JP2005257954A - Speech retrieval apparatus, speech retrieval method, and speech retrieval program - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech retrieval apparatus, a speech retrieval program, and a speech retrieval method that provide an efficient retrieval result by retrieving an optional word from speech data including a recognition error with a small information processing quantity. <P>SOLUTION: Speech retrieval is carried out by using the speech retrieval apparatus equipped with: a word/phrase expansion part which converts an inputted word to generate a phoneme string or syllable string; a phoneme string conversion part which adds or substitutes a new phoneme to or for the phoneme string or syllable string to generate a new phoneme string or new syllable string; a phoneme/syllable data storage part which stores phoneme/syllable data to be retrieved; and a collation part which collates the new phoneme string or new syllable string with the phoneme/syllable data to be retrieved. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  The present invention relates to a voice search device, a voice search method, and a voice search program, and more particularly to a technique for searching for a specific word from spoken voice.

  With the increase in the capacity of storage devices accompanying the advancement of computer technology and the spread of data communication using broadband due to the development of network technology, digitization of voice has begun. There is a demand for a technique that can appropriately search voice data that is an electronic voice. There may be a case where a recognition error is mixed in a phoneme string transcribed by speech recognition according to the prior art. For this reason, even when sounds that are determined to be the same reading when a human hears them by ear, a device (or software) that performs speech recognition may not generate the same phoneme string. A technique is known that enables a search using an input search keyword even when a recognition error exists (see, for example, Patent Document 1 and Non-Patent Document 1).

  FIG. 1 is a block diagram showing the configuration of the speech search apparatus described in Patent Document 1. Referring to FIG. 1, the speech search apparatus includes a speech / electrical signal conversion unit 101, a speech data storage unit 102, a phoneme or syllable recognition unit 103, a speech data start end storage unit 104, and a phoneme or syllable sequence storage unit. 105, search word / phrase phoneme string or syllable storage unit 106, matching unit 107, likelihood threshold storage unit 108, comparison unit 109, audio data reproduction start pointer 110, and electric signal / audio conversion unit 111 It consists of and.

  The conventional search device having such a configuration operates as follows. The sentence voice is converted into an electric signal by the voice / electrical signal conversion unit 101, the voice data is stored in the voice data storage unit 102, and the voice data is input to the phoneme or syllable recognition unit 103. The phoneme or syllable recognition unit 103 stores the phoneme or syllable sequence of sentence speech as the recognition result in the phoneme or syllable sequence storage unit 105, and sets the start position of the recognized speech data of each phoneme or syllable as the phoneme or syllable. Is stored in the voice data start-end storage unit 104.

  Next, the phoneme or syllable string of the word or phrase desired to be searched, which is input by characters using a keyboard or the like, is stored in the phoneme or syllable string storage unit 106 of the search word / phrase. The matching unit 107 matches the phoneme or syllable sequence of the sentence phoneme or syllable series storage unit 105 with the phoneme or syllable string of the word or phrase that the search word / phrase phoneme or syllable string storage unit 106 desires to search. To calculate the likelihood of matching with the phoneme or syllable string of the word or phrase desired to be searched in the phoneme or syllable string of the sentence speech.

  The matching likelihood calculation result and the likelihood threshold set in advance in the likelihood threshold storage unit 108 are input to the comparison unit 109 to detect the phoneme of the sentence speech exceeding the likelihood threshold or the section in the syllable series. , Output the position of the section. The voice data playback start pointer 110 indicates the position of the sentence voice data in the sentence voice data storage unit 102 in the search section by inputting the phoneme or syllable in the search result section and the start position of the voice data in the voice data start storage unit 104. Then, an electric signal which is audio data from this position is input to the electric signal / audio converter 111. As a result, the search result can be output as sound from the electrical signal / electrical converter 111. Since the above technique uses the likelihood to derive the search result, even when a recognition error exists, it is possible to execute a search with relatively high accuracy.

  The technique described in Non-Patent Document 1 searches for an audio file of an article including the voice of a query word from a news article reading voice corpus. The query is given as a phoneme sequence, and a phoneme sequence including a recognition error obtained by phoneme recognition of a speech corpus is searched. At this time, a certain amount of insertion, omission, and replacement is absorbed by using continuous DP matching that calculates the distance between the phoneme string starting from an arbitrary start point in the continuous phoneme string and the query. In addition, a confusion matrix that summarizes what phonemes are likely to be mistaken for is also used in conjunction with DP matching to enable more effective searches.

Japanese Examined Patent Publication No. 7-69708 Yuki Maeda, Akira Shimazu, Full-text search based on phoneme recognition, SIG-SLUD-A102, pp.1-6, 2001

  The problem to be solved by the present invention is that when an arbitrary word (or phrase) is searched from speech data in which recognition errors are mixed, a search is executed with a small amount of information processing, and an effective search result is obtained. An object is to provide a voice search device, a voice search program, and a voice search method.

  The means for solving the problem will be described below using the numbers used in [Best Mode for Carrying Out the Invention]. These numbers are added to clarify the correspondence between the description of [Claims] and [Best Mode for Carrying Out the Invention]. However, these numbers should not be used to interpret the technical scope of the invention described in [Claims].

  A phrase expansion unit (6) that converts an input word to generate a phoneme string or a syllable string, and adds or subtracts a new phoneme to the phoneme string or the syllable string, or the phoneme string or the syllable string A phoneme string conversion unit (7) for generating a new phoneme string or a new syllable string by replacing at least one phoneme constituting the phoneme, and a phoneme / syllable data for storing search target phoneme / syllable data The storage unit (9), the phoneme string or the syllable string, and the search target phoneme / syllable data are collated, and the new phoneme string or the new syllable string, the search target phoneme / syllable data, A voice search is performed on data including a recognition error using a voice search device including a matching unit (4) for matching the data.

  In the speech search apparatus, the search target phoneme / syllable data is composed of a plurality of phonemes, and the collation unit (4) finds a position that matches the new phoneme string or the new syllable string by the collation. A speech search is performed on data including recognition errors using a speech search device that detects from the search target phoneme / syllable data. Further, the phoneme string conversion unit (7) stores a development rule (25) that is a rule for generating the new phoneme string or the new syllable string, and based on the expansion rule (25), A voice search is performed on data including recognition errors using a new phoneme string or a voice search device that generates the new syllable string.

  Further, in the voice search device, the search target phoneme / syllable data is generated based on the voice data generated by sequentially executing information processing for converting the voice input by the voice input device into electronic data. Is done. The voice data may be voice data (30) stored in advance. In addition, the expansion rule (25) includes a plurality of phonemes constituting the search target phoneme / syllable data and phonemes constituting correct answer data (28) as a result of correctly recognizing the phoneme / syllable of the speech data (30). Perform a voice search for data containing recognition errors using a voice search device set based on comparison with syllables. Here, the speech search apparatus according to the present invention has no limitation on the data format that is the basis for generating the search target phoneme / syllable data.

  In the speech search device, the search target phoneme / syllable data is generated based on speech data generated from speech input by the speech input device as described above, or speech data (30) stored in advance. The expansion rule (25) executes a voice search for data including a recognition error using a voice search device set based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data. .

  In the speech search apparatus, the phrase expansion unit (6) includes morpheme analysis means (50) for analyzing morphemes constituting the input word, and the phoneme string conversion unit (7) includes the morpheme analysis means ( 50) speech for data including a recognition error using the speech search device that generates the new phoneme sequence or the new syllable sequence based on the analysis result output from 50) and the phoneme sequence or the syllable sequence. Perform a search. Moreover, the phrase expansion unit (6) further includes a registered phrase determination unit (51), and each of the morphemes that is a result of analysis by the morpheme analysis unit (50) The phoneme string conversion unit (7) determines whether the phoneme string is registered in advance, and uses the speech search device that generates the new phoneme string or the new syllable string based on the determination result. Perform a voice search for data containing recognition errors.

  In the speech search device, the search target phoneme / syllable data is based on speech data generated by sequentially executing information processing for converting speech input by the speech input device into electronic data, as described above. Generated. The voice data may be voice data (30) stored in advance. The search target phoneme / syllable data is generated based on a speech recognition result obtained by recognizing the speech data or a phoneme recognition result obtained by phoneme recognition of the speech data. Here, the voice data is generated from a set of spoken voices. A voice search is performed on data including a recognition error using the voice search apparatus having such a configuration. Further, in the speech search device, the search target phoneme / syllable data is generated based on a language model stored in advance, and the language model is a speech search device that is information describing connection restrictions of words. Use to perform a voice search for data containing recognition errors. Further, in the speech search device, the phrase expansion unit (6), when the input word is a word that does not need to generate the new phoneme string or the new syllable string, the phoneme string Alternatively, a voice search is performed on data including a recognition error using a voice search device that outputs the syllable string to the collation unit (4).

  In the speech search device, the speech search device has a predetermined storage area, and has phoneme dictionary data or syllable dictionary data for converting the input word in the storage area, and the phrase expansion unit (6) When generating a phoneme string, the phoneme dictionary data is used, and when generating a syllable string, the syllable dictionary data is used. Thereby, the phrase expansion unit (6) converts the inputted word into a phoneme or a syllable, and performs a voice search by a voice search device that generates the phoneme string or the syllable string based on the converted phoneme or syllable. Execute.

  When solving the above problem by a computer program, a step of converting an input word to generate a phoneme string or a syllable string, adding or subtracting a new phoneme to the phoneme string or the syllable string, or the phoneme Replacing the phoneme constituting the sequence or the syllable sequence with another phoneme, generating a new phoneme sequence or a new syllable sequence, reading the stored search target phoneme / syllable data, and the new A computer-executable program is installed in a predetermined computer and a method comprising the step of collating a phoneme string or the new syllable string with the search target phoneme / syllable data, and a voice is obtained by executing the program. Perform a search.

  In the program, the step of reading out the search target phoneme / syllable data constituted by a plurality of phonemes, and the search results in the search target phoneme / syllable data corresponding to the new phoneme string or the new syllable string by the collation. A program that can be executed by a computer is installed in a predetermined computer, and voice search is executed by executing the program.

  In the program, the step of reading the expansion rule (25), and the expansion rule (25) are rules for generating the new phoneme string or the new syllable string, and are based on the expansion rule (25). Then, a computer-executable program that includes the step of generating the new phoneme sequence or the new syllable sequence is installed in a predetermined computer, and voice search is executed by executing the program.

  In the program, the search target is based on voice data generated by sequentially executing information processing for converting voice inputted by the voice input device into electronic data, or voice data (30) stored in advance. Based on comparison between a step of generating phoneme / syllable data, a plurality of phonemes constituting the search target phoneme / syllable data, and a phoneme constituting correct answer data (28) as a result of correct phoneme recognition of the speech data. Then, a program that can be executed by a computer is installed in a predetermined computer and a voice search is executed by executing the program. Here, the speech search apparatus according to the present invention has no limitation on the data format that is the basis for generating the search target phoneme / syllable data.

  In the program, similar to the above, generating the search target phoneme / syllable data based on voice data generated from voice input by the voice input device or voice data stored in advance (30); A computer-executable program having a method of setting the expansion rule (25) based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data is installed in a predetermined computer. Perform a voice search by running the program.

  In the program, a step of analyzing a morpheme constituting an input word, and a step of generating the new phoneme sequence or the new syllable sequence based on the analysis result and the phoneme sequence or the syllable sequence A program that can be executed by a computer is installed in a predetermined computer, and voice search is executed by executing the program.

  In the program, a method comprising: determining whether each of the morphemes is registered in advance; and generating the new phoneme string or the new syllable string based on the determination result Is installed in a predetermined computer and a voice search is executed by executing the program.

  In the program, the retrieval target phoneme / syllable is read based on a step of reading out speech data (30) and a speech recognition result obtained by speech recognition of the speech data, or a phoneme / syllable recognition result obtained by phoneme / syllable recognition of the speech data. A program that can be executed by a computer is mounted on a predetermined computer, and a voice search is executed by executing the program. The voice data is a set of spoken voices, and is generated from voice data generated by sequentially converting voice input by a voice input device into electronic data in the same manner as described above. The voice data may be stored in advance.

  In the program, a step of reading a language model stored in advance, and the language model is information describing connection restrictions of words, and includes generating the search target phoneme / syllable data based on the language model. A program that can be executed by a computer is installed in a predetermined computer, and voice search is executed by executing the program. In the program, the step of determining whether or not the input word needs to generate the new phoneme string or the new syllable string; and as a result of the determination, the input word Is a word that does not require generation of the new phoneme string or the new syllable string, a program for outputting the phoneme string or the syllable string without making it a new phoneme string or a new syllable string. The voice search is executed by installing the program on a predetermined computer and executing the program.

  In the program, the step of reading phoneme dictionary data or syllable dictionary data for converting the inputted word stored in a predetermined storage area; and the input based on the phoneme dictionary data or the syllable dictionary data A computer-executable program comprising a step of converting a converted word into a phoneme or a syllable and a step of generating the phoneme sequence or the syllable sequence based on the converted phoneme or syllable Install and execute voice search by executing the program.

  Furthermore, as a method for solving the above-mentioned problem, a step of converting an input word to generate a phoneme string or a syllable string, adding or subtracting a new phoneme to the phoneme string or the syllable string, Replacing a phoneme constituting the syllable string with another phoneme to generate a new phoneme string or a new syllable string; reading a stored search target phoneme / syllable data; and the new phoneme string Alternatively, a speech search is performed on data including a recognition error using a speech search method including a step of collating the new syllable string with the search target phoneme / syllable data.

  In the speech search method, the step of reading out the search target phoneme / syllable data composed of a plurality of phonemes, and the matching to the new phoneme string or the location matching the new syllable string by the collation Perform a speech search on data containing recognition errors using a speech search method comprising detecting from syllable data.

  In the speech search method, the step of reading the expansion rule (25), and the expansion rule (25) are rules for generating the new phoneme string or the new syllable string, and the expansion rule (25) A speech search is performed on data including recognition errors using a speech search method comprising the step of generating the new phoneme sequence or the new syllable sequence based on

  In the voice search method, the search target phoneme / syllable data is generated based on voice data obtained by converting voice input by a voice input device into electronic data in real time or voice data (30) stored in advance. And a plurality of phonemes constituting the search target phoneme / syllable data and the expansion rule based on a comparison between the phoneme / syllable constituting the correct answer data (28) as a result of correctly recognizing the phoneme / syllable Perform a voice search on data containing recognition errors using a voice search method comprising the step of setting (25). In the speech search method according to the present invention, there is no limitation on the data format that is the basis for generating the search target phoneme / syllable data.

  In the voice search method, the search target phoneme / syllable data is generated based on voice data obtained by converting voice input by a voice input device into electronic data in real time or voice data (30) stored in advance. And using a speech search method comprising a step of setting the expansion rule (25) based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data. Perform a voice search.

  In the speech search method, the new phoneme string or the new syllable string is generated based on the step of analyzing the morpheme constituting the input word, the analysis result, and the phoneme string or the syllable string Performing a voice search on data including a recognition error using a voice search method comprising the steps of:

  In the speech search method, the method includes a step of determining whether each of the morphemes is registered in advance, and a step of generating the new phoneme sequence or the new syllable sequence based on the determination result. The voice search is performed on the data including the recognition error using the voice search method.

  In the voice search method, a step of reading voice data obtained by converting voice input by a voice input device into electronic data in real time, or voice data (30) stored in advance, and a voice recognition result obtained by voice recognition of the voice data Alternatively, a speech search is performed on data including a recognition error using a speech search method including a step of generating the search target phoneme / syllable data based on a phoneme recognition result obtained by phoneme recognition of the speech data. The voice data is a set of spoken voices, and the voice input by the voice input device is generated from the voice data that is sequentially converted into electronic data in the same manner as described above. The voice data may be stored in advance.

  In the speech search method, a step of reading a language model stored in advance, and the language model is information describing a word connection constraint, and generating the search target phoneme / syllable data based on the language model A voice search is performed on data including recognition errors using a voice search method comprising:

  In the speech search method, the step of reading phoneme dictionary data or syllable dictionary data for converting the input word stored in a predetermined storage area, and based on the phoneme dictionary data or the syllable dictionary data, Data including a recognition error using a speech search method comprising: converting an input word into a phoneme or syllable; and generating the phoneme sequence or the syllable sequence based on the converted phoneme or syllable Perform a voice search for.

  According to the present invention, when using a search keyword that is an arbitrary word (or phrase) for audio data that is a search target and specifying a location corresponding to the search keyword, the amount of information processing is small. There is an effect that a search result can be obtained.

  Furthermore, according to the present invention, when a phoneme string is generated from an arbitrary input word (or phrase), a search is performed in consideration of the possibility that a recognition error has occurred. There is an effect that a result is obtained.

  The best mode for carrying out the present invention will be described below with reference to the drawings.

[Configuration of First Embodiment]
FIG. 2 is a block diagram showing the configuration of the first embodiment for carrying out the present invention. According to this block diagram, the speech search apparatus in the first embodiment includes a search keyword input unit 1, a phoneme / syllable processing unit 2, a phoneme / syllable data output unit 3, a matching unit 4, and an output unit 5. It is shown that it consists of In the embodiment described below, a case where phoneme / syllable data to be searched is generated from previously stored speech data will be described as an example, but this restricts the state of speech data in the present invention. is not. For example, it is possible to perform a voice search in real time by sequentially executing recognition processing on input voice (spoken voice or the like) to create voice data.

  The search keyword input unit 1 is an information input device that outputs words, phrases, and sentences (hereinafter referred to as keywords) input by characters or voice as data that can be processed by a computer.

  The phoneme / syllable processing unit 2 is a data processing function block that converts a search keyword into a phoneme string or a syllable string (hereinafter referred to as a phoneme string). The phoneme / syllable processing unit 2 further includes a phrase expansion unit 6 and a phoneme string conversion unit 7. The phrase expansion unit 6 is a data processing function block that converts a keyword output from the search keyword input unit 1 into a phoneme string or the like. The phrase expansion unit 6 uses a phoneme / syllable dictionary (not shown) provided in the speech search device to determine which phoneme (or syllable) the keyword is composed of. The phrase expansion unit 6 converts the keyword into a phoneme string or the like by generating a phoneme string or the like using the phoneme (or syllable) obtained by the determination. The phoneme string conversion unit 7 is a data processing function block that generates an input phoneme string and a new phoneme string corresponding to a predetermined processing rule. Details regarding the phoneme string converter 7 will be described later.

  The phoneme / syllable data output unit 3 is a data output function block that outputs search target phoneme / syllable data that can be searched for speech. The phoneme / syllable data output unit 3 further includes a phoneme / syllable data generation unit 8 and a phoneme / syllable data storage unit 9. The phoneme / syllable data generation unit 8 is a data generation function block that generates speech data that can be processed by a computer from input speech and generates search target phoneme / syllable data from the speech data. The phoneme / syllable data storage unit 9 is an information storage function block that stores the search target phoneme / syllable data output from the phoneme / syllable data generation unit 8.

  The matching unit 4 is a collating function block that performs matching between the phoneme string output from the phoneme / syllable processing unit 2 and the search target phoneme / syllable data output from the phoneme / syllable data output unit 3. The matching unit 4 performs matching on the phoneme string output from the phoneme / syllable processing unit 2 and the search target phoneme / syllable data output from the phoneme / syllable data output unit 3, When the syllable data includes a portion that matches a phoneme string, information for identifying the matching portion (for example, a single sentence including the matching portion or a compound sentence including the single sentence) is extracted. When matching is performed, the phonemes or syllables included in the search target phoneme / syllable data output from the phoneme / syllable data output unit 3 and the phoneme string output from the phoneme / syllable processing unit 2 are selected. It is also possible to create an index as a set and perform a search more efficiently by matching the indexes. The output unit 5 is an information output device that receives the information output from the matching unit 4, converts the information into a data format that can be output, and outputs the data.

  FIG. 3 is a block diagram showing details of the configuration of the phoneme / syllable data generation unit 8 and the phoneme / syllable data storage unit 9 provided in the phoneme / syllable data output unit 3. Referring to FIG. 3, the phoneme / syllable data generation unit 8 is connected to the speech data storage unit 30, the phoneme / syllable recognition unit 31 connected to the speech data storage unit 30, and the phoneme / syllable recognition unit 31. And an acoustic model storage unit 32. The voice data storage unit 30 is an information storage function block that stores data generated by converting voice input in advance into a format that can be processed by a computer. The phoneme / syllable recognition unit 31 is an information processing function block that generates search target phoneme / syllable data corresponding to the speech data output from the speech data storage unit 30. The acoustic model storage unit 32 is an information storage function block that stores an acoustic model that is information of acoustic features of recognition units.

  The phoneme / syllable recognition unit 31 performs phoneme / syllable recognition based on the audio data output from the audio data storage unit 30 and the acoustic model output from the acoustic model storage unit 32. The phoneme / syllable recognition unit 31 outputs the phoneme / syllable data as the execution result to the phoneme / syllable data storage unit 9. The phoneme / syllable data storage unit 9 stores the phoneme / speech data output from the phoneme / syllable data generation unit 8 as search target phoneme / syllable data. The phoneme / syllable data output unit 3 outputs the search target phoneme / syllable data stored in the phoneme / syllable data storage unit 9 in response to a request from the matching unit 4.

  FIG. 4 is a block diagram showing in detail the configuration of the phoneme / syllable data generation unit 8 and the phoneme / syllable data storage unit 9 provided in the phoneme / syllable data output unit 3. Referring to FIG. 4, the phoneme / syllable data generation unit 8 includes a speech data storage unit 30, a speech recognition unit 33, a phoneme / syllable conversion unit 34, an acoustic model storage unit 35, and a language model storage unit 36. Is done. Since the voice data storage unit 30 and the phoneme / syllable data storage unit 9 shown in FIG. 4 are the same as those shown in FIG. 3, detailed description thereof will be omitted. The voice recognition unit 33 is an information processing function block that executes voice recognition of input voice data. The speech recognition unit 33 includes an acoustic model stored in the acoustic model storage unit 35 (information indicating acoustic characteristics of recognition units) and a language model stored in the language model storage unit 36 (connection restrictions between words to be recognized). Large vocabulary continuous speech recognition is performed using the information indicating The phoneme / syllable conversion unit 34 is an information processing function block that generates search target phoneme / syllable data based on the recognition result obtained by the large vocabulary continuous speech recognition processing executed by the speech recognition unit 33.

  The phoneme / syllable data generation unit 8 uses the acoustic model stored in the acoustic model storage unit 32 (or the acoustic model storage unit 35) described above to search for the phoneme / syllable data storage unit 9 to be searched. Generate data. As the acoustic model stored in the acoustic model storage unit 32 (or the acoustic model storage unit 35), any acoustic model can be selected. For example, it is possible to generate phoneme / syllable data using HMM (Hidden Markov Model), which is usually widely used for speech recognition, and to use other acoustic models as necessary. A precision search is possible. As the language model stored in the language model storage unit 36, any language model can be selected similarly to the acoustic model. For example, the phoneme and syllable data is generated using the word N-Gram model that is normally widely used for speech recognition, and other language models are used as necessary, so that higher accuracy can be achieved. Can be searched.

  By configuring the phoneme / syllable data generation unit 8 to include the language model storage unit 36 shown in FIG. 4, large vocabulary continuous speech recognition can be performed. The search target phoneme / syllable data obtained by the large vocabulary continuous speech recognition processing is a known word (a word registered in the language model) compared to the phoneme / syllable data generation unit 8 having no language model. High-precision processing can be performed in the portion related to. By providing such phoneme / syllable data generation unit 8 and switching between them according to the importance of processing speed and accuracy of phoneme / syllable recognition, a highly convenient search device can be configured.

  FIG. 5 is a block diagram showing in detail the configuration of the phoneme string conversion unit 7. Referring to FIG. 5, the phoneme string conversion unit 7 includes an expansion rule output unit 20 and an expansion execution unit 21, which are connected to each other. The expansion rule output unit 20 is an information output function block that outputs a expansion rule that is a processing rule for information processing executed by the expansion execution unit 21. The expansion execution unit 21 is an information processing function block that transforms the phoneme sequence output from the phrase expansion unit 6 based on the expansion rule output from the expansion rule output unit 20.

  The expansion rule output unit 20 further includes an expansion rule creation phoneme / syllable data storage unit 22, a statistical processing unit 23, an expansion rule generation unit 24, and an expansion rule storage unit 25. The expansion rule creation phoneme / syllable data storage unit 22 is an information storage function block for storing expansion rule creation phoneme / syllable data. The phoneme / syllable data stored in the expansion rule creating phoneme / syllable data storage unit 22 can be arbitrarily changed. For example, by extracting the stored phoneme / syllable data from the phoneme / syllable data storage unit 9 and generating a development rule, it is possible to create a development rule corresponding to specific speech data.

  The statistical processing unit 23 is a statistical processing functional block that executes statistical processing of phoneme / syllable data stored in the phoneme / syllable data storage unit 22 for developing rules. The development rule creation unit 24 is an information creation function block that creates a development rule based on the statistical processing result output from the statistical processing unit 23.

  Hereinafter, the expansion rule creation process executed by the expansion rule creation unit 24 will be described in detail by taking as an example the case where the statistical processing unit 23 performs n-gram statistics. The statistical processing unit 23 investigates the phoneme n-gram statistics of the phoneme string included in the phoneme / syllable data for rule creation in advance, and arranges the expansion rules in the order of the frequency of the phoneme n-gram. The statistical processing unit 23 outputs the statistical processing result to the development rule creation unit 24, and the development rule creation unit 24 sets a predetermined threshold for the output statistical processing result. Further, the expansion rule creation unit 24 creates a plurality of expansion rules while changing the front and rear phoneme conditions using the phoneme / syllable data stored in the expansion rule creation phoneme / syllable data storage unit 22. The development rule creation unit 24 narrows down the plurality of statistical rules based on the created plurality of development rules and the statistical processing result provided with the threshold value. For example, a threshold value that is considered to be rarely applied to keywords with a low frequency of phoneme n-grams is provided, and a development rule that is equal to or lower than the threshold value is deleted. Conversely, the expansion rule that deletes the new phoneme string generated by the expansion execution unit 21 includes a context with a high frequency of n-grams. By executing such processing, a development rule set with high efficiency and accuracy is created.

  FIG. 6 is a block diagram showing the configuration of the phoneme string conversion unit 7 in detail. Referring to FIG. 7, the phoneme string conversion unit 7 includes an expansion rule output unit 20 and an expansion execution unit 21, which are connected to each other. The expansion rule output unit 20 further includes an expansion rule creation voice data storage unit 26, a development rule generation phoneme / syllable data storage unit 27, a correct rule storage unit 28, a statistical processing unit 23, an expansion rule generation unit 24, and an expansion rule. And a storage unit 25. The expansion execution unit 21, the statistical processing unit 23, the expansion rule creation unit 24, and the expansion rule storage unit 25 are the same as those shown in FIG.

  The expansion rule creation voice data storage unit 26 is an information storage function block that stores voice data for development rule creation. The audio data stored in the expansion rule creating audio data storage unit 26 can be arbitrarily changed. For example, by creating a development rule by extracting the voice data stored in the voice data storage unit for development rule creation 26 from the voice data storage unit 30, a development rule corresponding to specific voice data is created. It becomes possible to do. The expansion rule creation phoneme / syllable data storage unit 27 is an information storage function block that stores the expansion rule generation phoneme / syllable data created based on the speech data stored in the expansion rule creation speech data storage unit 26. is there. The stored expansion rule creating phoneme / syllable data is a phoneme / syllable recognition result obtained by executing phoneme / syllable recognition on the speech data stored in the expanded rule creating speech data storage unit 26, or This is data obtained by converting a speech recognition result obtained by executing speech recognition into a phoneme string or a syllable string. This expansion rule creation phoneme / syllable data may include a recognition error.

  The correct rule storage unit 28 is an information storage function block that stores correct answer data. The correct answer data is the correct phoneme string (or syllable string) of the expansion rule creation phoneme / syllable data stored in the expansion rule creation phoneme / syllable data storage unit 27. This correct answer data may be data created for the entire development rule creating voice data, or may be a part thereof.

  The statistical processing unit 23 is an information processing function block that performs matching with the phoneme string (or syllable string) of the rule creating phoneme / syllable data and the phoneme string (or syllable string) of the correct answer data. As a result of the matching, the statistical processing unit 23 obtains statistics about the phoneme string (or syllable string) in the correct answer data and the phoneme string (or syllable string) in the development rule creation phoneme / syllable data, and the statistics The result is output to the expansion rule creation unit 24.

  The expansion rule creation unit 24 is an information generation function block that generates expansion rules based on the statistical results. The expansion rule creation unit 24, for example, performs DP matching between the phoneme sequence of the rule creation syllable / speech data and the phoneme sequence of the correct data, and the phoneme in the correct data is a rule for each phoneme that looks at the preceding and following phoneme conditions. The frequency of which phoneme in the phoneme / syllable data for creation is replaced (or dropped or inserted) is investigated.

An example of creating an expansion rule based on replacement will be described below. It should be noted that the expansion rule can be created by the same process for dropping and inserting. The following list example is a part of the phoneme that looks at the phoneme conditions before and after in the correct answer data, the phoneme that the phoneme became in the development rule creation phoneme / syllable data, and the number of replacements. is there.
(A) t (o) k 50
(A) t (o) d 40
(A) t (o) p 20
(I) k (a) g 400
(I) k (a) t 100
(E) s (u) z 3
(E) s (u) c 2
In the above list example, for example, “(a) t (o) k 50” is the phoneme t in the correct answer, the phoneme t in which the front phoneme is a and the back phoneme is o, in the phoneme / syllable data for rule creation. It shows that the frequency of becoming k was 50.

Further, the creation of the expansion rule set is performed based on specific conditions. For example, let us consider a case where a condition of “arrange from the most frequently appearing items and adopt a certain threshold value or more” is set as a creation condition. In this case, for example, if the threshold is 50 or more,
(I) k (a) → g, (i) k (a) → t, (a) t (o) → k
Adopt the deployment rules.
Further, a case is considered in which “the frequency ratio is investigated for each phoneme as a replacement destination of phonemes in correct answer data, and a phoneme having a certain threshold value or more is adopted”. At this time, if the phoneme “t” in the correct answer data is extracted and investigated, and (a) t (o) appears 200 times in total,
(A) t (o) → k 50/200 = 0.25,
(A) t (o) → d 40/200 = 0.20,
(A) t (o) → p 20/200 = 0.10,
It becomes.
Here, if the threshold is 0.20 or more, the expansion rules adopted are (a) t (o) → k, (a) t (o) → d in the case of phoneme (a) t (o). Become.
[Operation of First Embodiment]
FIG. 7 is a flowchart showing the operation of the first exemplary embodiment of the present invention. Referring to FIG. 7, the operation of the first embodiment starts when content to be searched is determined. In step S101, a search keyword is input to execute a search for the phoneme string or syllable string in the determined content. The input search keyword is output from the search keyword input unit 1 and input to the phoneme / syllable processing unit 2.

  In step S <b> 102, the phoneme / syllable processing unit 2 inputs the search keyword to the phrase expansion unit 6. The phrase expansion unit 6 to which the search keyword is input converts the search keyword into a phoneme string or a syllable string (hereinafter referred to as a phoneme string or the like as in [Configuration of the embodiment]), and the process proceeds to step S103. In step S <b> 103, the phrase expansion unit 6 determines whether the converted phoneme string or the like is to be output to the phoneme string conversion unit 7. As a result of the determination, if it is determined that the phoneme string or the like does not need to be processed by the phoneme string conversion unit 7, the process proceeds to step S106. For example, if the input search keyword is a word with relatively few recognition errors in speech recognition, the phrase expansion unit 6 outputs the phoneme sequence of the word to the matching unit 4 without outputting it to the phoneme sequence conversion unit 7 To do. This makes it possible to reduce the amount of data processing corresponding to the input search keyword.

  In step S104, the phoneme string conversion unit 7 to which the phoneme string output from the phrase expansion unit 6 is input outputs the phoneme string to the expansion execution unit 21. The expansion execution unit 21 receives the phoneme string output from the phoneme string conversion unit 7 and extracts the expansion rule stored in the expansion rule storage unit 25 in response to the phoneme string. In step S105, the expansion executing unit 21 that has extracted the expansion rule generates a new phoneme string or the like corresponding to the phoneme string of the search keyword sent from the phrase expansion unit 6 based on the expansion rule.

  In step S <b> 106, the phoneme string conversion unit 7 outputs the new phoneme string generated by the expansion execution unit 21 to the matching unit 4. The matching unit 4 responds to an input of at least one of a search keyword phoneme string or the like or a new phoneme string (syllable string) and outputs a search target phoneme / syllable data of a search target content to a phoneme / syllable data output unit. 3 is requested (step S107). The phoneme / syllable data output unit 3 extracts the corresponding search target phoneme / syllable data from the phoneme / syllable data storage unit 9 in response to the request and outputs it to the matching unit 4.

  In step S108, the matching unit 4 executes matching between the phoneme string output from the phoneme / syllable processing unit 2 and the search target phoneme / syllable data output from the phoneme / syllable data output unit 3, and the execution. The result is output to the output unit 5.

As a specific example of the expansion rule, for example, a rule that inserts the same vowel so as to be able to cope with fluctuations of a short vowel and a long vowel, that is,
“A” → “aa”, “i” → “ii”, “u” → “uu”, “e” → “ee”, “o” → “oo”
And an expansion rule that removes the same vowel, ie
“Aa” → “a”, “ii” → “i”, “uu” → “u”, “ee” → “e”, “oo” → “o”
The operation for generating a new phoneme string or the like will be described by taking as an example the case where is set.

  When "interphone" is input as a search keyword, expanded phoneme strings such as "iNtahouN", "iNtahoN", and "iNtahoN" are created by applying the above expansion rules to "iNtahooN" that is converted to a phoneme string. . If the search target is the phoneme recognition result and the recognition result corresponding to “interphone” is “iNtahouN” and stored in the phoneme / syllable data storage unit 9, the search is performed with the above-described expanded phoneme string “iNtaahooN”. be able to. Also, if the search target is a large vocabulary continuous speech recognition result converted to a phoneme string, only “interphone” is registered in the recognition dictionary instead of “interphone”, and “interphone” becomes the recognition result, and the phoneme string Can be searched with the above-described expanded phoneme string “iNtaahoN”.

  Other expansion rules include consonants “t” → “d”, “k” → “g”, “d” → “r”, “sh” → “j” replacement, “w” “y” of insertion, such as the deletion of "a", "i", "u", "e", "o" vowel, a case that has been set if the error likely in advance which phoneme is any phoneme as a developing rules as an example, a new phoneme A sequence (or syllable sequence) generation operation will be described.

  For example, it is assumed that “Rebun Island” is to be searched, and “Rebun Island” is stored as “debuNto” in the phoneme / syllable data storage unit 9 in the phoneme recognition result to be searched.

  In this case, first, the search keyword “Rebun Island” is converted into a phoneme string “rebuNtoo”, and the phoneme is expanded for this phoneme string. In this case, it is possible to search for a desired portion in the content by using “debuNto” expanded phoneme string by replacing “d” → “r” and deleting “o”.

  As another example, for example, when it is desired to search for “Prime Koizumi”, in the phoneme string to be searched, “Koizumi Prime” is “koizumijusho” obtained by converting the speech recognition result “Koizumi Award” into a phoneme string. And In this case, first, the search keyword “Prime Minister Koizumi” is converted into a phoneme string “koizumishusho”, and phonemes are expanded for this phoneme string. In this case, a desired part in the content can be searched by using the expanded phoneme string “koizumijusho” by replacing “sh” → “j”.

  By performing a voice search with such a configuration and operation, even when there is a recognition error with respect to the content to be searched, an effective search can be executed without increasing the burden on processing.

[Configuration of Second Embodiment]
FIG. 8 is a block diagram showing the configuration of the second exemplary embodiment of the present invention. The second embodiment of the present invention includes an expansion adjustment unit that determines which rule is applied to execute a search when there are a plurality of expansion rules. Referring to FIG. 8, the configuration of the second embodiment includes a phoneme / syllable number counting unit 40 connected to the phrase expansion unit 6, and each of the phoneme / syllable number counting unit 40 and the phoneme string conversion unit 7. And an expansion adjustment unit 41 connected to the terminal, and adjust the expansion according to the number of phonemes and syllables of the keyword.

  The phoneme / syllable number counting unit 40 is a counter that counts the number of phonemes (or the number of syllables) included in the keyword phoneme string output by the phoneme / syllable conversion means 2. The development adjustment unit 41 is based on the phoneme sequence of the keyword output by the phrase expansion unit 6 and the counted phoneme number (or the number of syllables) output by the phoneme / syllable number counting unit 40. It is an information processing function block which adjusts the expansion | deployment rule used by. The expansion adjustment unit 41 uses a restriction rule or an extension for the expansion rule used in the phoneme string conversion unit 7 based on the number of phonemes (or the number of syllables) output from the phoneme / syllable number counting unit 40 and a preset threshold value. Perform adjustments by adding rules.

  The specific operation of applying this restriction rule will be described below. In the following example, the development adjustment unit 41 stores a threshold value applied to the number of phonemes (or the number of syllables) counted by the phoneme / syllable number counting unit 40, and when the number of phonemes of the search keyword is less than the threshold value Consider the case where the restriction rule is “no expansion”. In general, the shorter the search keyword is, the more false detections occur. Therefore, by providing the development adjustment unit 41 that can add such a restriction, more effective voice search can be performed. The above restriction rule can be arbitrarily changed other than “no expansion”. For example, it is possible to use a restriction rule such as “no phoneme is deleted” or “the number of phonemes / syllable strings after expansion is less than a certain threshold”. It is also possible to store a plurality of these restriction rules and use them in combination.

  Further, the expansion adjustment unit 41 stores a threshold value applied to the number of phonemes (or the number of syllables) counted by the phoneme / syllable number counting unit 40. When the number of phonemes of the search keyword is larger than the threshold value, the expansion adjustment unit 41 Makes it possible to perform effective voice search by using an extended rule “insert, replace, delete at multiple locations” in the keyword.

  FIG. 9 is a block diagram showing another configuration in the second exemplary embodiment of the present invention. Referring to FIG. 9, in another configuration in the second embodiment, a context investigation unit 43 connected to the phrase expansion unit 6 is provided, and is connected to each of the context investigation unit 43 and the phoneme string conversion unit 7. And a deployment adjusting unit 44. The expansion adjustment unit 44 shown in FIG. 9 adjusts the expansion based on the phoneme / syllable context included in the keyword.

  The context investigation unit 43 is an information processing function block that examines a context such as a phoneme string of a search keyword output from the phrase expansion unit 6. The expansion adjustment unit 44 is an information processing function block that adjusts expansion rules used in the phoneme string conversion unit 7 based on the phoneme (or syllable) context output by the context examining unit 43. The expansion adjustment unit 44 adjusts whether or not to expand the context portion in accordance with the phoneme / syllable context included in the search keyword.

  The development adjustment unit 44 stores in advance statistical information of the phoneme / syllable context included in each known word or unknown word from a specific sentence set. The expansion adjustment unit 44 analyzes the context of phonemes and syllables in the search keyword output from the context investigation unit 43, and if there is a context part that is included in many unknown words in the context, expands that part. Context parts that are included in many known words are not expanded.

  As a result, when a speech recognition result converted into phonemes is used as a search target, an appropriate speech search can be executed. A known word, which is a word registered in the language model used for speech recognition, is likely to appear correctly in the phoneme string to be searched, and the processing burden is reduced by not expanding in that case. can do. In addition, unknown words that are not registered in the language model are likely to be misrecognized, and the phoneme string often deviates greatly from the original phoneme string due to the misrecognition. Therefore, an effective search can be performed by performing the expansion and performing a voice search. Furthermore, by examining the frequency of contexts included in a specific sentence set in advance and avoiding the expansion of phoneme strings that contain a large number of contexts that may cause false detection after expansion, errors due to expanded phoneme strings are prevented. It is also possible to prevent an increase in detection.

  FIG. 10 is a block diagram showing another configuration in the second exemplary embodiment of the present invention. Referring to FIG. 10, another configuration in the second embodiment is a configuration including a map 45, a development rule set storage unit 46, a development adjustment unit 47, and a context investigation unit 48. The expansion adjustment unit 47 is connected to the word expansion unit 6, the phoneme string conversion unit 7, the map 45, the expansion rule set storage unit 46, and the context investigation unit 48, and the context investigation unit 48 is connected to the phoneme / syllable data storage unit 9. Is done. In the following description, a case where the nature of content can be identified from the distribution of phonemes / syllable n-grams created from the content will be described as an example.

The expansion adjustment unit 47 is an information processing function block having a function of selecting an expansion rule to be applied to the phoneme string conversion unit 7 from a plurality of expansion rules based on the map 45 and the expansion rule set storage unit 46. The map 45 is an information storage function block that stores a map indicating information “what kind of expansion rule set is valid for what kind of content search target”. The expansion rule set storage unit 46 is an information storage function block that stores a plurality of expansion rules (46 ₁ , 46 ₂ ... 46 _n ) in advance. The context investigation unit 48 is an information processing function block that examines a context such as a phoneme string of a search keyword output from the phoneme / syllable data storage unit 9.

  The context investigation unit 48 investigates the phoneme / syllable n-gram of the search target phoneme / syllable data storage unit 9, selects an expansion rule set suitable for the search target, and selects the expansion selected for the keyword phoneme / syllable string. Apply the rules. Thereby, even when the search target changes, an expansion rule set suitable for the search target can be automatically selected from a plurality of prepared expansion rules. Further, when applying the expansion rules, the expansion pattern may be further adjusted by the keyword phoneme / syllable number or the phoneme / syllable context as described above.

[Operation of Second Embodiment]
FIG. 11 is a flowchart illustrating an example of the operation of the second embodiment. Referring to FIG. 11, the operation of the second embodiment starts when content to be searched is determined. In step S101, a search keyword is input to execute a search for a phoneme string (or syllable string) in the determined content. The input search keyword is output from the search keyword input unit 1 and input to the phoneme / syllable processing unit 2.

  In step S <b> 102, the phoneme / syllable processing unit 2 inputs the search keyword to the phrase expansion unit 6. The phrase expansion unit 6 to which the search keyword is input converts the search keyword into a phoneme string or the like, and proceeds to step S103. In step S <b> 103, the phrase expansion unit 6 determines whether the converted phoneme string should be output to the phoneme string conversion unit 7. As a result of the determination, if it is determined that the phoneme string or the like does not need to be processed by the phoneme string conversion unit 7, the process proceeds to step S106. For example, if the input search keyword is a word with relatively few recognition errors in speech recognition, the phrase expansion unit 6 outputs the phoneme sequence of the word to the matching unit 4 without outputting it to the phoneme sequence conversion unit 7 To do. As a result of the determination in step S103, if it is determined that the phoneme string or the like requires processing by the phoneme string conversion unit 7, the process proceeds to step S201.

  In step S201, the expansion adjustment corresponding to the search keyword converted into a phoneme string or the like is executed. The development adjustment unit (41, 44, 47) outputs the processing result in step S201 to the phoneme string conversion unit 7. In step S104, the phoneme string conversion unit 7 extracts expansion rules based on the processing results output from the expansion adjustment units (41, 44, 47), and the following processing operates in the same manner as in the first embodiment. .

  This makes it possible to apply expansion patterns suitable for keywords and search targets, such as restrictions and expansions to expansion rules, and the use of various expansion rules, enabling more accurate and efficient searches. it can.

[Configuration of Third Embodiment]
FIG. 12 is a block diagram showing the configuration of the third exemplary embodiment of the present invention. Referring to FIG. 12, the third embodiment of the present invention includes a morpheme analysis unit 50 and a known word / unknown word determination unit 51. The morpheme analysis unit 50 is an information processing function block that takes the search keyword output from the search keyword input unit 1 as input, divides it into morphemes, and generates a morpheme string. The morpheme analysis unit 50 is connected to the known word / unknown word determination unit 51 and outputs the generated morpheme string to the known word / unknown word determination unit 51. The known word / unknown word determination unit 51 is an information processing function block that receives the morpheme string output from the morpheme analysis unit 50 and determines whether each morpheme is a known word or an unknown word. The known word / unknown word determination unit 51 matches the input morpheme sequence with a language model used when creating phoneme / syllable data stored in the phoneme / syllable data storage unit 9, and each morpheme is a known word. Or an unknown word, and the determination result of each morpheme is output to the phrase expansion unit 6.

[Operation of Third Embodiment]
FIG. 13 is a flowchart showing an example of the operation of the third exemplary embodiment of the present invention. Referring to FIG. 13, the operation of the third embodiment starts when the content to be searched is determined. In step S101, a search keyword is input to execute a search for a phoneme string (or syllable string) in the determined content. The input search keyword is output from the search keyword input unit 1 and input to the morpheme analysis unit 50.

  In step S301, the morpheme analysis unit 50 divides the input search keyword into morphemes, and generates a morpheme string. The generated morpheme string is output to the known word / unknown word determination unit 51. In step S302, the known word / unknown word determination unit 51 matches the input morpheme sequence with a language model used when creating the search target phoneme / syllable data stored in the phoneme / syllable data storage unit 9, It is determined whether the morpheme is a known word or an unknown word, and the determination result of each morpheme and the search keyword are output to the phrase expansion unit 6.

  In step 102, the phrase expansion unit 6 receives the morpheme-separated keyword output from the known word / unknown word determination unit 51 and the determination result of the known word / unknown word of each morpheme, and converts each morpheme into a phoneme or syllable string. The result is converted and output to the phoneme string conversion unit 7 together with the morpheme-delimited phoneme or syllable string and the determination result of each morpheme known word / unknown word. Hereinafter, the operation is the same as in the first embodiment or the second embodiment.

  As a result, it is possible to specify an expansion rule to be applied based on a known word / unknown word included in the keyword, and a search can be performed with high accuracy and efficiency. Furthermore, by combining with the second embodiment, it is also possible to adjust the expansion pattern and the number of expansions to be applied depending on the known words / unknown words included in the keyword, so that more accurate and efficient search is possible. It can be performed.

  For example, it is assumed that “inflation targeting” is input as an actual keyword. By applying this to morphological analysis, “inflation” and “targeting” are obtained. The known word / unknown word determination unit 51 determines whether each is a known word or an unknown word by matching each with a language model used for speech recognition. In this case, it is assumed that “inflation” is a known word and “targeting” is an unknown word. The known word is likely to appear correctly in the recognition result, and the unknown word is likely to be misaligned from the keyword phoneme string. The phoneme string “iNfure” of the “inflation” that is a known word is developed only in a small amount, and the phoneme string “taagetyiNgu” of the “targeting” that is an unknown word is expanded a lot. As a result, it is possible to perform a highly accurate search while reducing the burden on processing.

FIG. 1 is a block diagram showing a configuration of a conventional voice search apparatus. FIG. 2 is a block diagram showing the configuration of the first embodiment. FIG. 3 is a block diagram showing details of the configuration of the phoneme / syllable data output unit 3. FIG. 4 is a block diagram showing in detail another configuration of the phoneme / syllable data output unit 3. FIG. 5 is a block diagram showing in detail the configuration of the phoneme string conversion unit 7. FIG. 6 is a block diagram showing in detail another configuration of the phoneme string conversion unit 7. FIG. 7 is a flowchart showing the operation of the first embodiment. FIG. 8 is a block diagram showing the configuration of the second embodiment. FIG. 9 is a block diagram showing another configuration in the second embodiment. FIG. 10 is a flowchart illustrating the operation of the second embodiment. FIG. 11 is a block diagram illustrating a configuration of the third embodiment. FIG. 12 is a block diagram showing another configuration according to the third embodiment. FIG. 13 is a flowchart showing the operation of the third embodiment.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 ... Search keyword input part 2 ... Phoneme / syllable processing part 3 ... Phoneme / syllable data output part 4 ... Matching part 5 ... Output part 6 ... Phrase expansion part 7 ... Phoneme sequence conversion part 8 ... Phoneme / syllable data generation part 9 ... Phoneme / syllable data storage unit 20 ... expansion rule output unit 21 ... expansion execution unit 22 ... expansion rule creation phoneme / syllable data storage unit 23 ... statistical processing unit 24 ... expansion rule generation unit 25 ... expansion rule storage unit 26 ... expansion rule Creation voice data storage unit 27 ... Development rule creation phoneme / syllable data storage unit 28 ... Correct rule storage unit 30 ... Speech data storage unit 31 ... Phoneme / syllable recognition unit 32 ... Acoustic model storage unit 33 ... Speech recognition unit 34 ... Phoneme / syllable conversion unit 35 ... Acoustic model storage unit 36 ... Language model storage unit 40 ... Phoneme / syllable number counting unit 41 ... Development adjustment unit 43 ... Context investigation unit 44 ... Exhibition Open adjustment unit 45 ... map 46 ... expansion rule set storage unit,
46 ₁ , 46 _{2 to} 46 _n ... Expansion rule 47... Expansion adjustment unit 48 .. context investigation unit 50 .. morpheme analysis unit 51 .. known word / unknown word determination unit

Claims (33)

A phrase expansion unit that converts input words to generate phoneme strings or syllable strings;
A phoneme that generates a new phoneme string or a new syllable string by adding or subtracting a new phoneme to the phoneme string or the syllable string, or replacing a phoneme constituting the phoneme string or the syllable string with another phoneme A column conversion unit;
Phoneme / syllable data storage unit for storing search target phoneme / syllable data;
A collation unit that collates the phoneme string or the syllable string with the search target phoneme / syllable data, and collates the new phoneme string or the new syllable string with the search target phoneme / syllable data; Voice search device provided. The voice search device according to claim 1,
The search target phoneme / syllable data is composed of a plurality of phonemes,
The collation unit detects, from the search target phoneme / syllable data, a part that matches the new phoneme string or the new syllable string by the collation. The voice search device according to claim 1 or 2,
The phoneme string conversion unit stores an expansion rule that is a rule for generating the new phoneme string or the new syllable string, and based on the expansion rule, the new phoneme string or the new syllable string Generate voice search device. The voice search device according to claim 3.
The search target phoneme / syllable data is generated based on the input voice data,
The expansion rule is set based on a comparison between a plurality of phonemes constituting the search target phoneme / syllable data and a phoneme constituting correct data which is a result of correct phoneme recognition of the speech data. The voice search device according to claim 3.
The search target phoneme / syllable data is generated based on the input voice data,
The expansion rule is set based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data. The voice search device according to any one of claims 1 to 5,
The phrase expansion unit includes morpheme analysis means for analyzing morphemes constituting the input word,
The phoneme string conversion unit generates the new phoneme string or the new syllable string based on the analysis result output from the morpheme analyzing unit and the phoneme string or the syllable string. The voice search device according to claim 6.
The phrase expansion unit includes a registered phrase determination unit,
The registered word determination unit determines whether each morpheme that is a result of analysis by the morpheme analysis unit is registered in advance,
The phoneme string conversion unit generates the new phoneme string or the new syllable string based on the determination result. The voice search device according to any one of claims 1 to 7,
The search target phoneme / syllable data is generated based on a speech recognition result obtained by speech recognition of input speech data or a phoneme recognition result obtained by phoneme recognition of the speech data,
The voice search device is a set of spoken voices. The voice search device according to claim 8.
The search target phoneme / syllable data is generated based on a language model stored in advance,
The language model is information describing connection restrictions of words. The voice search device according to any one of claims 1 to 9,
The phrase expansion unit outputs the phoneme sequence or the syllable sequence to the collation unit when the input word is a word that does not need to generate the new phoneme sequence or the new syllable sequence. Voice search device. The voice search device according to any one of claims 1 to 10,
Phoneme dictionary data or syllable dictionary data for converting the input word,
The phrase expansion unit converts the input word into phonemes or syllables based on the phoneme dictionary data or the syllable dictionary data,
A speech search device that generates the phoneme string or the syllable string based on the converted phoneme or syllable. Converting input words to generate phoneme sequences or syllable sequences;
A step of generating a new phoneme string or a new syllable string by adding or subtracting a new phoneme to the phoneme string or the syllable string or replacing a phoneme constituting the phoneme string or the syllable string with another phoneme When,
Reading stored search target phoneme / syllable data;
Collating the phoneme string or the syllable string with the search target phoneme / syllable data, and collating the new phoneme string or the new syllable string with the search target phoneme / syllable data. A program that can be executed on a computer. The program according to claim 12,
Reading the search target phoneme / syllable data composed of a plurality of phonemes;
A computer-executable program comprising a step of detecting, from the search target phoneme / syllable data, the new phoneme string or a part that matches the new syllable string by the collation. The program according to claim 12 or 13,
A step of reading an expansion rule; and the expansion rule is a rule for generating the new phoneme string or the new syllable string,
A computer-executable program comprising the step of generating the new phoneme string or the new syllable string based on the expansion rule. The program according to claim 14, wherein
Generating the search target phoneme / syllable data based on the input voice data;
A method of setting the expansion rule based on a comparison between a plurality of phonemes constituting the search target phoneme / syllable data and a phoneme constituting correct data which is a result of correct phoneme recognition of the speech data. A program that can be executed on a computer. The program according to claim 14, wherein
Generating the search target phoneme / syllable data based on the input voice data;
A computer-executable program comprising: a step of setting the expansion rule based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data. The program according to any one of claims 12 to 16,
Analyzing the morphemes that make up the input word;
A computer-executable program comprising a step of generating the new phoneme string or the new syllable string based on the analysis result and the phoneme string or the syllable string. The program according to claim 17, wherein
Determining whether each of the morphemes is pre-registered;
A computer-executable program comprising a step of generating the new phoneme string or the new syllable string based on the determination result. The program according to any one of claims 12 to 18,
Reading the input voice data, and the voice data is a set of spoken voices;
A computer-executable program comprising: generating the search target phoneme / syllable data based on a voice recognition result obtained by voice recognition of the voice data or a phoneme recognition result obtained by phoneme recognition of the voice data. The program according to claim 19, wherein
A step of reading a language model stored in advance, and the language model is information describing a connection constraint of words;
A computer-executable program comprising the step of generating the search target phoneme / syllable data based on the language model. The program according to any one of claims 12 to 20,
Performing a determination as to whether the input word needs to generate the new phoneme string or the new syllable string;
As a result of the determination, if the input word is a word that does not require generation of the new phoneme string or the new syllable string, the method includes the step of outputting the phoneme string or the syllable string. A computer executable program. The program according to any one of claims 12 to 21,
Reading phoneme dictionary data or syllable dictionary data for converting input words;
Converting the input word into phonemes or syllables based on the phoneme dictionary data or the syllable dictionary data;
A computer-executable program comprising a step of generating the phoneme string or the syllable string based on the converted phoneme or syllable. Converting input words to generate phoneme sequences or syllable sequences;
A step of generating a new phoneme string or a new syllable string by adding or subtracting a new phoneme to the phoneme string or the syllable string or replacing a phoneme constituting the phoneme string or the syllable string with another phoneme When,
Reading stored search target phoneme / syllable data;
Collating the phoneme string or the syllable string with the search target phoneme / syllable data, and comparing the new phoneme string or the new syllable string with the search target phoneme / syllable data. Voice search method. 24. The voice search method according to claim 23.
Reading the search target phoneme / syllable data composed of a plurality of phonemes;
A speech search method comprising a step of detecting, from the search target phoneme / syllable data, a part that matches the new phoneme string or the new syllable string by the collation. The voice search method according to claim 23 or 24,
A step of reading an expansion rule; and the expansion rule is a rule for generating the new phoneme string or the new syllable string,
A speech search method comprising the step of generating the new phoneme string or the new syllable string based on the expansion rule. 26. The voice search method according to claim 25.
Generating the search target phoneme / syllable data based on the input voice data;
A speech search comprising a step of setting the expansion rule based on a comparison between a plurality of phonemes constituting the search target phoneme / syllable data and a phoneme constituting correct data which is a result of correct phoneme recognition of the speech data. Method. 26. The voice search method according to claim 25.
Generating the search target phoneme / syllable data based on the input voice data;
A speech search method comprising a step of setting the expansion rule based on statistics of appearance frequencies of a plurality of phonemes constituting the search target phoneme / syllable data. The voice search method according to any one of claims 23 to 27,
Analyzing the morphemes that make up the input word;
A speech search method comprising a step of generating the new phoneme string or the new syllable string based on the analysis result and the phoneme string or the syllable string. The voice search method according to claim 28, wherein
Determining whether each of the morphemes is pre-registered;
A speech search method comprising a step of generating the new phoneme string or the new syllable string based on the determination result. The voice search method according to any one of claims 23 to 29,
Reading the input voice data, and the voice data is a set of spoken voices;
Generating a search target phoneme / syllable data based on a speech recognition result obtained by speech recognition of the speech data or a phoneme recognition result obtained by phoneme recognition of the speech data. The voice search method according to claim 30, wherein
A step of reading a language model stored in advance, and the language model is information describing a connection constraint of words;
A speech search method comprising: generating the search target phoneme / syllable data based on the language model. The voice search method according to any one of claims 23 to 31,
Performing a determination as to whether the input word needs to generate the new phoneme string or the new syllable string;
As a result of the determination, if the input word is a word that does not require generation of the new phoneme string or the new syllable string, a step of outputting the phoneme string or the syllable string. The voice search method according to any one of claims 23 to 32,
Reading phoneme dictionary data or syllable dictionary data for converting input words;
Converting the input word into phonemes or syllables based on the phoneme dictionary data or the syllable dictionary data;
A speech search method comprising: generating the phoneme string or the syllable string based on the converted phoneme or syllable.

Images