model closest to the topic is selected. To. Alternatively, a general (general purpose) acoustic model or language model may be selected.

  Next, in step S203, voice recognition processing is executed. Specifically, the feature amount extraction unit 14 extracts the acoustic feature amount of the selected input speech data, and the speech recognition unit 16 performs recognition processing of the input speech data using the selected model. Then, the voice recognition unit 16 writes the recognition result data in the recognition result storage unit 18. As already described, the text of the speech recognition result is written in the recognition result storage unit 18 in association with the original speech data and the extracted feature data at the time.

  Next, in step S204, the control unit 40 determines whether or not to correct the recognition result based on an operation from the user. As a result of the determination, if the user gives an instruction that no correction is necessary (step S204: NO), the process proceeds to step S205. If an instruction to correct is given from the user (step S204: YES), the process proceeds to step S206.

When the process proceeds to step S205, the correction result output unit 22 saves and outputs the recognition result text as a transcribed result as it is. In this case, the correction unit 20 passes the recognition result data read from the recognition result storage unit 18 to the correction result output unit 22 without correction. That is, when the recognition result text is not corrected based on the operation by the user, the correction unit 20 and the correction result output unit 22 regard the recognition result data as it is as the correction result data, and save / output the data.
By adopting such processing, the user can give priority to the speed of outputting the transcription result rather than the accuracy of the recognition result. Further, even when the recognition result is not corrected as described above, the model can be adapted using the recognition result. When the speech recognition process is performed again later, the model adapted this time can be used.

  Next, when the process proceeds to step S206, the correction unit 20 corrects the text in the same step. Specifically, the correction unit 20 first displays the recognition result text on the screen in association with the input voice data. At this time, the correction unit 20 graphically displays the input voice data in the form of a voice waveform on the screen. An example of the screen display will be described later. Further, the correcting unit 20 reproduces and outputs the sound from the designated position (time) in the input sound data based on an instruction from the user using a pointing device or the like. The user corrects the text by specifying an error location based on the displayed information and reproduced voice, and inputting characters from a keyboard or the like. The corrected text is stored and output as a transcription result by the correction result output unit 22 in the format described with reference to FIG. Specifically, the correction result output unit 22 outputs the data of the correction result to an external device or the like, or writes it to a storage medium in the transcription device.

  Note that the correction result output unit 22 passes the correction result data to the pronunciation providing unit 24 as character transcription data, regardless of whether the process proceeds to step S205 or to step S206. As the transcription result data, speech waveform data, feature amount data, and corrected text data are stored in a form that can be associated with each other according to time. After completion of the processing in step S205 or step S206, in either case, the process proceeds to the next step S207.

  Next, in step S207, the pronunciation providing unit 24 adds the text of the correction result (including the case where the text of the recognition result is regarded as the text of the correction result as it is because correction has not been performed) to the phonetic sequence (phoneme string). ). Then, the aligning unit 26 performs an aligning process (alignment process) for associating in which time interval in the input voice data the pronunciation is made.

  The specific procedure of the process for the pronunciation giving unit 24 to give the pronunciation series is as follows. That is, the pronunciation providing unit 24 converts the correction result text given as a kanji kana mixed sentence into a text of only hiragana. For example, a morphological analyzer such as “MeCab” is used to convert a kanji-kana mixed sentence text into a hiragana-only text. “MeCab” is an open source (generally provided in the form of source program) morphological analysis engine program [Reference URL: http://mecab.googlecode.com/svn/trunk/mecab/doc/index .html]. After conversion to the text of only hiragana, the pronunciation providing unit 24 refers to the correspondence table of hiragana and phonetic symbols stored in advance, thereby converting the hiragana text into phonetic symbols (a string of symbols representing phonemes). Convert to Note that when sound is imparted using such an existing technique, the accuracy is about 95%. That is, the obtained phoneme string includes an error of about 5%, but the numerical value of 95% is a sufficiently good value as the accuracy of the information used for adapting the model.

  An example of a process in which the aligning unit 26 aligns the input speech data and the phoneme symbol sequence is as follows. That is, the alignment unit 26 searches for a state transition sequence of the acoustic model (HMM) that maximizes the probability of outputting the phoneme sequence provided by the sound generation assigning unit 24. This processing is called viterbi alignment and is one of existing technologies. By executing this Viterbi alignment, time information corresponding to each phoneme included in the assigned phoneme string can be obtained. That is, the alignment unit 26 corresponds to associating the correction result text, the phoneme string corresponding to the text, the input voice data, and the feature amount of the voice data on the same time axis. Output data. The size of the “time interval” is set to 10 milliseconds, which is equal to the length of the time window when extracting the feature value from the input voice data. In other words, the minimum unit of alignment processing by the alignment unit 26 is 10 milliseconds.

  Next, in step S208, the adaptation unit 28 uses the model (acoustic model or language model, or those models) based on the alignment result output by the alignment unit 26, the transcription text (corrected result text), and the input speech data. Both). Specifically, the adaptation unit 28 creates a new model and registers it in the model storage unit 15 or updates the original model (the used model selected in the immediately preceding speech recognition process). Do some processing. A specific processing example for the adaptation unit 28 to adapt the model is as follows.

  That is, with respect to the acoustic model, when an acoustic model that matches the actual speaker is selected, the adaptation unit 28 performs the MAP (Maximum a posteriori) method or the MLLR (Maximum Likelihood Linear Regression) method. Use to adapt the model. Both the MAP method and the MLLR method are existing technologies that can be used. When the acoustic model does not match the actual speaker, the adaptation unit 28 learns a new acoustic model using the Baum-Welch algorithm.

  As for the language model, when a language model that matches the actual topic to be recognized is selected, the adaptation unit 28 adapts the model using the MAP method or the model mixture method. . The model mixing method counts the occurrence frequency of n-grams in the existing language model, taking into account the occurrence frequency of n-grams in the newly obtained text, and the appearance of newly calculated n-grams This is a method of updating a language model with probability. If the actual topic and the language model do not match, the adaptation unit 28 learns a new language model based on the acquired text.

  In step S209, the control unit 40 determines whether to complete the transcription process based on an operation from the user. As a result of the determination, when the user gives an instruction to complete the process (step S209: YES), the process of this flowchart is terminated. If the user gives an instruction not to complete (step S209: NO), the process returns to step S203, and the transcription process is continued using the adapted model (including the newly learned model). To do. That is, a series of processing from step S203 to S208 is repeated.

  In this way, it becomes possible to perform detailed transcription gradually by repeating a plurality of times. As an example, when writing a text to be used for manuscripts or subtitles of news footage based on the speaker's utterance at a press conference, in the first stage, priority is given to promptness and only important parts of the utterance It is enough if is able to transcribe. That is, even if the recognition result of the other part is wrong, it may not be used at that time. At a later stage, more parts of the utterance need to be transcribed more accurately. That is, there is a situation in which the control unit 40 performs control so that transcription of the input voice data can be repeated a plurality of times, and it is not always necessary to transcribe all utterances completely and accurately from the first time. The control unit 40 controls the process according to the above-described procedure so that the use can be performed in consideration of these circumstances.

  FIG. 7 is a schematic diagram illustrating an example of a user interface screen displayed by the correction unit 20. In the figure, reference numeral 80 denotes a correction screen. The correction screen is displayed on the display device as one window managed by basic software of a personal computer, for example. Reference numeral 81 denotes a part of the screen 80, which is an area for displaying input speech data to be transcribed as a waveform. Reference numeral 82 denotes a part of the screen 80, which is an area for displaying the text of the recognition processing result output from the voice recognition unit 16. 83 is also a part of the screen 80, and is an editing area for correcting the text of the recognition processing result. The user can correct and edit the text included in the editing area 83 by using a pointing device or a keyboard. The area 81 for displaying the speech waveform, the area 82 for displaying the text of the recognition process result, and the editing area 83 for the correction process are associated with each other using a vertical broken line. . In the figure, the horizontal direction corresponds to the time axis. A plurality of broken lines drawn vertically between the area 81 and the area 82 and a plurality of broken lines drawn vertically between the area 82 and the editing area 83 indicate a predetermined time.

  Note that the function of the transcription apparatus in the above-described embodiment may be realized by a computer. In that case, the program for realizing the control function may be recorded on a computer-readable recording medium, and the program recorded on the recording medium may be read into a computer system and executed. Here, the “computer system” includes an OS and hardware such as peripheral devices. The “computer-readable recording medium” refers to a storage device such as a flexible disk, a magneto-optical disk, a portable medium such as a ROM and a CD-ROM, and a hard disk incorporated in a computer system. Furthermore, a “computer-readable recording medium” dynamically holds a program for a short time, like a communication line when transmitting a program via a network such as the Internet or a communication line such as a telephone line. In this case, a volatile memory inside a computer system serving as a server or a client in that case may be included, and a program that holds a program for a certain period of time. The program may be a program for realizing a part of the above-described functions, or may be a program that can realize the above-described functions in combination with a program already recorded in a computer system.

  Although a plurality of embodiments have been described above, the present invention can also be implemented in the following modifications.

(Modification 1) In the above-described configuration, the pronunciation giving unit gives a pronunciation (phoneme string) to all of the correction result text output by the correction result output unit 22, and the adaptation unit 28 uses all of them. It was decided to adapt the model. In the present modification, the sounding imparting unit 24 provides sounding only for the utterance selected by the user. The adapting unit adapts the model using the correction result text, pronunciation, and input speech data (and feature amount) corresponding to only the utterance selected by the user. In this case, the correction result output unit 22 or the pronunciation providing unit 24 displays a screen that allows the user to select an utterance based on the correction result data shown in FIG. 5 and accepts a selection operation from the user. Like that.
By configuring as in this example, not all recognition results and all correction results are used for model adaptation, but model adaptation using only a part of them is performed. Thereby, the calculation amount required for adaptation can be reduced.

(Variation 2) It has already been described that the accuracy of automatic sound generation by the sound generation unit 24 is about 95%. In this modification, the sound generation unit 24 displays the result of sound generation by automatic processing on a screen or the like, and can correct the sound generation (phoneme string) by an input operation from the user. In the subsequent processing, the model is adapted using the pronunciation automatically given by the pronunciation giving unit 24 and corrected by the user. Thereby, the adaptation of the model can be performed better, and the accuracy of speech recognition is further improved.

(Modification 3) In the embodiment described above, speech recognition processing is performed (offline) on a set of recorded input speech data, and transcription processing using the recognition result text and the correction result text is performed. It was. In this modification, speech recognition processing is performed online based on a speech stream input from the outside, and transcription processing using recognition result text and correction result text is performed. Based on the correction result text, pronunciation assignment, alignment, and model adaptation are performed by the same method as in the above-described embodiment. That is, the model is adapted in parallel with the online speech recognition process.

In the above-described embodiment and its modifications, the control by the control unit 40 is summarized below.
(A) After performing the voice recognition process, the control unit 40 controls whether or not the recognition result text correction process is performed based on an instruction from the user.
(B) When the correction process of the recognition result text is performed, the control unit 40 performs adaptation using the entire correction result text or uses only a part of the correction result text based on an instruction from the user. Controls whether adaptation is performed.
(C) Based on an instruction from the user, the control unit 40 adapts the model based on the correction result text only once with respect to certain input voice data, or repeatedly adapts the model based on the correction result text several times. Control what to do.
(D) The control unit 40 controls whether the speech recognition process is performed online or the speech recognition process is performed offline based on an instruction from the user.

  The embodiment of the present invention has been described in detail with reference to the drawings. However, the specific configuration is not limited to this embodiment, and includes designs and the like that do not depart from the gist of the present invention.

  The present invention can be used to improve the efficiency of transcription work. Specifically, it can be used to create conference minutes, to create records based on press conferences and interviews, and to transcribe audio from broadcast programs.

DESCRIPTION OF SYMBOLS 1 Transcription apparatus 11 Voice acquisition part 12 Input voice memory | storage part 14 Feature-value extraction part 15 Model memory | storage part 16 Speech recognition part 18 Recognition result memory | storage part 20 Correction part 22 Correction result output part 24 Pronunciation imparting part 26 Alignment part 28 Adaptation part 40 Control unit 151 Acoustic model storage unit 152 Language model storage unit

Claims (4)

A model storage unit for storing an acoustic model and a language model;
Referring to the acoustic model and the language model stored in the model storage unit, performing a recognition process based on the feature amount extracted from the input speech data, and outputting a recognition result text;
A correction unit that enables correction of the recognition result text based on an operation by a user and creates a correction result text obtained as a result of the correction;
A pronunciation giving unit that gives a phoneme string corresponding to the pronunciation of the correction result text;
An alignment unit that aligns the phoneme sequence provided by the pronunciation providing unit and the input speech data in association with each other at time;
The model storage unit is updated by adapting or creating an acoustic model based on the correspondence relationship between the phoneme sequence aligned by the alignment unit and the input speech data, and the feature amount of the input speech data. Adapting or creating a language model based on the corrected result text and updating the model storage unit;
A transcription apparatus comprising: When the correction unit does not correct the recognition result text based on an operation by a user, the correction result text is directly output as the correction result text.
The transcription apparatus according to claim 1. A recognition process by the speech recognition unit, a process by which the correction unit creates the correction result text, a process by which the pronunciation giving unit gives the phoneme string, and an alignment unit by the phoneme string and the input speech data A control unit that controls a series of processes including a process of aligning and a process in which the adaptation unit updates the model storage unit to repeat a plurality of times,
The transcription apparatus according to claim 1, further comprising: Computer
A model storage unit for storing an acoustic model and a language model;
Speech recognition means for performing recognition processing based on the feature amount extracted from the input speech data and outputting a recognition result text while referring to the acoustic model and the language model stored in the model storage unit;
Correction means for enabling correction of the recognition result text based on an operation by a user and creating a correction result text obtained as a result of the correction;
Pronunciation giving means for giving a phoneme string corresponding to the pronunciation of the correction result text;
An aligning means for associating and aligning the phoneme sequence given by the pronunciation giving means and the input voice data in time;
The model storage unit is updated by adapting or creating an acoustic model based on the correspondence between the phoneme sequence aligned by the alignment unit and the input speech data and the feature amount of the input speech data. Adapting means for adapting or creating a language model based on the correction result text and updating the model storage unit;
Program to function as.

Images