JP2010009446A - System, method and program for retrieving voice file - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To facilitate registration of a new word and input of a keyword without being conscious of the contents of a voice recognition dictionary as far as possible in order to connect voice recognition to succeeding language processing or retrieval. <P>SOLUTION: In the registration of a new word or the input of a keyword, "reading" is input by a user at first. The reading is converted from pronunciation into notation by the same language model as a language model for voice recognition, and thereby a Kana-Kanji notation is obtained. Then, the obtained Kana-Kanji notation is properly compared with a corrected word and an original character string to identify the unknown word of the voice recognition dictionary. A converted keyword can be used for retrieving retrieval data formed by voice recognition of a voice file. The unknown word portion can be properly registered in the voice recognition dictionary. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to processing of text created by voice recognition, and more particularly to retrieval of text created by voice recognition from a voice file.

  In recent years, we have recorded conversations at call centers and sales offices for the purpose of business integrity, etc., and performed voice recognition on the recorded voices, and then processed search and text mining. An attempt has been made.

  In text mining and keyword search, language processing using a dictionary is generally performed. At that time, when an unknown word or a new word is encountered, the word is newly registered in the dictionary, and the dictionary is updated.

  In particular, in Japanese, there are large variations in notation such as whether to use kana, kanji, or alphabet, so it is common to register synonyms of various notations in the language processing dictionary. However, in the speech recognition, if synonyms having the same pronunciation are registered, there is no means for distinguishing which one is output, which causes a reduction in recognition accuracy due to an increase in vocabulary. Therefore, it is required to register only a single notation as much as possible and to take care to reduce the entropy of the language model.

  For example, as a result of the speech recognition of “Kikumisai”, there can be two types of “structured bonds” and “structured bonds”, but they are unified into “structured bonds”.

  Also, in speech recognition and subsequent processing, it is often necessary to use different units for each convenience, but in this case, the problem is further complicated. In other words, it is necessary to perform dictionary registration and keyword input for language processing, always conscious of what units and notations are used in the speech recognition in the previous stage.

  However, since the details of the language model with respect to the result of speech recognition are unknown, it is difficult for the user to perform dictionary registration and keyword input for language processing suitable for speech recognition in the previous stage.

  Japanese Patent Application Laid-Open No. 7-152756 discloses that in a process using a dictionary such as kana-kanji conversion, speech synthesis, and speech recognition, the dictionary for each purpose is put together to facilitate maintenance and reduce capacity.

  Japanese Patent Laid-Open No. 8-314915 discloses an information expression mode conversion device that constitutes dictionary means corresponding to a plurality of conversion functions and performs a search using a plurality of indexes for the dictionary means.

  Japanese Patent Application Laid-Open No. 2000-339305 discloses a technique for creating a document by improving input accuracy and input operability by using two input methods of keyboard input and voice input.

  Japanese Patent Application Laid-Open No. 2004-145014 discloses an automatic voice response device and an automatic voice response method that can easily manage grammar and dictionary and can respond faithfully to input voice.

  [NAGATA 1999] (details below) discloses techniques related to unknown word search.

  [MORI 1999] (details below) discloses kana-kanji conversion technology using a probabilistic model.

JP 7-152756 A JP-A-8-314915 JP 2000-339305 A JP 2004-145014 A [NAGATA 1999] Nagata, M.: A part of speech estimation method for Japanese unknown words using a statistical model of morphology and context, Proc. Of the 37th ACL, pp277-284, 1999 [MORI 1999] Mori Shinsuke, Tsuchiya Masami, Yamaji Osamu, Nagao Makoto. Kana-Kanji Conversion by Stochastic Model. IPSJ Transactions. Vol.40, No.7, pp.2946-2953. 1999.

  Even if the above-described conventional techniques are combined, the problem of the necessity of keyword input in consideration of the language model of speech recognition in order to appropriately search the speech file is still not solved.

  Accordingly, an object of the present invention is to connect speech recognition to subsequent language processing and search, and to easily register new words and input keywords without being conscious of the contents of the speech recognition dictionary as much as possible.

  Normally, new word registration and keyword input are performed using the computer's kana-kanji conversion function. In other words, “reading” is input at that stage, but the reading is discarded after the notation is confirmed. On the other hand, a dictionary for speech recognition, that is, a language model, always requires reading, and maintains notation / reading and word co-occurrence probabilities.

  The present invention has been made paying attention to this point, and a new word registration and search in the subsequent stage is performed on the same language model as a speech recognition dictionary, that is, a language model, used for attaching a text keyword to an audio file. It is used for conversion of “reading” of time.

  In order to convert an audio file into text, an analog audio signal is first converted into a digital signal, from which a frequency domain signal is generated for discrete Fourier transform in a predetermined time window, from which a logarithmic spectrum is generated. Further, a cepstrum is generated by discrete cosine transform. Further, the amplitude, fundamental frequency, power spectrum envelope, and the like of the waveform are extracted from the cepstrum by a known technique, and these become acoustic feature quantities. The acoustic model is used to determine which word each part of the utterance may be based on this acoustic feature amount. At this time, the acoustic model uses a probabilistic model to obtain probable words stochastically.

  Thus, once a possible word sequence is determined, the language model is applied to it. The language model uses a probability model to predict and determine what word string is most likely from the context (adjacent words).

  According to the present invention, in the new word registration and keyword input, first, “reading” is input by the user. This reading is converted into pronunciation and notation by the same language model as the language model for speech recognition, thereby obtaining kana-kanji notation. Next, an unknown word in the speech recognition dictionary is identified by comparing the obtained kana / kanji notation with the corrected word and the original character string as appropriate. The conversion result keyword can be used to search index data formed by voice recognition of a voice file. The unknown word part can be registered in the speech recognition dictionary as appropriate.

  According to the present invention, it is possible to easily register new words and input keywords without being aware of the contents of the speech recognition dictionary in speech recognition and subsequent language processing and search.

  The configuration and processing of an embodiment of the present invention will be described below with reference to the drawings. In the following description, the same elements are referred to by the same reference numerals throughout the drawings unless otherwise specified. It should be understood that the configuration and processing described here are described as an example, and the technical scope of the present invention is not intended to be limited to this example.

 FIG. 1 is a schematic block diagram showing an embodiment of a hardware configuration for carrying out the present invention. The configuration shown in FIG. 1 is preferably a plurality of client systems 110 for an individual user to perform a search, a voice storage server 120 for storing voice data files so as to be searchable, and voice recognition from voice data. Then, according to the acoustic model and the language model, the speech recognition server 130 for creating index data, and the client system 110, the speech storage server 120, and the network 140 for connecting the speech recognition server 130 are included.

  The network 140 can use any connection form such as LAN, WAN, the Internet, and an intranet.

  In addition, such a configuration in which the client system 110, the voice storage server 120, and the voice recognition server 130 are separately separated and connected by the network 140 is not essential, and the voice data and the index are copied to the local system. Accordingly, the system of the present invention can be configured stand-alone.

  A web browser 112 is installed in the client system 110. The web browser 112 has a function of interpreting an in-content script, such as JavaScript (trademark), accepts an input from the user, and communicates on the client system 110 side and the voice storage server 120 side communication interface 122. The search operation is performed in cooperation with a program such as Perl or PHP on the voice storage server 120 side. A combination of a general HTML embedded form and CGI can also be used without using JavaScript (trademark).

  Alternatively, on the voice storage server 120 side, a search function may be constructed using a server-side Java (trademark) mechanism such as a servlet or JSP.

  The voice storage server 120 has voice data 124 and search data 126 for searching for individual voice files stored in the voice data 124. Here, the audio data 124 is, for example, a call center conversation, a broadcast program, podcast data, or the like. Since it is difficult to search in the form of the voice data as it is, the voice data is sequentially sent to the voice recognition server 130 via the communication interface 122, the network 140, and the communication interface 136.

  The speech recognition server 130 has an acoustic model 132 and a language model 134 for recognizing speech files and generating index text. The generated index text is provided as index data 126 to the voice storage server 120 via the communication interface 136 and the communication interface 122.

  The search function on the voice storage server 120 side can be realized by a database search system such as PostgreSQL or MySQL.

  The language model 134 of the voice recognition server 130 can also be accessed from the client system 110 and the voice storage server 120 via the network 140 and each communication interface.

  The language model 134 may also accept operations such as word registration, editing, and deletion by operations from the client system 110 of an authenticated special user.

  FIG. 2 generically shows a more detailed block diagram of the hardware configuration of the client system 110, the voice storage server 120, and the voice recognition server 130.

  The configuration in FIG. 2 includes a main memory 204 and a CPU 206, which are connected to the bus 202. The CPU is preferably based on 32-bit or 64-bit key architecture, such as Intel Pentium (R) 4, Xeon (R), Core 2 DUO, AMD Athlon (R), etc. can do. A display 210 such as an LCD monitor is connected to the bus 202 via a display controller 208. In the client system 110, the display 210 is used for a user to perform a search while looking at the web browser 112. In the voice storage server 120, a system administrator writes a program such as JavaScript (trademark) or PHP and registers it so as to be callable from the client system 110 or a user ID of a user who accesses through the client program 110. And used to register passwords.

  A hard disk 214 and a DVD drive 216 are also connected to the bus 202 via an IDE controller 212.

  In the case of the client system 110, the hard disk 214 stores an operating system, a web browser 112, and other programs that can be loaded into the main memory 204. Suitable operating systems include, but are not limited to, any operating system that supports TCP / IP networking functions, such as, but not limited to, Windows (R) XP, Windows (R) Vista, Linux (R), Mac OS, etc. Can be used.

  In the case of the voice storage server 120, the hard disk 214 stores an operating system, a voice data file 124, and an index data file 126. Suitable operating systems include, but are not limited to, any operating system that supports TCP / IP networking functions, such as Windows® 2003 Server, Linux®, Mac OS, etc. it can. The hard disk 214 further includes programs such as Apache and Tomcat for causing the voice recognition server 120 to function as a database server.

  In the case of the speech recognition server 130, the hard disk 214 stores an operating system, an acoustic model file 132, and a language model file 134. Suitable operating systems include, but are not limited to, any operating system that supports TCP / IP networking functions, such as Windows® 2003 Server, Linux®, Mac OS, etc. it can. The hard disk 214 further includes programs such as Apache and Tomcat for causing the voice recognition server 120 to function as an application server.

  The DVD drive 216 is used to additionally introduce a program from the CD-ROM or DVD disk to the hard disk 214 as necessary. Further, a keyboard 220 and a mouse 222 are connected to the bus 202 via a keyboard / mouse controller 220.

  The communication interface 224 preferably conforms to the Ethernet protocol, and plays a role of physically connecting the computer main body and the network 140 to the TCP / IP communication protocol of the communication function of the computer operating system. Providing a network interface layer. Here, the illustrated configuration is a wired connection configuration, but may be a wireless LAN connection based on a wireless LAN connection standard such as IEEE802.11a / b / g.

  Further, the communication interface 224 is not limited to the Ethernet protocol, and may conform to an arbitrary protocol such as a token ring, for example, and is not limited to a specific physical communication protocol.

  FIG. 3 is a functional block diagram for explaining the function of the voice recognition process executed by the voice recognition server 130. This processing program is stored in the hard disk drive of the speech recognition server 130 and is called up to the main memory as necessary. In FIG. 3, the input signal 302 is preferably provided via the network 140 as a separate audio file from the audio data file 124 of the audio storage server 122. In the voice recognition server 130, a voice data file may be temporarily saved in a hard disk and played by applying to a voice file player program, and the playback analog signal may be used as the input signal 302.

  In the acoustic processing block 304, the analog input signal 302 is once converted into a digital signal by A / D conversion. If the input signal 302 is a digital signal in advance, A / D conversion is not necessary.

  The acoustic processing block 304 further generates a frequency domain signal for discrete Fourier transform in a predetermined time window, generates a logarithmic spectrum therefrom, and further generates a cepstrum by discrete cosine transform. Further, the amplitude, fundamental frequency, power spectrum envelope, and the like of the waveform are extracted from the cepstrum by a known technique, and these become acoustic feature quantities.

  In the decoding block 306, the acoustic model 308 and the language model 310 are applied to the acoustic feature amount input from the acoustic processing block 304, thereby obtaining a text resulting from speech recognition of the input signal 302.

  More specifically, the acoustic model 308 is used to obtain a sequence of phonemes with high likelihood using a probabilistic model such as an HMM.

  On the other hand, the language model 310 is used to determine what word string is most likely using the vocabulary dictionary 312 from the phoneme sequence obtained by applying the acoustic model 308. For example, “post is red”, and the most audible word sequence can be “post is red”, “cost is high”, “host circle Nakai”, and so on. The language model 310 determines that “post is red” is the maximum likelihood.

  The speech recognition result text (word string) obtained in this way can be used in any other application program 314. In this embodiment, the text data resulting from the speech recognition is sent from the speech recognition server 130 to the speech storage server 120, and is stored in the index data file 126 in association with the speech file. The technique for associating the text obtained as a result of voice recognition of the voice file with the voice file as an index is not limited thereto, but is described in JP 2000-348064 A and JP 2006-178087 A related to the present applicant. ing.

  Next, processing according to an embodiment of the present invention will be described with reference to the flowchart in FIG. In step 402 of FIG. 4, the user inputs “reading” of the word / phrase using a predetermined graphic user interface (GUI). An example of this GUI is shown in FIG. It should also be noted that this GUI is executed on the client system 110 of FIG.

  Here, for example, if it is assumed that the user inputs the word “compound lending transaction” (compound word), the text field 602 shown as a reading field in FIG. input.

  In step 404, pronunciation / notation conversion using a language model is executed. That is, the phonetic recognition dictionary and its language model are used to convert pronunciation to notation. The processing program for this purpose is preferably in the voice recognition server 130, and the client system 110 simply calls the program on the voice recognition server 130 with a mechanism such as CGI or JSP. A speech recognition dictionary and its language model on the speech recognition server 130 are accessed and used via the network 140. The dictionary for speech recognition and its language model are generically shown as language model 134 in FIG. In this case, if necessary, the voice storage server 120 may create a voice recognition dictionary on the voice recognition server 130 and a replica of the language model, and the reason may be used.

As a method for obtaining an optimum output sequence based on a language model from a certain input symbol sequence, for example, the method described in [MORI 1999] described above can be used. There, the word string W is selected so as to maximize the conditional probability of the word string mixed with kana and kanji, which uses the input kana string Y as a condition. The formula is as follows.

  A detailed flowchart of step 404 in FIG. 4 is shown in FIG. In step 502, the input kana string is converted into a phoneme symbol string by the processing program so as to be associated with the speech recognition dictionary. Here, the phoneme symbol is a symbol corresponding to each of the types of sounds constituting the speech. In Japanese, there are about 50 types including vowels and consonants. Since conversion from readings (Sounds-like) to phonetic symbol strings is not necessarily one-to-one, a plurality of symbol strings are output in that case. However, the number is limited to a few at most, and the same processing is performed for each, finally the probability of each is compared, and it is obtained by selecting the maximum one. Describe.

For example, the following phonetic symbol string (H) can be obtained from “Kabuken Taisakutoriki”.
H = / k / a / b / u / k / e / _n / t / a / i / sy / a / k / u / t / o / r / i / h / i / k / i /

After that, the optimum word string W is selected by the above formula (1) with Y replaced by the phonetic symbol string H. For this reason, the formula (1) is modified as follows.

  Here, P (H | W) is a probability indicating how each word is read, and P (W) is an appearance probability of the word string, and is calculated by, for example, an N-gram model. These two values can be calculated from information that the speech recognition engine generally holds as a dictionary / language model. In other words, the maximum probability value W is selected based on the right side of the above expression, and this processing is the result of speech recognition when the sound model is ideal and the sound model is ideal (word string). Which are the following steps 504, 506 and 508. In the implementation, as shown in FIG. 3, the phoneme string obtained as a result of the acoustic processing in the speech recognition engine is input to the decoding block 306 as being uniquely determined.

  In step 504, the pointer serving as an index to the symbol of interest is set to the leftmost (/ k / in this example) of the phonetic symbol string.

この式で、右辺の第１項であるP(N|W_unk)は当該単語がN個の音素からなる読みをもつ確率であり、第２項は入力の部分列を構成する当該音素列が出現する確率を音素の1-gramにより近似している。なお、音素列は記号列の１つと考えられるが、記号列の出現確率を効率的に近似する手法は他にもさまざま存在する。例えば、第１項をポアソン分布、第２項をより高次のN-gramとするなどの手法を適用することができる。 In step 506, dictionary lookup is performed on the right subsequence starting from the pointer, and a word that matches the pronunciation is obtained as a candidate. “Section / ka”, “Stock / kabu”, “Stock certificate / kabuke_n”, etc. are candidates here. The above is a case where it is assumed that there is no unknown word, but it is normal that an unknown word exists in any dictionary. Therefore, in this step, a partial phoneme sequence that did not exist in the dictionary, for example “kabuke”, may be added to the candidate as an unknown word W _unk = <kabuke>, assuming that it was an unknown word. . In that case, the problem is how to calculate P (H | W) for a word whose notation is unknown. Many unknown word models have been proposed in so-called morphological analysis (eg, [NAGATA 1999] described above), but there is a merit in pronunciation / notation conversion (that part is pointed out as an unknown word and not converted). There is almost no discussion because there is no. Therefore, here, the “phoneme sequence” of the unknown word W _unk part is h (= h ₁ h ₂ ,..., H _N ), and the appearance probability of each phoneme (h _i ) is P (h _i ) as follows: A simple model.

In this equation, P (N | W _unk ), the first term on the right side, is the probability that the word has a reading consisting of N phonemes, and the second term is the phoneme sequence constituting the input subsequence. The probability of appearing is approximated by a 1-gram phoneme. The phoneme string is considered as one of the symbol strings, but there are various other methods for efficiently approximating the appearance probability of the symbol string. For example, a technique such as a Poisson distribution for the first term and a higher order N-gram for the second term can be applied.

In step 508, the language model is referred to, and the occurrence probability is calculated for the word (sequence) candidate obtained in step 506. For example, if it is based on N-gram, refer to the language model,
P (`` Section '') = P (Start symbol → `` Section '') = 0.0001
P (`` share '') = P (start symbol → `` share '') = 0.0005
P (`` Stock certificate '') = P (Start symbol → `` Stock certificate '') = 0.0025
Such a calculation is performed.

  In step 510, words (sequences) determined to have sufficiently small absolute values of probabilities or relative values compared to other candidates are excluded, and subsequent iterations are not performed. Otherwise, for each candidate, the pointer is updated at step 512 and the process from step 506 is repeated. In the above example, for example, “stock” and “stock certificate” which are the top two with the highest probability are left, and “section” is rejected. If “stock” is selected, the pointer is placed at “k” (5th phoneme from the left) immediately after “/ k / a / b / u /”, and the processing of step 506 is performed from that position. A candidate word that is repeated after 5 phonemes, for example, “fatigue / ke_ntai” becomes a candidate word. Various repetitions of candidate word strings are obtained by such repetition, but most of them are rejected in this process because their probabilities are sufficiently low. In the case of “stock certificate”, the pointer is advanced to “t” which is the next phoneme of “/ k / a / b / u / k / e / _n /” in the same manner, and step 506 and subsequent steps are repeated.

  As a result, one or a plurality of conversion results using the speech recognition dictionary and the language model are listed as a result of the processing of FIG. 5, and the user selects one of them by operating the mouse. As a result, as shown in FIG. 6, “stock certificate body transaction”, which is the selected result, is displayed in the text field 604 below the text field 602 in which “Kabuken Taikakutoriki” is displayed. Is done.

  In actual implementation, various ideas are made to speed up the above based on Viterbi and Dynamic Programming, but this is a well-known technique and will not be described in detail here.

  Returning to FIG. 4, if there is an erroneous part or it is determined that the word is an unknown word, in step 406, the user places the cursor on the part and preferably the normal kana-kanji conversion function. Use to correct. The conversion here is performed by clicking the “Conversion” button 606 in FIG. 6, but a conversion key on the keyboard may be hit. The kana-kanji conversion function here may be provided in the client computer 110. For example, in FIG. 6, since “Tasakaku” is “physical”, it is corrected to “loan”.

  In step 408, the first converted "stock certificate transaction" and the revised "stock certificate lending transaction" are compared by the system, so that "stock certificate / kabugen" and "transaction / toriki" parts Is correctly detected that the part corresponding to “borrowing / squeezing” is an unknown word for the speech recognition dictionary.

  When the user clicks the “confirm” button 608 in the next step 410, the unknown word “borrowing / squeezing” for the speech recognition dictionary is changed to “stock certificate” or “transaction”, which is a known word. Along with this, it is registered in the speech recognition dictionary in the context of “stock certificate → lending” or “lending → transaction”. Once registered in the speech recognition dictionary, the speech model is used by the language model 310 in the next speech recognition processing of the speech file.

  On the other hand, when the user clicks the “search” button 610 in step 410, “borrowing” is an unknown word in the speech recognition dictionary at this stage, and therefore there is a possibility that the search cannot be performed correctly. -Warn by creating a window. In any case, when the user clicks the “search” button 610, the keyword selected by the user from the conversion result based on the speech recognition dictionary is read by the client computer 110 from the conversion result based on the speech recognition dictionary. The voice storage server 120 searches the index data 126 based on the sent keyword.

  When the audio storage server 120 finds a hit in the keyword in the index data 126, it returns a list of audio files associated with the hit index data to the client computer 110.

  The client computer 110 displays the received audio file list in a separate window, and when the user of the client computer 110 clicks the link (not shown) of the audio file in the list as appropriate, Listen to the contents of audio files.

For reference, the following is a sample of pronunciation / notation conversion.
In the example below, the number on the left is -kΣlogP where P is the probability of the result (where k is a coefficient to make an integer, here it is 256, and the base of the logarithm is 10) ), And <..> u indicates an unknown word.
<Example 1>
Input >>
4565 <Stock Certificate Taiho Trading>
4640 <Stock certificate retirement transaction>
4673 <Retired stock certificates Nine transactions>
4732 <Stock Certificate 癪 Transaction>
4867 <Trading against stock certificates>
4937 <Thai Stock Trading>
<Example 2>
Input >> Tripple no Meigara <<
2247 <Triple A Brand>
3239 <Triple yeah brand>
3514 <Triple A No Brand>
3792 <Brand of Triple A>
3921 <Triple A Brand>
3942 <Triple D_A Brand>
4188 <triple yeah stocks>
<Example 3>
Input >>
4165 <Stock Certificate <Taishaku> u Transaction>
4565 <Stock Certificate Taiho Trading>
4640 <Shares <Transaction> u Trading>
4673 <Retired stock certificates Nine transactions>
4732 <Stock Certificate 癪 Transaction>
4758 <Transaction> u Transaction>

  As described above, the technique of the present invention has been described using a specific embodiment. However, the technical scope of the present invention is not limited to this specific embodiment, and various modifications are possible. For example, it may be a stand-alone configuration rather than a configuration connected by a network as shown in FIG. 1, and the voice storage server may be configured by the same server as the voice recognition server.

1 is an overall schematic diagram of a hardware configuration. It is a figure which shows the more detailed structure of the computer used by the structure of FIG. It is a block diagram which shows the function of a speech recognition system. It is a figure of the flowchart of the process which converts a reading using the language model for speech recognition. It is a figure of the flowchart of the process which converts a reading using the language model for speech recognition. It is a figure which shows the window for inputting and displaying a reading and a conversion result.

Claims (9)

A system for searching for data stored by associating an audio file with an index text obtained as a result of voice recognition of the audio file by computer processing,
Means for accepting reading input from a user by processing of the computer;
Means for converting the reading into a phonetic symbol string by the processing of the computer;
Means for converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file by the processing of the computer;
Means for searching the index text using the keyword by the processing of the computer;
Audio file search system. Means for correcting the keyword in accordance with a user operation by the processing of the computer;
Means for identifying an unknown word based on the corrected part of the keyword by the processing of the computer;
Means for notifying the user of the presence of the unknown word;
The audio file search system according to claim 1. A system for updating, by computer processing, a speech recognition dictionary for recognizing speech files and creating index texts for retrieving speech files,
Means for accepting reading input from a user by processing of the computer;
Means for converting the reading into a phonetic symbol string by the processing of the computer;
Means for converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file by the processing of the computer;
Means for correcting the keyword in accordance with a user operation by the processing of the computer;
Means for identifying an unknown word based on the corrected part of the keyword by the processing of the computer;
Means for registering a user correction word corresponding to the unknown word in the speech recognition dictionary by the processing of the computer;
A speech recognition dictionary update system. A method for searching for data stored by associating an audio file with an index text obtained as a result of voice recognition of the audio file by computer processing,
Accepting reading input from a user by processing of the computer;
Converting the reading into a phonetic symbol string by processing of the computer;
Converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file by the processing of the computer;
A step of searching the index text using the keyword by the processing of the computer;
How to search for audio files. Correcting the keyword in accordance with a user operation by a process of the computer;
Identifying an unknown word based on the corrected portion of the keyword by processing of the computer;
Further notifying the user of the presence of the unknown word,
The method for searching for an audio file according to claim 4. A method for updating, by computer processing, a speech recognition dictionary for recognizing speech files and creating index texts for searching speech files,
Accepting reading input from a user by processing of the computer;
Converting the reading into a phonetic symbol string by processing of the computer;
Converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file by the processing of the computer;
Correcting the keyword in accordance with a user operation by a process of the computer;
Identifying an unknown word based on the corrected portion of the keyword by processing of the computer;
Registering a user correction word corresponding to the unknown word in the speech recognition dictionary by the processing of the computer;
How to update the speech recognition dictionary. A program for searching for data stored by associating an audio file with an index text obtained as a result of voice recognition of the audio file by computer processing,
The computer,
Accepting reading input from the user;
Converting the reading into a phonetic symbol string;
Converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file;
Performing a step of searching the index text using the keyword;
An audio file search program. The computer,
Correcting the keyword in accordance with a user operation by a process of the computer;
Identifying an unknown word based on the corrected portion of the keyword by processing of the computer;
Further informing the user of the presence of the unknown word,
The audio file search program according to claim 7. A program for updating a speech recognition dictionary for creating an index text by recognizing the speech file in order to search for the speech file by computer processing,
The computer,
Accepting reading input from a user by processing of the computer;
Converting the reading into a phonetic symbol string by processing of the computer;
Converting the phonetic symbol string into a keyword using a language model substantially the same as the language model used for speech recognition of the speech file by the processing of the computer;
Correcting the keyword in accordance with a user operation by a process of the computer;
Identifying an unknown word based on the corrected portion of the keyword by processing of the computer;
Registering a user correction word corresponding to the unknown word in the speech recognition dictionary by the processing of the computer;
Voice recognition dictionary update program.

Images