JP2012181307A - Voice processing device, voice processing method and voice processing program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To increase efficiency in recognition processing of voice data.SOLUTION: A voice processing device includes storage means for storing input voice data, voice division means for dividing voice data stored in the storage means, voice recognition means for recognizing a plurality of partial voice data items generated by division by the voice division means using at least two voice recognition engines and converting them to character data, and integration means for integrating character data as recognition results by the voice recognition means to generate document data.

Description

  The present invention relates to a technology for recognizing speech.

  In the above technical field, as disclosed in Patent Document 1, a technique for dividing and recognizing input voice data is known.

JP 2000-089786 A

  However, in the above prior art, the voice data is sequentially divided by the voice recognition processing means, and the processing efficiency is poor.

  The objective of this invention is providing the technique which solves the above-mentioned subject.

In order to achieve the above object, an apparatus according to the present invention provides:
Storage means for storing the input voice data;
Audio dividing means for dividing the audio data stored in the storage means;
Voice recognition means for recognizing a plurality of partial voice data generated by the division by the voice division means using at least two voice recognition engines and converting them into character data;
Integration means for integrating character data as a recognition result by the voice recognition means to generate document data;
It is provided with.

In order to achieve the above object, the method according to the present invention comprises:
An audio dividing step for dividing the audio data stored in the storage means;
A voice recognition step of recognizing a plurality of partial voice data generated by the division in the voice division step using at least two voice recognition engines and converting the data into character data;
An integration step of generating document data by integrating character data as a recognition result in the voice recognition step;
It is characterized by including.

In order to achieve the above object, a program according to the present invention provides:
An audio dividing step for dividing the audio data stored in the storage means;
A voice recognition step of recognizing a plurality of partial voice data generated by the division in the voice division step using at least two voice recognition engines and converting the data into character data;
An integration step of generating document data by integrating character data as a recognition result in the voice recognition step;
Is executed by a computer.

  ADVANTAGE OF THE INVENTION According to this invention, the recognition process of audio | voice data can be made efficient.

It is a block diagram which shows the structure of the audio processing apparatus which concerns on 1st Embodiment of this invention. It is a block diagram which shows the structure of the audio processing apparatus which concerns on 2nd Embodiment of this invention. It is a block diagram which shows the structure of the speech recognition part which concerns on 2nd Embodiment of this invention. It is a figure which shows the structure of the audio | voice data memorize | stored in the audio | voice processing apparatus which concerns on 2nd Embodiment of this invention. It is a figure which shows the division | segmentation result of the audio | voice data in the audio | voice processing apparatus which concerns on 2nd Embodiment of this invention. It is a figure for demonstrating the division | segmentation method of the audio | voice data in the audio | voice processing apparatus which concerns on 2nd Embodiment of this invention. It is a figure for demonstrating the division | segmentation method of the audio | voice data in the audio | voice processing apparatus which concerns on 2nd Embodiment of this invention. It is a figure which shows the graphical user interface which the audio processing apparatus which concerns on 2nd Embodiment of this invention displays. It is a block diagram which shows the structure of the speech recognition part which concerns on 3rd Embodiment of this invention.

  Hereinafter, exemplary embodiments of the present invention will be described in detail with reference to the drawings. However, the components described in the following embodiments are merely examples, and are not intended to limit the technical scope of the present invention only to them.

(First embodiment)
A speech processing apparatus 100 as a first embodiment of the present invention will be described with reference to FIG. The voice processing device 100 is a device that recognizes voice data and converts it into document data.

  As shown in FIG. 1, the speech processing apparatus 100 includes a storage unit 101, a speech division unit 102, a speech recognition unit 103, and a recognition result integration unit 105. The storage unit 101 stores input voice data. The audio dividing unit 102 reads audio data from the storage unit 101 and divides it into a plurality of partial audio data. Further, the voice recognition unit 103 recognizes the plurality of partial voice data generated by the division by the voice division unit 102 using at least two voice recognition engines 131 to 13n, and converts them into character data. Further, the recognition result integration unit 105 integrates the character data output from the voice recognition unit 103 to generate document data.

  According to the above configuration, since speech data is processed in parallel by a plurality of speech recognition engines, speech recognition can be performed very efficiently.

(Second Embodiment)
A second embodiment of the present invention will be described with reference to FIG. FIG. 2 is a block diagram showing a configuration of the voice processing system 200 according to the present embodiment. In FIG. 2, the audio processing system 200 is connected to a microphone 210, a speaker 220, a display 230, an operation unit (mouse or keyboard) 240, and the like. The voice processing system 200 includes a voice storage unit 201, a voice recognition unit 202, a document processing unit 203, a voice playback unit 207, and an operation unit 240. The voice processing system 200 recognizes the voice input from the microphone 210 and outputs the input voice and the recognition result to the display 230 and the speaker 220, and corrects or edits an error in the document as the recognition result. It is. Of the components of the voice processing system 200, the voice storage unit 201 stores voice data input from the microphone 210. The voice recognition unit 202 recognizes the voice data stored in the voice storage unit 201 and converts it into document data. Further, the document processing unit 203 inserts the document data generated by the voice recognition unit 202 into the prepared GUI form, and generates display data. The generated display data is displayed on the display 230.

  A user operation is accepted from the operation unit 240, and the document data generated by the document processing unit 203 is displayed on the display 230, and the document data is edited and corrected.

  FIG. 3 is a diagram illustrating a detailed configuration of the voice recognition unit 202. The voice recognition unit 202 includes a voice input unit 310, a voice division unit 320, a plurality of recognition engines 331 to 33n, a recognition result combining unit 340, and a document output unit 350.

  The audio input unit 310 reads out audio data (audio file) from the audio storage unit 201 and passes it to the audio dividing unit 320. The audio dividing unit 320 divides the received audio data into partial audio data. The partial voice data is sent to a plurality of recognition engines 331 to 33n, and is subjected to voice recognition processing and converted into character data.

  Here, a recognition process may be added to one partial voice data by a plurality of voice recognition engines. In that case, the most reliable speech recognition result can be adopted. Alternatively, two adjacent partial speech data may be input to each speech recognition engine, and the speech recognition engine may recognize a combination of the two partial speech data. For example, if a series of partial voice data 1 to 3 is generated by division, the voice recognition engine 331 recognizes a combination of the partial voice data 1 and the partial voice data 2, and the voice recognition engine 332 recognizes the partial voice data 2 and the partial voice data. The recognition accuracy may be improved by a method of recognizing the combination of data 3.

  The recognition result combining unit 340 uses a single or a plurality of recognition results and integrates them in time series to create an integrated recognition result. If there is an overlap in the recognition results for the same partial speech data (such as when one partial speech file is recognized by multiple recognition engines using different dictionaries or recognition methods), the recognition results are compared and the one with higher reliability (features) Adopt the one with more amount matching) and create the integrated recognition result. Then, the integrated recognition result is output to the document output unit 350. The document output unit 350 generates a graphic interface that allows the user to edit the document as the recognition result and outputs the integrated recognition result.

  FIG. 4 is a diagram showing audio data stored in the audio storage unit 201. Here, since it is premised on a system that supports creation of minutes, it is assumed that one audio data table 400 is created for one meeting. In addition to the storage sequence number 401 for identifying the conference, the voice data table 400 is given a comment 402 such as a place, a speaker, and contents, and a digital voice together with a time stamp indicating the date and time when each voice file was generated. A file 403 is stored.

  FIG. 5 shows a partial audio data table 500 for managing the divided audio data. Each partial audio data has a length of 10 ms or 1 s, for example, and a sequential number is added together with a time stamp. This makes it possible to accurately integrate recognition results. Each partial speech data is assigned a recognition engine number that identifies a recognition engine for recognizing the data. Furthermore, the partial speech data table 500 has an area for storing the recognition results of the recognition engines 331 to 33n, and specifies the characters indicated by each syllable.

  FIG. 6 is a diagram for explaining a method of dividing audio data. As a method of dividing the audio data, there can be considered a method 601 for equally dividing the audio data before the division including the silent portion, and a method 602 for equally dividing the audio portion excluding the silent portion from the audio data before the division. The method 601 has an advantage that the division processing speed is fast, and the method 602 has an advantage that the voice recognition speed is fast. You may receive the instruction | indication from a user about the space | interval of a division | segmentation. In other words, the method 602 can be said to be a method of searching for a silent section existing in an audio file and dividing the voice at a timing when the silent section starts or ends. You may change a division | segmentation time for every sound section according to the length of a sound section. Further, the dividing process may be performed after connecting the sound sections.

  Further, the ratio of the silent part in the pre-division voice data is calculated in advance, and the method 601 is adopted when the silent ratio is smaller than a predetermined value, and the method 602 is adopted when the silent ratio is larger than the processing value. You may switch to do.

  Further, as shown in FIG. 7, a division method may be employed in which a partial section of the audio of the preceding and succeeding time is provided before or behind the divided audio, or both. According to such an overlap method, each divided sound has a portion where it is overlapped, and each partial sound data can be set longer, so that it is possible to avoid misrecognition due to poor division timing. Furthermore, if partial speech data is set longer, speech recognition can be performed more accurately in consideration of the context before and after. Of course, the non-overlapping method can also be adopted, and the non-overlapping method does not have overlapping sections. Since the same voice is not converted a plurality of times, resources such as CPU consumption time and memory usage can be reduced.

  FIG. 8 is a diagram showing a graphical user interface 800 generated by the document processing unit 203 and displayed on the display 230. This is an example, and the present invention is not limited to this. The graphical user interface 800 includes an audio waveform display field 801, a title display field 802, and a document data display field 803 as a recognition result. Furthermore, the graphical user interface 800 includes a time stamp (elapsed time from the start time of all sounds) display field 804 indicating the start time of the sound corresponding to each document data.

  In addition to the playback button 805 and volume button 806, a repeat button 807 and the like are prepared below the document data display field 803, and are used for playback operation, volume change operation, repeat operation, and the like, respectively. Each character displayed in the document data display field 803 is associated with the position of the voice data that is the material for recognizing the character. Therefore, if the cursor is moved to any position of the document displayed in the document data display field 803 and the playback button 805 is clicked in that state, the audio data corresponding to that position is played back. That is, each sentence displayed in the document data does not represent a voice data unit, but constitutes a table line by line for the convenience of document editing. That is, the audio data is not divided for each row, and there is only one audio data to be reproduced in this graphical user interface 800.

  Since the operation panel is arranged below the document data display field 803 as shown in FIG. 8, there is an advantage that the user can easily reproduce the audio data while checking and editing the document data. In the document data display field 803, the document data is arranged in time series from top to bottom, but may be displayed in time series from the bottom to the top. In that case, it is desirable to arrange the operation button groups 805 to 807 above the document data display field 803. Alternatively, when the user can change the display order of the document data, the position of the operation button may be changed according to the display order setting.

  As described above, according to this embodiment, since voice data is divided and parallel processing is performed by a plurality of voice recognition engines, voice recognition processing can be performed very efficiently.

(Third embodiment)
A speech processing system according to the third embodiment of the present invention will be described with reference to FIG. FIG. 9 is a diagram illustrating an internal configuration of the speech recognition unit 902 included in the speech processing system according to the present embodiment. Since the configuration of the speech processing system other than the speech recognition unit 902 is the same as that of the second embodiment, description thereof is omitted here.

  The speech recognition unit 902 is different from the speech recognition unit 202 in the second embodiment in that it includes one or more speech recognition engines with learning functions 931 to 93n and a recognition engine learning control unit 960. Since other configurations are the same as those of the second embodiment, the same components are denoted by the same reference numerals, and detailed description thereof is omitted.

  The recognition engines 931 to 93n have a function of converting input speech into a character string. The voice is input from the voice dividing unit 320, and the converted character string is input to the recognition result combining unit 340. Each of the recognition engines 931 to 93n has a function of learning the gender and habit of the voice every time the voice is converted and autonomously improving the recognition rate. The recognition engine learning control unit 960 controls the learning performance of each recognition engine. Instead of starting multiple speech recognition engines from the beginning of the system, initially start one speech recognition engine, and after a certain amount of speech recognition processing, duplicate the learned results and start up multiple engines. Also good.

  According to the present embodiment, since the speech recognition engines 931 to 93n are equipped with a learning function, by performing a certain amount of speech recognition processing, they learn gender, wrinkles, and the like to help improve the recognition rate.

(Other embodiments)
As mentioned above, although embodiment of this invention was explained in full detail, the system or apparatus which combined the separate characteristic contained in each embodiment how was included in the category of this invention.

  Further, the present invention may be applied to a system constituted by a plurality of devices, or may be applied to a single device. Furthermore, the present invention can also be applied to a case where an information processing program that realizes the functions of the embodiments is supplied directly or remotely to a system or apparatus. Therefore, in order to realize the functions of the present invention on a computer, a program installed in the computer, a medium storing the program, and a WWW (World Wide Web) server for downloading the program are also included in the scope of the present invention. .

Claims (11)

Storage means for storing the input voice data;
Audio dividing means for dividing the audio data stored in the storage means;
Voice recognition means for recognizing a plurality of partial voice data generated by the division by the voice division means using at least two voice recognition engines and converting them into character data;
Integration means for integrating character data as a recognition result by the voice recognition means to generate document data;
An audio processing apparatus comprising: The voice recognition means
The speech processing apparatus according to claim 1, wherein recognition processing is performed on at least two speech recognition engines on one partial speech data generated by the speech partitioning unit.   The audio processing apparatus according to claim 1, wherein the audio dividing unit equally divides the audio data before division including a silent part.   The audio processing apparatus according to claim 1, wherein the audio dividing unit equally divides an audio portion obtained by removing a silent portion from audio data before division.   The voice processing apparatus according to claim 1, wherein the voice dividing unit receives an instruction from a user and changes a division interval according to the instruction.   6. The voice dividing unit searches for a silent section existing in the voice data, and divides the voice data at a timing when the silent section starts or ends. Voice processing device.   The speech processing apparatus according to claim 6, wherein the voice dividing unit changes a division interval for each voiced section according to a length of a voiced section existing in the voice data.     The voice dividing means calculates a ratio of a silent section in the voice data, and if the silent ratio is smaller than a predetermined value, the voice data before the division is divided equally including a silent part, and the silent ratio is processed. 8. The audio processing apparatus according to claim 1, wherein when the number is larger than the value, the audio part obtained by removing the silent part from the audio data before division is equally divided.   The audio processing according to any one of claims 1 to 8, wherein the audio dividing unit divides the audio data so that end portions of the partial audio data have overlapping audio data. apparatus. An audio dividing step for dividing the audio data stored in the storage means;
A voice recognition step of recognizing a plurality of partial voice data generated by the division in the voice division step using at least two voice recognition engines and converting the data into character data;
An integration step of generating document data by integrating character data as a recognition result in the voice recognition step;
A speech processing method comprising: An audio dividing step for dividing the audio data stored in the storage means;
A voice recognition step of recognizing a plurality of partial voice data generated by the division in the voice division step using at least two voice recognition engines and converting the data into character data;
An integration step of generating document data by integrating character data as a recognition result in the voice recognition step;
A sound processing program for causing a computer to execute.

Images