JP4928514B2 - Speech recognition apparatus and speech recognition program - Google Patents

Description

  The present invention relates to a speech recognition apparatus and a speech recognition program, and more particularly to a speech recognition apparatus and a speech recognition program for realizing highly accurate speech recognition using a continuously updated latest model.

  Conventionally, the announcer's manuscripts read out by news announcers in news programs, etc., are those that the director printed electronic manuscripts submitted by reporters, and were handwritten and revised immediately before or during the broadcast, depending on the length of the broadcast and the flow of the story. Used.

  Also, in speech recognition used for news program subtitle production, in order to deal with new proper nouns and topics, recognition errors may occur by using the electronic manuscript that is the source of the reading manuscript as adaptive learning data for the language model. Is known to be important (see, for example, Patent Document 1).

  In addition, in the conventional caption production system using speech recognition, for example, it took about 8 minutes to learn a language model, and therefore, an electronic manuscript published 10 minutes before the start of broadcasting was used as adaptive learning data (for example, Non-patent document 1).

  Also, conventionally, in order to improve the recognition accuracy of speech recognition by updating the acoustic model and language model without imposing a heavy burden on the user, the acoustic model management server in speech recognition acquires updated acoustic data. A method is known in which the constructed acoustic model is transmitted to a speech recognition device via a network, and the acoustic model that the speech recognition device refers to during speech recognition is updated with the acoustic model transmitted by the acoustic model management server. (For example, refer to Patent Document 2). Thus, Patent Document 2 also mentions the importance of the latest model in speech recognition.

Furthermore, if only a new word is added to the dictionary, the new word is added to the word pronunciation dictionary network without stopping the activated speech recognition decoder, and the N-gram assigned to the unknown word is added. Probability can also be used (see, for example, Patent Document 3 and Non-Patent Document 2).
Japanese Patent No. 3836607 Akio Ando et al., “Broadcast news subtitle production system using voice recognition,” IEICE, Vol. J84-D-II, No. 6, pp. 877-887, 2001.6. JP 2002-91477 A JP 2002-207495 A Ryuichi Nishimura et al., “Examination of speech recognition methods in the speech input Web system w3voice,” Sound Lecture, 1-10-17, pp. 51-52, 2008.3.

  However, in the above-described conventional technology, for example, when a reporter's submission in news or the like is delayed from the time for ending speech recognition, the electronic manuscript is not reflected in the language model, which is one of the causes of recognition errors. It was one. In addition, since the conventional speech recognition system generally includes only one speech recognition decoder, even if the language model is updated to the latest one, the speech recognition decoder is stopped once and manually activated again. It was necessary to load the latest language model.

  Therefore, for example, in speech recognition in subtitle production or the like, if subtitle broadcasting starts and speech recognition that is being activated is stopped, subtitle broadcasting is interrupted, which is not preferable in operation. Also, it takes time to restart the speech recognition decoder after updating the language model.

  Furthermore, in the production of news program subtitles, vocabulary (size 60,000 words) entries are also changed as needed, and it is difficult to dynamically update the language model and pronunciation dictionary while moving one speech recognition decoder.

  In other words, as mentioned above, the importance of the latest model in speech recognition has been mentioned in the past, but here again it is assumed that speech recognition is stopped and restarted, and operation can be continued without interrupting speech recognition. It was not a thing.

  The present invention has been made in view of the above-described problems, and an object thereof is to provide a speech recognition device and a speech recognition program for realizing highly accurate speech recognition using the latest model continuously updated. And

  In order to solve the above problems, the present invention employs means for solving the problems having the following characteristics.

According to the first aspect of the present invention, in the speech recognition apparatus that recognizes an input speech and converts it into a character, model learning means that learns at least one of a language model, a pronunciation dictionary, an acoustic model, and a speech recognition parameter as needed. A model update notification means for notifying that the model learning means has been updated to the latest model, an acoustic analysis means for extracting an acoustic feature quantity of the input speech, and an acoustic feature quantity obtained by the acoustic analysis means, A plurality of speech recognition decoders that read a language model, a pronunciation dictionary, an acoustic model, and speech recognition parameters that are stored in advance or updated by the model learning unit, and perform speech recognition of the acoustic feature amount; and a plurality of speech recognition decoders Among them, a speech recognition decoder that performs speech recognition based on update information notified by the model update notification means. -Option have a decoder control means for said decoder control means, in addition to the speech recognition decoder running the old models, and activates the voice recognition decoder latest models simultaneously, acoustic features obtained from the acoustic analysis means The voice recognition decoder is switched to the latest model at the timing of detecting the utterance start included in the volume .

According to the first aspect of the present invention, it is possible to realize highly accurate speech recognition using the latest model continuously updated. In addition, high-accuracy speech recognition can always be realized continuously using the latest model.

The invention described in claim 2 is characterized in that the decoder control means causes all of the plurality of speech recognition decoders to sequentially read and restart the latest model without interruption.

According to the second aspect of the present invention, it is possible to quickly update to the latest model without interrupting speech recognition and perform speech recognition using the latest model.

According to a third aspect of the present invention, after the decoder control means restarts, the speech recognition decoder recognizes the input speech, and the speech features included in the acoustic feature quantity obtained from the acoustic analysis means. It is characterized in that the starting edge is sequentially received at the timing when it is detected .

According to the third aspect of the present invention, it is possible to reduce the total processing time of complicated speech recognition that requires processing time while using the latest model.

According to a fourth aspect of the present invention, there is provided character correction means for correcting a speech recognition result, outputting the corrected history information to the model learning means, and using it as learning data.

According to the fourth aspect of the present invention, the possibility of the same voice recognition error occurring can be reduced.

According to a fifth aspect of the present invention, there is provided a speech recognition program for causing a computer to execute speech recognition processing for recognizing input speech and converting it into a character. The computer comprises a language model, a pronunciation dictionary, an acoustic model, and speech recognition. Model learning means for learning at least one of parameters as needed, model update notification means for notifying that the model has been updated to the latest model, acoustic analysis means for extracting acoustic feature quantities of the input speech, A plurality of speech recognitions that perform speech recognition of the acoustic feature values by reading the acoustic feature values obtained by the analysis means and the language model, pronunciation dictionary, acoustic model, and speech recognition parameters stored in advance or updated by the model learning means Among the plurality of speech recognition decoders, the decoder is notified by the model update notification means. Based on the update information, the performs selection of the speech recognition decoder for speech recognition, in addition to the speech recognition decoder running the old models, and activates the voice recognition decoder latest models simultaneously, the acoustic analysis means To function as decoder control means for switching to the latest model speech recognition decoder at the timing when the utterance start point included in the acoustic feature value obtained from the above is detected .

According to the fifth aspect of the invention, it is possible to realize highly accurate speech recognition using the latest model continuously updated. Moreover, the data classification process can be easily realized by installing the program.

  According to the present invention, highly accurate speech recognition can be realized using the latest model updated continuously.

<Outline of the present invention>
The present invention simultaneously activates another speech recognition decoder for reading the latest updated model without interrupting speech recognition processing in addition to the speech recognition decoder already activated by reading an old model, By switching the speech recognition decoder for performing the latest model to the latest model, high-accuracy speech recognition is always realized continuously using the latest model.

  Hereinafter, preferred embodiments of a voice recognition device and a voice recognition program according to the present invention will be described with reference to the drawings.

<Voice recognition device: functional configuration example>
FIG. 1 is a diagram illustrating an example of a functional configuration of the speech recognition apparatus according to the present embodiment. A speech recognition apparatus 10 shown in FIG. 1 includes an acoustic analysis unit 11, a decoder control unit 12, speech recognition decoders 13-1 and 13-2, a character correction unit 14, a model learning unit 15, and a model update notification unit. 16, storage means 17, and learning data 18.

  The acoustic analysis unit 11 extracts an acoustic feature amount from the input voice. As the acoustic feature amount, for example, various acoustic feature amounts such as frequency characteristics, sound power, and sex attributes are extracted. In addition, these feature quantities can be used in a general speech recognition technique. For example, about 12-dimensional Mel Frequency Cepstrum Coefficients (MFCC) (for example, Shikano et al., “ Voice recognition system "(Ohm, 2001, etc.)), feature quantities obtained by quantifying the shape of the vocal tract, such as linear prediction coefficients, feature quantities such as prosody (pitch, inflection, etc.), and feature quantities thereof By analyzing statistical information such as the average value and variance of the image, various feature quantities can be acquired. The acoustic analysis unit 11 outputs various acoustic feature amounts obtained by the analysis to the decoder control unit 12.

  The decoder control means 12 activates a plurality of speech recognition decoders 13 (in the example shown in FIG. 1, speech recognition decoders 13-1 and 13-2) and the like, selects which speech recognition decoder 13 to select (speech → character Control for acquiring the acoustic recognition result, such as whether to perform decoding).

  Specifically, the decoder control unit 12 is performing speech recognition among the plurality of speech recognition decoders 13 in accordance with update information notified from the model update notification unit 16 described later and indicating that the model has been updated. Select any speech recognition decoder that is not present and load the selected speech recognition decoder with the latest model and restart it. In addition, the decoder control unit 12 transmits the acoustic feature amount to the speech recognition decoder 13 in charge of speech recognition, and outputs the obtained speech recognition result such as character information to the character correction unit 14.

  The speech recognition decoder 13 reads and activates all of the latest language model, pronunciation dictionary, acoustic model, and speech recognition parameters that have been stored in the storage unit 17 at that time or updated by the model learning unit 15 described later. And voice recognition is possible.

  When the speech recognition decoder 13 acquires the acoustic feature amount from the decoder control unit 12, the speech recognition decoder 13 sequentially performs speech recognition and outputs a speech recognition result such as character information to the decoder control unit 12. In the example of FIG. 1, two speech recognition decoders are provided. However, the present invention is not limited to this, and three or more speech recognition decoders may be provided. Note that the sequential speech recognition can use a conventional method such as an early-determined type unit as disclosed in Japanese Patent No. 3834169, for example.

  The character correction means 14 is input with an accurate character, for example, for an error such as a personal name by a user's check on the speech recognition result obtained by the decoder control means 12, an automatic sentence proofreading process, or the like. Convert a given part of the corresponding sentence. In addition, the character correction means 14 can perform addition and deletion of a character by inputting instruction information.

  When the learning data 18 is newly updated with the latest text, voice, or the like, the model learning unit 15 performs processing of the voice recognition decoder 13 at a timing such as switching at a predetermined time, data update, program change, or the like. Is completely asynchronous, and automatically or manually learns at least one of the language model, pronunciation dictionary, acoustic model, and parameter files (speech recognition parameters) used for speech recognition, etc. to update the data. Do.

  Note that the speech recognition parameters include variables that adjust the accuracy and processing speed of speech recognition, such as the maximum number of words that should be retained in the speech recognition process, weighting factors that adjust the balance between the scores of the language model and the acoustic model, etc. A list of

  As a result, the model stored in the storage unit 17 is updated to the latest language model, pronunciation dictionary, acoustic model, and parameter file. Further, the model learning unit 15 outputs a signal indicating that the model has been learned to the model update notification unit 16.

  The model update notification unit 16 outputs an update signal related to the data update input by the model learning unit 15 to the decoder control unit 12. Here, the update signal is data identification information indicating which data has been updated among the language model, pronunciation dictionary, acoustic model, and parameter file, an update date, an update version, and the like.

  Thereby, for example, in the embodiment shown in FIG. 1, for example, the decoder control means 12 first activates the speech recognition decoder 13-1 by reading the learning data 18 of the latest model. When the learning data 18 of the next latest model is generated, the decoder control means 12 receives notification of update information from the model update notification means 16 and sends the learning data 18 of the latest model to the speech recognition decoder 13-2. After confirming that it has been read and activated and is in a recognizable state, the speech recognition processing target is recognized from the speech recognition decoder 13-1 at a predetermined timing based on, for example, the acoustic feature amount obtained from the acoustic analysis means 11. Switch to the decoder 13-2.

  In addition, when both the speech recognition decoders 13-1 and 13-2 use the latest model, the decoder control means 12 can alternately perform speech recognition processing for each sentence using both.

  The accumulating unit 17 accumulates data necessary for realizing the speech recognition processing in the present embodiment, and performs reading according to the necessity of the speech recognition processing and writing according to the necessity of the model learning unit 15. Specifically, the accumulated data is a model that is stored in advance or automatically or manually updated necessary for the speech recognition processing in the speech recognition decoder 13, and includes, for example, a language model, a pronunciation dictionary, an acoustic model, and a speech recognition parameter. It is all.

  Here, as the language model, for example, a general N-gram model expressing the ease of connection between words can be used, and for example, the word “global warming” can be used next to the word “earth”. The probability that “is connected” can be expressed numerically as 0.8 mag.

  The pronunciation dictionary is a file in which the pronunciation of each word is represented by a combination of vowels and consonants. For example, the pronunciation of the word “Earth” is described as “/ chi ki u: /” or the like.

  The acoustic model represents the frequency characteristics of each vowel / consonant voice and can be represented by a general hidden Markov model (HMM).

  The speech recognition parameters are a list of variables that adjust the accuracy and processing speed of speech recognition, such as the maximum number of words that should be retained during the speech recognition process and weighting factors that adjust the balance of each score between the language model and the acoustic model. . In addition, the part that learns the latest model is included in the speech recognition system, or it can be activated independently and transmitted to the speech recognition system by some communication means. I do not care.

  The learning data 18 stores various data relating to a predetermined field such as text and voice. In addition, when the learning data 18 is updated in various data and the content corresponds to what is currently voice-recognized, or when various models are updated, the learning and the data itself are model-learned. Output to means 15.

<Regarding Update and Control Method of Speech Recognition Decoder 13 in Decoder Control Unit 12>
Here, a method for updating and controlling the speech recognition decoder 13 in the decoder control means 12 will be described.

<Decoder update: Example 1>
The decoder control means 12 starts the speech recognition decoder 13 at the same time, and switches to the latest model speech recognition decoder without interruption. The decoder control means 12 receives from the model update notification means 16 a notification of update information indicating that the language model and the pronunciation dictionary have been automatically (or manually) updated with the latest electronic manuscript submitted, for example. If the recognition decoder 13-1 is executing speech recognition, the speech recognition decoder 13-2 is newly activated with the latest model.

  Then, after confirming that the speech recognition decoder 13-2 is in a recognizable state, the decoder control unit 12 performs speech recognition at a predetermined timing such as a non-speech interval in the input speech acquired by the acoustic analysis unit 11, for example. Is switched from the speech recognition decoder 13-1 to the speech recognition decoder 13-2.

  In addition, the decoder control means 12 can continuously perform speech recognition using the latest model by repeatedly updating the model and starting, selecting and switching the decoder in the same manner.

<Decoder update: Example 2>
The decoder control unit 12 activates all of the plurality of speech recognition decoders 13 in advance with the latest model at that time, and predetermined timing based on the acoustic feature amount obtained from the input speech (for example, every sentence, 1 news) The speech recognition processing is performed by arbitrarily switching a plurality of speech recognition decoders for each theme, for each program, every predetermined time, and the like.

  Next, when the speech recognition model is updated to the latest model, a certain speech recognition decoder is stopped from among a plurality of activated speech recognition parameters, and the model is activated after the model is updated to the latest state. Let Similarly, the speech recognition decoder that has not been updated to the latest model is also updated after being sequentially updated.

  Note that the speech recognition decoder being updated does not perform speech recognition processing at that time, and performs speech recognition with the latest model after restarting. As a result, multiple speech recognition decoders can be started in parallel, reducing the load on the decoder even when speech that takes a long time for speech recognition is input, and improving the overall speech recognition processing speed. Can be made.

  According to the above-described embodiment, it is possible to provide a speech recognition device for realizing highly accurate speech recognition using the latest model continuously updated. By switching the speech recognition decoder that performs speech recognition to the latest model without interrupting speech recognition, high-accuracy speech recognition can always be realized continuously using the latest model.

  As a result, for example, when using speech recognition for the purpose of providing subtitles in real time on a live TV program, even if speech recognition is already in operation during broadcasting, speech recognition is stopped even for a moment. Instead, the latest model can always be read and high-accuracy speech recognition can be realized continuously.

<Hardware configuration>
Here, as described above, the speech recognition apparatus 10 can perform speech recognition processing according to the present invention with a dedicated device configuration, but an execution program that can cause a computer to execute speech recognition processing in each configuration described below. For example, the speech recognition processing according to the present invention can be realized by installing the program in a general-purpose personal computer, workstation, or the like.

  Here, an example of a hardware configuration of an executable computer in the present invention will be described with reference to the drawings. FIG. 2 is a diagram illustrating an example of a hardware configuration capable of realizing the speech recognition process according to the present invention.

  2 includes an input device 21, an output device 22, a drive device 23, an auxiliary storage device 24, a memory device 25, a CPU (Central Processing Unit) 26 for performing various controls, and a network connection device. 27, and these are connected to each other by a system bus B.

  The input device 21 has a pointing device such as a keyboard and a mouse operated by a user, and inputs various operation signals such as execution of a program from the user. The output device 22 has a display for displaying various windows and data necessary for operating the computer main body for performing the speech recognition processing according to the present invention. Can be displayed.

  Here, in the present invention, the execution program installed in the computer main body is provided by a portable recording medium 28 such as a USB (Universal Serial Bus) memory or a CD-ROM, for example. The recording medium 28 on which the program is recorded can be set in the drive device 23, and the execution program included in the recording medium 28 is installed in the auxiliary storage device 24 from the recording medium 28 via the drive device 23.

  The auxiliary storage device 24 is a storage means such as a hard disk, and can store an execution program according to the present invention, a control program provided in a computer, etc., and perform input / output as necessary.

  The memory device 25 stores an execution program read from the auxiliary storage device 24 by the CPU 26. The memory device 25 includes a ROM (Read Only Memory), a RAM (Random Access Memory), and the like.

  The CPU 26 performs processing of the entire computer, such as various operations and data input / output with each hardware component, based on a control program such as an OS (Operating System) and an execution program read and stored by the memory device 25. By controlling the above, it is possible to realize each processing in the voice recognition described above. Various kinds of information necessary during the execution of the program can be acquired from the auxiliary storage device 24 and can also be stored.

  The network connection device 27 acquires an execution program from another terminal connected to the communication network by connecting to a communication network or the like, or an execution result obtained by executing the program or an execution in the present invention The program itself can be provided to other terminals.

  With the hardware configuration as described above, it is possible to realize a highly accurate voice recognition process at a low cost without requiring a special device configuration. In addition, by installing an execution program (speech recognition program or the like), speech recognition processing can be easily realized.

<Voice recognition processing procedure>
Next, an example of a speech recognition processing procedure in the present embodiment will be described using a flowchart. In the following description, it is assumed that up to D speech recognition decoders can be activated in the processing of the entire speech recognition apparatus that supports automatic model updating, model learning according to learning data update, update processing, speech Since the selection and switching process of the speech recognition decoder that performs additional activation and recognition of the recognition decoder is performed in an asynchronous parallel operation, these processes will be described separately.

<Model learning / update process>
First, a model learning / updating process procedure according to the present embodiment will be described with reference to a flowchart. FIG. 3 is a flowchart illustrating an example of a model learning / updating process procedure according to the present embodiment.

  In FIG. 3, first, when the model learning means receives an update event indicating that the learning data such as text and speech has been updated from the outside (S01), the model (language model / pronunciation dictionary) stored in the storage means. The latest model is learned for the acoustic model and various parameter files), and the model is updated (S02).

  Next, the model update notification means notifies the decoder control means of a model update event (model update information) indicating that the model has been updated (S03). In addition, the decoder control means selects an updatable speech recognition decoder from among the plurality of speech recognition decoders based on the update information obtained in S03, and loads and activates the selected speech recognition decoder with the latest model (S04). . The speech recognition decoder updated at this time is a speech recognition decoder that has not yet been activated as a spare device, or a speech recognition decoder that is sequentially selected in a predetermined order from a plurality of decoders that are activated in parallel. In response, stop, update, and restart processes are performed.

  It should be noted that the updated latest model d ′ is assumed to be “d ′ = (d + 1)% D” in order to specifically describe the processing described later. This expression is expressed as a remainder d 'obtained by adding 1 to the number d for identifying a speech recognition decoder and dividing by D (the maximum number of speech recognition decoders that can be activated).

<Additional activation of speech recognition decoder and recognition target switching processing>
Next, additional activation of the speech recognition decoder and recognition target switching processing will be described using a flowchart.

  FIG. 4 is a flowchart showing an example of additional activation of the speech recognition decoder and recognition target switching processing. In the processing shown in FIG. 4, when the operation of the entire speech recognition apparatus is started, initial value setting is first performed (S11). Specifically, 0 is set in the number d of the speech recognition decoder, the latest model (language model, pronunciation dictionary, acoustic model, parameter file, etc.) at that time is read, and the speech recognition decoder d is activated (S12). .

  Here, when the voice to be recognized starts to be input (S13), the extraction of the acoustic feature quantity by the acoustic analysis is started (S14), and the utterance start point of, for example, a human voice is detected therefrom (S15).

  Here, if learning / updating processing of a model operating asynchronously and in parallel with speech recognition is performed, as shown in FIG. 3 described above, 1 is added to the number d of the speech recognition decoder and D The model (language model, pronunciation dictionary, acoustic model, parameter file) is learned and updated with the remainder divided by d ′ = (d + 1)% D, and a model update event is notified to the decoder control means 12. Assume that the speech recognition decoder d ′ is activated with its latest model.

  In this state, the decoding control means 12 determines whether or not the speech recognition decoder d ′ has been activated (S16). If the speech recognition decoder d ′ has been activated (YES in S16), speech recognition is performed. The decoder d is stopped, the speech recognition decoder d is updated by the speech recognition decoder d ′, and the decoder number d in charge of speech recognition processing is switched to d ′ (S17). Further, in the process of S16, when the speech recognition decoder d 'has not been activated (NO in S16), the number d of the speech recognition decoder remains unchanged.

  Then, the decoder control means 12 receives the acoustic feature quantity of the input speech from the acoustic analysis means and transmits it to the speech recognition decoder d (S18). The speech recognition decoder d searches for a correct word (S19), and transmits a recognition result character string from the speech recognition decoder d to the decoder control unit (S20). Then, the decoder control means 12 outputs the character string of the recognition result to the outside (S21), and this is used for applications such as caption production for live broadcast programs.

  Here, it is determined whether or not the input signal is the utterance end (S22), and if the input voice has not reached the utterance end (NO in S22), the acoustic control is received by the decoder control means 12 in the process of S18. Returning to the transmission to the voice recognition decoder d, the voice recognition is repeated until the end of the utterance without changing the voice recognition decoder.

  Also, in the process of S22, if the input voice has reached the end of the utterance (YES in S22), it is then determined whether or not the entire voice recognition process is to be terminated (S23), and the voice recognition is terminated. If not (NO in S23), in order to continue the speech recognition, the process returns to the detection of the utterance start end in the processing of S15, and the subsequent processing is repeated until the speech recognition processing is completed. When the voice recognition is finished (YES in S23), the entire voice recognition process is finished.

  With the processing procedure described above, it is possible to provide a speech recognition program for realizing highly accurate speech recognition using the latest model updated continuously. Specifically, by switching the speech recognition decoder that performs speech recognition to the latest model without interrupting speech recognition, high-accuracy speech recognition is always realized continuously using the latest model. . For example, when using voice recognition for the purpose of giving subtitles to a live TV program in real time, even if voice recognition is already in operation during broadcasting, voice recognition is not stopped even momentarily, The latest model is always read and highly accurate speech recognition is possible continuously.

<About start and end detection of speech>
In the processing described above, the processing procedure for detecting the start and end of speech performed during acoustic analysis is, for example, online speech segment detection (eg, Japanese Patent Application Laid-Open No. 2007-233148) with little time delay due to endless phoneme recognition. Can be used. This outline will be described below.

<Speech segment detection by phoneme recognition>
In speech segment detection for real-time speech recognition, even if some non-speech segments may be mistaken as speech segments, rather than making fine speech / non-speech determinations in units of frames, It is important to cut the speech into sections of an appropriate length to contribute to the improvement of the recognition rate. In addition, for caption display, the delay time from voice input to voice start / end detection is also required to be as small as possible.

  For example, in speech recognition in a caption production system, phoneme recognition using a gender-dependent gender-dependent acoustic model is performed endlessly, taking into account not only the sound power but also the frequency characteristics, and the utterance interval is detected from the likelihood at that time. I have to. Since phoneme recognition can be applied regardless of tasks, it is simpler than a method using a task-dependent language model, and even if the acoustic models are operated in parallel with men and women, the amount of computation hardly becomes a problem.

  Therefore, in the present embodiment, gender parallel phoneme recognition that allows transition between men and women and is common to pruning is always performed, and the beginning and end of the utterance are detected early using the ratio of cumulative phoneme likelihood. Thereby, in the voice segment detection experiment for the news program, the FRR (False Rejection Rate: the ratio of the voice segment erroneously determined as non-voice) by the short-time power was 4.6%, It has been confirmed that the above-described method is as very small as 0.53%, and the delay time until the start / end detection of the utterance is sufficiently short.

  It should be noted that the processing procedure for detecting the start and end of the utterance described above can be operated by any known utterance section detection method, and the processing procedure for searching for the correct word in the speech recognition decoder d is any known speech recognition. It is possible to operate in a manner.

<Specific Example of Speech Recognition Processing>
Next, a specific embodiment of the voice recognition process described above will be described with reference to the drawings. FIG. 5 is a diagram illustrating a specific example to which the speech recognition method according to the present embodiment is applied. FIG. 5 shows an example of a caption production system 30 that uses the speech recognition apparatus 10. Specifically, the caption production system 30 switches input of program audio of a direct method (for example, reading a manuscript by an announcer, a reporter's site report, etc.) and a repetitive audio of a risk peak method (for example, an interview, etc.), and A / D The input voice obtained by conversion or the like is input to the voice recognition device 10 described above.

  The speech recognition apparatus 10 uses a speech recognition decoder A and B to create a male by using various model data such as a language model, pronunciation dictionary, etc., which is learned from news electronic manuscript as learning data, and an unspecified speaker acoustic model. Voice recognition is performed using an HMM or a female HMM, and subtitles are confirmed and corrected, and a character string is displayed on the subtitle screen.

  As described above, by applying the speech recognition method according to the present invention, the speech recognition decoder for performing speech recognition is switched to the latest model without interrupting speech recognition at all in the caption production system 30, so that the latest model is always updated. High-accuracy speech recognition can be continuously realized by using. In other words, users who perform speech recognition processing (program producers, etc.) do not need to worry about whether the model is the latest one, and it is guaranteed that voice recognition is always activated automatically with the latest model. The

<Results of comparison with conventional method in this embodiment>
Here, in order to investigate the effect of the speech recognition apparatus that supports automatic model update, learning is performed when the electronic manuscript corresponding to each news item in the broadcast program is not adaptively learned (model one hour before the broadcast). The results of examining the effect of this case (the model learned immediately before broadcasting) by a caption production experiment by voice recognition (real-time manual correction of recognition errors) will be described.

  FIG. 6 is a diagram illustrating an example of the effect of the update. FIG. 6 shows the effect of updating the language model and the pronunciation dictionary as an example. For example, as a result of an experiment with two news programs, as shown in FIG. 6, when a language model (vocabulary size 60,000 words) is adaptively learned with a related manuscript, a test set perplexity that is an indicator of the complexity of the language model. Tee and unknown word rate (percentage of words not registered in the pronunciation dictionary) are greatly reduced, trigram hit rate (language model coverage) is increased, and speech recognition errors are reduced to about 1/3. It was. Even in the model one hour before the broadcast, there was almost no subtitle error, but proper nouns such as personal names became unknown words, requiring manual correction. Therefore, it can be said that the decoder control method of the present invention with low possibility of remaining errors in subtitles and small display delay of subtitles is preferable in operation.

  As described above, according to the present invention, highly accurate speech recognition can be realized using the latest model continuously updated. In addition, according to the present invention, the voice recognition decoder in charge of the voice recognition processing is automatically switched to the latest model without interrupting the voice recognition at all, so that the voice recognition using the latest model is always performed with high accuracy. Is realized continuously.

  Specifically, the present invention continuously realizes highly accurate speech recognition using the latest model by switching the speech recognition decoder that performs speech recognition to the latest model without interrupting speech recognition. To do. For example, when using voice recognition for the purpose of giving subtitles to a live TV program in real time, even if voice recognition is already in operation during broadcasting, voice recognition is not stopped even momentarily, The latest model can always be read and highly accurate voice recognition can be realized continuously.

  Although the preferred embodiment of the present invention has been described in detail above, the present invention is not limited to the specific embodiment, and various modifications, within the scope of the gist of the present invention described in the claims, It can be changed.

It is a figure which shows an example of a function structure of the speech recognition apparatus in this embodiment. It is a figure which shows an example of the hardware constitutions which can implement | achieve the speech recognition process in this invention. It is a flowchart which shows an example of the learning / update process procedure of the model in this embodiment. It is a flowchart which shows an example of the additional starting of a speech recognition decoder, and a recognition target switching process. It is a figure which shows the specific Example to which the speech recognition method in this embodiment is applied. It is a figure which shows an example of the effect of an update.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 Speech recognition apparatus 11 Acoustic analysis means 12 Decoder control means 13 Speech recognition decoder 14 Character correction means 15 Model learning means 16 Model update notification means 17 Storage means 18 Learning data 21 Input device 22 Output device 23 Drive device 24 Auxiliary storage device 25 Memory Device 26 CPU
27 Network connection device 28 Recording medium 30 Subtitle production system

Claims (5)

In a speech recognition device that recognizes input speech and converts it into characters,
Model learning means for learning at least one of a language model, pronunciation dictionary, acoustic model, and speech recognition parameter as needed;
Model update notification means for notifying that the model learning means has been updated to the latest model;
Acoustic analysis means for extracting an acoustic feature of the input speech;
A plurality of acoustic features obtained by the acoustic analysis unit and a language model, pronunciation dictionary, acoustic model, and speech recognition parameter that are stored in advance or updated by the model learning unit are read, and speech recognition of the acoustic feature amount is performed. A speech recognition decoder;
Wherein the plurality of speech recognition decoder, based on the update information notified by the model update notification means, we have a decoder control means for selecting the speech recognition decoder for the speech recognition,
The decoder control means simultaneously activates the latest model speech recognition decoder in addition to the speech recognition decoder activated in the old model, and detects the utterance start included in the acoustic feature amount obtained from the acoustic analysis means. A speech recognition apparatus characterized by switching to the latest model speech recognition decoder . The decoder control means includes
The speech recognition apparatus according to claim 1 , wherein all of the plurality of speech recognition decoders sequentially read and restart the latest model without interruption. The decoder control means includes
After the restarting, the respective speech recognition decoders sequentially recognize the input speech at the timing when the utterance start point included in the acoustic feature amount obtained from the acoustic analysis unit is detected. Item 3. The speech recognition device according to Item 2 . Correct speech recognition result, according the modified history information to any one of claims 1 to 3, characterized in that it has a character modifying means for use as an output to the learning data to the model learning unit Voice recognition device. In a speech recognition program for causing a computer to execute speech recognition processing for recognizing input speech and converting it into characters,
Computer
Model learning means for learning at least one of a language model, a pronunciation dictionary, an acoustic model, and a speech recognition parameter as needed;
Model update notification means for notifying that the model learning means has been updated to the latest model,
Acoustic analysis means for extracting an acoustic feature of the input speech;
A plurality of acoustic features obtained by the acoustic analysis unit and a language model, pronunciation dictionary, acoustic model, and speech recognition parameter that are stored in advance or updated by the model learning unit are read, and speech recognition of the acoustic feature amount is performed. A speech recognition decoder, and
Among the plurality of speech recognition decoders, a speech recognition decoder for performing speech recognition is selected based on update information notified by the model update notification means, and in addition to a speech recognition decoder activated in an old model Voice recognition for simultaneously starting a speech recognition decoder of the latest model and functioning as a decoder control means for switching to the speech recognition decoder of the latest model at the timing when the utterance start point included in the acoustic feature value obtained from the acoustic analysis means is detected program.