JP2011175046A - Voice search device and voice search method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice search device and a voice search method, for detecting desired voice from a large amount of voice data including unknown words or incorrect recognition which does not exist in a dictionary, by voice and text input. <P>SOLUTION: An index is attached to the unknown word, and a plurality of detection candidates are generated for the index of the unknown word by dividing the unknown word or search word for the voice data of the unknown word into phonemes or syllables, in large vocabulary continuous voice recognition for recognizing the voice data from search input, by the search input of voice or texts. Thereby, a search result is presented from the large amount of voice data including unknown words or incorrect recognition which does not exist in the dictionary. <P>COPYRIGHT: (C)2011,JPO&INPIT

Description

  The present invention relates to a voice search apparatus and a voice search method for detecting a desired voice from voice data by voice and text input.

  On the Internet, there is a large amount of audio information such as news audio, video postings, podcasts (these are referred to as audio documents), and the amount is increasing year by year. In addition, voice data held by individual organizations such as conference voices, lecture voices, and call center voices is increasing. It is required to detect information desired by a user at high speed from such enormous audio data. In addition, many search request words include proper nouns and new coined words (eg, swine flu), and many words do not exist in existing large-scale dictionaries. In the conventional text search method, such a word is correctly expressed as a character string. Therefore, if a search word is input as a character string, it is often searched correctly and there are few problems.

  On the other hand, for speech documents, it is usually easiest to use a text search technique after converting speech into a word string with a large vocabulary continuous speech recognizer. Words (unknown words) that do not exist in words to 100,000 words cannot be searched. In addition, there are many voice recognition errors, and a normal text search cannot be performed.

  There is a method for enlarging a recognition dictionary and reducing unknown words for an unknown word search word in an audio document (for example, Non-Patent Document 1). However, it is impossible to register all proper nouns as a dictionary, and recognition errors cannot be avoided even if the dictionary is enlarged. Is not resolved.

Therefore, for unknown words, speech documents are converted into symbolic strings such as phonemes and syllables (which may also be other basic units), which are the basic units of spoken language, and expressed with these symbolic strings. Basically, matching is performed with the generated symbol string (for example, Non-Patent Document 2). In order to deal with recognition errors in this phoneme / syllable string, it is common to efficiently represent the recognition results of a plurality of candidates in a graph structure and search for this (for example, Patent Document 1 and Non-Patent Document 3). . Various approaches have been made to speed up the phoneme / syllable string matching of this and the search word (Non-Patent Documents 3 and 4), but it takes time in proportion to the amount of speech documents to be searched. There's a problem.

JP 2009-271117 A

Kurijo Rio, Ito Yoshiaki, Kojima Kazunori, Ishigame Masaaki, Tanakayo: Search term detection by query tone using vocabulary on the Web, IEICE, Spoken Technical Report, SP2009-84 (2009.12) Takaaki Hori, et al .: Open vocabulary utterance search using confusion network, Proc. Of Acoustical Society of Japan, 1-3-10 (2007.9) Yoshiaki Ito, et al .: Integration of multiple subwords in vocabulary-free spoken document retrieval, information processing is the society journal, Vol.50, No.2, pp.524-533 (2009.2) Naoyuki Kanda, Takashi Sumiyoshi, Masato Togami, Yasunari Otsuki: Performance Evaluation of Multi-stage Rescoring Method for Arbitrary Vocabulary Speech Search, Proceedings of the Second Speech Document Processing Workshop, pp.73-78 (2008.2)

  Although a fast search robust to unknown words is desired for a large amount of speech documents, as described above, the conventional method uses speech recognition technology to convert speech documents to phonemes, which are the basic units of speech languages. Basically, after conversion to a symbol string such as a syllable, matching with a symbol string represented by these symbol strings is performed. In order to deal with this recognition error, it is a conventional method to efficiently represent the recognition results of a plurality of candidates in a graph structure and perform a search for this.

  However, there is a problem that it takes time in proportion to the amount of the audio document to be searched and a recognition error that does not exist in a plurality of recognition result candidates cannot be dealt with.

  Two methods have been proposed to deal with this problem. One is a method in which a phoneme symbol string is indexed by a transposed index method, and an index table is searched in consideration of recognition errors such as substitution errors at the time of search. Although this method can increase the speed, there is a problem that it takes time to search for an index.

  The other method is an indexing method in consideration of recognition errors of phonemes / syllable strings. As a result, search can be performed at high speed, but search request words are often detected excessively. Therefore, in the conventional method, symbol strings are collated in detail with the conventional method to narrow down candidates. . This method has a problem that the detailed collation portion takes time in proportion to the time length of the voice document.

In order to solve the above-described problems in the prior art, the inventor has invented a voice search device and a voice search method having the following characteristics.

The voice search device according to claim 1 is:
Search term input part by voice or text, continuous voice database storage part to be searched,
A large vocabulary continuous speech recognition unit for recognizing speech data from the input unit and the database storage unit;
A continuous speech data recognition result storage unit for storing a recognition result of the large vocabulary continuous speech recognition unit;
An unknown word index creating unit for indexing unknown words in the continuous speech database;
A phoneme / syllable recognition unit that recognizes speech data of unknown words by dividing them into phonemes or syllables, which are basic units of speech,
A phoneme / syllable data storage unit for storing a recognition result of the phoneme / syllable recognition unit;
A phoneme / syllable string search unit that presents at least one search candidate, wherein the phoneme / syllable recognition unit detects a plurality of unknown words assigned the index It has a function of generating candidates.

The voice search device according to claim 2,
The voice search device according to claim 1,
The phoneme / syllable recognition unit has a function of assuming a substitution error and an insertion error from at least one recognition result candidate, assigning the recognition result as an index, and presenting a detection candidate.

The voice search device according to claim 3 is:
The voice search device according to claim 1 or 2,
The phoneme / syllable recognition unit has a function of assuming a drop error from at least one recognition result candidate, assuming the drop error as a search word, adding it as an index, and presenting a detection candidate. And

The voice search device according to claim 4 is:
The voice search device according to claim 3,
The phoneme / syllable recognition unit assumes a substitution error, an insertion error, and / or a drop error from at least one recognition result candidate, assigns the recognition of the error as an index, and uses information of divided search terms And a function of selecting detection candidates from the detection candidates obtained based on a preset threshold value.

The voice search device according to claim 5 is:
The voice search device according to claim 1, wherein:
The phoneme / syllable recognition unit assigns an index using at least one recognition result candidate using a virtual distance between phonemes and between syllables,
A phoneme / syllable recognition unit having a function of presenting detection candidates based on a distance from the second recognition result candidate or the third recognition result candidate based on the acoustic similarity with the first recognition result candidate And Various definitions of the distance between phonemes and syllables are conceivable, and usually the one corresponding to the scale used in the recognition system is used.

The voice search device according to claim 6 is:
The voice search device according to claim 1, wherein:
The phoneme / syllable recognition unit assigns an index to the at least one recognition result candidate using the log likelihood defined by Equation 1,

認識結果の対数尤度に基づいて、検出候補を提示する機能を具備することを特徴とする。

It has a function of presenting detection candidates based on the log likelihood of the recognition result.

The voice search method according to claim 7 comprises:
A search word input step by speech or text, a continuous speech database storage step to be searched, a large vocabulary continuous speech recognition step for recognizing speech data from the input unit and the database storage unit,
A step of storing a continuous speech data recognition result for storing a recognition result of the large vocabulary continuous speech recognition unit;
An unknown word index for indexing unknown words in the continuous speech database;
A phoneme / syllable recognition step for recognizing the speech data of unknown words by dividing them into phonemes or syllables which are the basic units of speech;
Storing a phoneme / syllable recognition result storing step for storing a recognition result of the phoneme / syllable recognition unit;
A phoneme / syllable search step for presenting at least one search candidate,
The phoneme / syllable recognition step has a function of generating a plurality of detection candidates for an unknown word to which the index is assigned.

The voice search method according to claim 8 comprises:
The voice search method according to claim 7,
The phoneme / syllable recognition step has a function of assuming a substitution error and an insertion error from at least one recognition result candidate, assigning the recognition as an index, and presenting a detection candidate.

The voice search method according to claim 9 comprises:
The voice search method according to claim 7 or 8,
The phoneme / syllable recognition step has a function of assuming a drop error from at least one recognition result candidate, assuming recognition of the drop error as a search word, giving it as an index, and presenting a detection candidate. Features.

The voice search method according to claim 10 comprises:
The voice search method according to claim 9,
The phoneme / syllable recognition step assumes a substitution error, an insertion error, and / or a drop error from at least one recognition result candidate, assigns the recognition of the error as an index, and uses information of divided search terms It has the function to select a detection candidate from the detection candidate obtained by using a preset threshold as a reference.

The voice search method according to claim 11 comprises:
The voice search method according to claim 7, wherein:
In the phoneme / syllable recognition step, at least one or more detection candidates are indexed using a virtual distance between phonemes and between syllables,
It has a function of presenting a detection candidate by a distance from the second detection candidate or the third detection candidate based on the acoustic similarity with the first detection candidate. Various definitions of the distance between phonemes and syllables are conceivable, and usually the one corresponding to the scale used in the recognition system is used.

The voice search method according to claim 12 includes:
12. The voice search method according to claim 7, wherein:
In the phoneme / syllable recognition step, at least one or more detection candidates are indexed using a log likelihood defined by Equation 1,

認識結果の対数尤度に基づいて、検出候補を提示する機能を有することを特徴とする。

It has a function of presenting detection candidates based on the log likelihood of the recognition result.

From the above, for known words, it is converted into a word string by large vocabulary continuous speech recognition, and for unknown words or recognition error words, a phoneme string or syllable string that is a basic unit rather than a word by phoneme / syllable recognition From the recognition, it is possible to provide a voice search device and a voice search method based on voice and text input from a large amount of voice data containing unknown words or recognition errors that do not exist in the dictionary.
In addition, when a search candidate is presented using the divided search information, the search efficiency can be improved by selecting the detection candidate based on a threshold set in advance.

It is the block diagram (a) of a speech search device, and the block diagram (b) explaining the algorithm of a speech search method. It is explanatory drawing which shows the preparation procedure of a trigram array. It is explanatory drawing which shows the preparation procedure of a trigram array in case a substitution error is included. It is explanatory drawing which shows the preparation procedure of a trigram array in case an insertion error is included. It is explanatory drawing which shows the division | segmentation to the trigram of a search word. It is explanatory drawing which shows the preparation procedure of a trigram array in case a drop error is included. It is explanatory drawing which shows the replacement error countermeasure by combined use of an insertion error countermeasure and a drop error countermeasure. It is explanatory drawing which shows the internal representation of a trigram array.

  The present invention provides a means capable of searching for speech at a high speed with respect to a large amount of speech database including recognition errors and unknown words.

  Specifically, a known word is converted into a word string by a conventional large vocabulary recognition apparatus and then searched by a normal text search method. On the other hand, for unknown words and recognition error words, a syllable string or phoneme string, which is a basic unit rather than a word, is recognized, and a search is performed on the result. Since unknown words cannot be recognized correctly in phoneme strings or syllable strings, it is assumed that there is a recognition error. Recognition errors include substitution errors, insertion errors, and omission errors. For replacement errors, recognition results of a plurality of candidates are used. If this is not enough, it can be dealt with by using the following countermeasure against insertion error and dropping error. For insertion errors, search is performed in consideration of insertion of recognition results. To cope with omission errors, the phoneme / syllable string of the search word is dropped.

  The speech database to be searched is recognized in advance offline, and the result is n-phonemes / syllables taking into account errors, for example, in the case of 3 syllables, a triple array (hereinafter referred to as trigram array) And a distance or likelihood indicating the position of the recognition result and the degree of recognition error is added to this and stored as an index in dictionary order. For the search word, the word is divided into three syllable units, the trigram array index is searched by a binary search method or the like, the divided search results are integrated, and the final search result is output. By this method, it is possible to search at high speed without lowering the search accuracy compared to the conventional method. The search time is within one second for 10,000 hours of audio data. The storage amount of the trigram array index is also smaller than the original audio file data amount, which is practical.

  Both the text input and the voice input are possible for the input of the search word to the apparatus. In the case of speech input, it is converted into text by a large vocabulary continuous speech recognition device or a phoneme / syllable continuous speech recognition device. The latter recognition error is dealt with by the same method as the speech recognition error countermeasure for the retrieval target speech data.

  The core of the present invention is that the distance or likelihood corresponding to the recognition error is attached to the index, and the search can be performed at high speed with the same performance without performing detailed matching with the search character string. That is.

  Embodiments according to the present invention will be described with reference to the drawings. The following are merely illustrative examples, and details of implementation are not limited to the illustrative examples and figures.

  FIG. 1 shows a block diagram of an audio document retrieval apparatus according to the present invention and an overall processing flow. First, the recognition and indexing of voice documents performed offline will be described.

  The search target speech is taken out from the storage unit (a) of the speech document data to be searched, and the large vocabulary continuous speech recognition device (a) for known word search is performed and converted into a word string. The vocabulary continuous speech recognition result is stored in the storage unit (c) (S3). For the recognition result of the word string of known words, the index creation unit (d) for searching for known words tabulates the words and their appearance positions with the data structure of the transposed index.

  In parallel with this, the phoneme / syllable recognition unit (e) for recognizing a recognition error word / unknown word also recognizes a phoneme or syllable string which is a basic recognition unit of speech. In particular, since recognition of phonemes / syllables is difficult, multiple recognition candidates are output and output in a lattice format (this output format is optional) and stored in the storage unit (f) for phoneme / syllable recognition results. (S4). The phoneme / syllable string recognition result for the recognition error word / unknown word search is also indexed in the same way as the known word in the phoneme / syllable index creation unit (ki) for the unknown word / recognition word search. It is converted into a table and indexed by the following method.

  A unit of n phonemes / syllable strings (hereinafter referred to as n-gram) is used as a unit, and the head position and the distance in the recognition result of the n-gram and the phoneme / syllable string are used as an index and are summarized in a table. Create n-grams by shifting one phoneme and syllable from beginning to end of a voice document. Arranged in the order of the dictionary so that the sentence search can be performed at high speed (S5).

  The procedure of S5 will be described using the example of FIG. This example is a case where there is only one recognition candidate. The n-gram created using only the first candidate phoneme / syllable recognition result has a distance of zero.

  For simplicity of explanation, n = 3. Since the first 3 gram (trigram) of the syllable unit of the voice document is “fu uri” and the head position is 0, the index is 0, the distance is 0, and the insertion error is 0. The next 3 grams (trigram) is “u ri e”, the index is 1, the distance is 0, and the insertion error is 0. Thereafter, the same operation is performed. After this is done to the end of the audio document, the trigrams are arranged in dictionary order. When the same trigram exists in a plurality of places, a plurality of the same trigrams are arranged when arranged in dictionary order. There are various variations of the storage method when the same trigram continues. For example, when the same trigrams are arranged, only one can be represented and the other can be stored in a separate table.

  FIG. 3 shows a case where recognition results of a plurality of candidates are used for countermeasure against substitution error. Here, the number of syllable recognition candidates is three. The trigram index is created in the same way as in FIG. 2, but as shown in the example of FIG. 3, the trigram head is selected as the first candidate, the middle as the second candidate, and the tail as the third candidate. This shows the case of creating a gram. The distance in this case is obtained as the sum of the distances from the first candidate. That is, d (e, u) + d (ki, ri). Here, d (syllable i, syllable j) indicates the distance between syllable i and syllable j. The distance between syllables is defined in advance and stored in a table. The distance between the phonemes and syllables is defined by the batcha rear distance. The distance between the syllable a and the syllable b is expressed as in Equation 2.

バタチャリヤ距離は、多次元正規分布間の距離を表わすもので、音節のモデルは複数個の正規分布の和からなるＭ個の状態で表わされる。 (Batachariya distance)

The batcha rear distance represents a distance between multi-dimensional normal distributions, and a syllable model is represented by M states composed of a sum of a plurality of normal distributions.

  In this example, it is 0 because there is no insertion error. In this way, trigram indexes are created for all combinations in consideration of a plurality of recognition result candidates.

  FIG. 4 shows an example of countermeasure against insertion error. For simplicity, only the recognition result of the first candidate is shown. A trigram index is created for this sequence assuming an insertion error. The example of FIG. 4 shows an example of “fu uri” in which a trigram is created by skipping this syllable, assuming that the recognition result “ku” at position 1 is an insertion error. In this case, since it was created assuming an insertion error, the insertion column is 1. The insertion error is not assumed to be unlimited, but is made to match the insertion error tendency of an actual speech recognition apparatus. For example, assume that one syllable of three syllables can be inserted.

  Actually, since both replacement errors and insertion errors are dealt with simultaneously, the operations shown in FIGS. 3 and 4 are all performed on the recognition results of a plurality of candidates.

  Using the above method, search-response voice documents are indexed offline. In contrast, an online search for a search term will be described.

  When the search word from the input part (g) (S1) of the search word input by typing or voice input is a known word (S2), a normal text search technique is performed by the search part (g) for the known word. (S6), and a search result is obtained (S7). The search result of the known word is displayed to the user by the search result display unit (sa) of the known word.

  On the other hand, when the search word is an unknown word search (when it is not in the dictionary for speech recognition) (S2), it is converted into a phoneme / syllable string by the search unit (co) for unknown words / recognized words. (S8), divided into n consecutive units (that is, n-grams) (here, m-divided), and independently for each n-gram unit, the audio document is indexed in n-gram units by the above-described method. The index table that has been converted is searched at a high speed by the binary search method (S9), and a search result is obtained (S10). The search result of the unknown word / recognition error word is displayed to the user by the display unit (b) of the search result of the unknown word / recognition error word.

  Various modifications are conceivable for the above-described high-speed search for unknown / recognized words. For example, since the search unit is limited to a fixed length of trigram, it is possible to use a calculation method or a table in which the given trigram is in one-to-one correspondence with the table. The retrieved result is generally composed of the appearance positions and scores (with distances and likelihoods) of a plurality of places for every m n-grams. Of the n n-gram search results candidates that have been independently searched, those in which the appearance positions continue without overlapping each other are regarded as correct search position candidates. Among them, m consecutive scores are added, and a search result satisfying a preset threshold value is set. At this time, the score reflects whether the search result is based on an n-gram assuming an insertion error, and whether the search result is based on an n-gram assuming a drop-off error in a search word to be described later. There are various ways to reflect the information. If the search term cannot be divided into n-gram units, it is divided while allowing overlap. For example, when a word consisting of 7 syllables is divided into 3 grams, it is divided into 3 at positions 1 to 3, 3 to 5 and 5 to 7). FIG. 5 shows another method of dividing the syllable length.

  For missing errors in the recognition result, the phoneme / syllable string of the search word is allowed to be dropped, and the search is performed in the same manner as a new search word. However, the dropout error is not assumed to be unlimited, but is matched with the dropout error tendency of the speech recognition apparatus. Usually, dropout is assumed at a rate of one in three consecutive syllables. FIG. 6 shows a 3 gram creation method that copes with a search term omission error. As described above, when an unknown word search word is given, assuming a drop error, the search word is divided into a plurality of n-gram units, and these are independently searched in the index table. Based on the search results, integration is performed to obtain a search term result. In the trigram when dropping is considered, this is stored as shown in FIG.

  FIG. 7 shows an example of replacement error countermeasures by using a combination of countermeasures against insertion errors of recognition results and dropout error countermeasures using search words. By using both of these, substitution errors can be dealt with.

Here, the recognition error has been described using the scale of the distance from the first candidate, but the likelihood (logarithmic posterior probability) representing the probability of the recognition result is attached to the output of the recognition device. This value can also be used. The log posterior probability is defined by Equation 1 above. Number 1 represents the logarithmic probability of the input feature parameter sequence _{a i a i + 1 ··· a} j from the i time segment of speech input pattern to the j-th time interval by HMM syllable string S.

  The method of the present invention was carried out on an academic speech database for 44 hours of speech utterance length of a Japanese spoken corpus. The voice search device of FIG. 1 is stored in C language on a personal computer (Intel (registered trademark) Xeon (registered trademark) X5365, 3 GHz, memory 33 GB) to which a CPU, a memory, an external storage device and the like are significantly electrically connected. Constructed using. In particular, in order to evaluate retrieval of unknown words, first, continuous syllable recognition is performed, and a recognition result is output up to the third candidate (referred to as a syllable lattice). The trigram array was indexed for this recognition result. FIG. 8 shows an internal representation in the storage device of the trigram array (SIL indicates a beginning symbol). In the case of Japanese syllables, there are 116 types of syllables including foreign language expressions, so the combination of index and 3 syllables can be stored as 4 bytes = 1 long byte integer type. The storage capacity of the index table of the audio data for 44 hours was 1.5 Gbytes. This is less than the storage amount of the original speech waveform (3600 × 44 hours × 16 kHz × 2 bytes) = 5 Gbytes.

  The distance between syllables is defined as the distance between the normal distributions of the speech feature vectors in each state in the HMM in syllable units. When an insertion error is considered, the search score is added to the score by α times the number of insertions, and when a drop error is assumed, β times the number of deletions.

  The search term is uttered only 4 times or less (once every 10 hours) in the 44 hours of speech data (about 220,000 words uttered), and does not exist in the dictionary of the large vocabulary continuous speech recognition apparatus of 20000 words. 43 words were used as unknown words. The total number of appearances is 142, and the average number of appearances is 3 times per word (once every 15 hours, in other words, only once during 70,000 words). It can be easily understood that this search is a very difficult problem.

  First, Table 1 shows the performance of the syllable recognition rate.

here,
Accuracy rate = 1.0-replacement rate-dropout rate,
Recognition accuracy = 1.0-replacement rate-dropout rate-insertion rate,
It is. This performance is the performance of the speech recognition apparatus itself, and the performance is improving year by year, but is not directly related to the present invention. However, the higher the syllable recognition performance, the better the search performance, so care must be taken in interpreting the evaluation results of the present invention. From Table 1, the percentage of correctly recognized syllables uttered by the third candidate is 87%. The insertion error rate is 3% and the drop error rate is 6%.

  Next, Table 2 shows the search results for unknown words.

  For comparison, Table 3 shows the results obtained by a method of searching with a simple trigram without using a trigram with distance, which is the basis of the present invention, and deleting excessive search candidates by detailed DP matching.

here,
Reproducibility = number correctly detected / total number of search terms,
Match rate = number correctly detected / number detected. In the table, “without narrowing down” indicates a result when searching with a conventional trigram array without a distance. Table 3 shows a case where candidate sections are narrowed down by DP matching and in detail by collating syllables. Comparing Table 2 and Table 3, it can be seen that the performance is almost the same. Roughly speaking, when an unknown word that appears 4 times in 44 hours (once every 10 hours) is detected, 20 candidate locations (once every 2 hours) are detected, and 2 of them are correct search results. Is performance.

On the other hand, the search time is 2.5 ms per search word in the method according to the present invention, and 15 ms in our conventional method using DP matching together. The method of the present invention is proportional to the logarithm of the speech time length to be searched, while in our conventional method, it is linearly proportional. For example, when searching for 10,000 hours of speech data, the method of the present invention can be searched in about several tens of ms per search word, and it takes about 5 seconds with our conventional method.

A: Storage unit for speech document data to be searched b: Large vocabulary continuous speech recognition unit for known word search C: Storage unit for large vocabulary continuous speech recognition results D: Index creation unit for known word search E: Recognition Phoneme / syllable recognition unit for error word / unknown word search K: Phoneme / syllable recognition result storage unit K: Phoneme / syllable index creation unit for unknown word / recognition error word search K: Search word input Part (typing input or voice input)
K: Search section for known words E: Search section for unknown words / recognized error words S: Search result display section for known words S: Display section for search results of unknown words / recognized error words

Claims (12)

Search term input part by voice or text, continuous voice database storage part to be searched,
A large vocabulary continuous speech recognition unit for recognizing speech data from the input unit and the database storage unit;
A continuous speech data recognition result storage unit for storing a recognition result of the large vocabulary continuous speech recognition unit;
An unknown word index creating unit for indexing unknown words in the continuous speech database;
A phoneme / syllable recognition unit that recognizes speech data of unknown words by dividing them into phonemes or syllables, which are basic units of speech,
A phoneme / syllable recognition result storage unit for storing a recognition result of the phoneme / syllable recognition unit;
A phoneme / syllable string search unit that presents at least one search candidate, wherein the phoneme / syllable recognition unit detects a plurality of unknown words assigned the index A voice search device comprising a function for generating candidates. The voice search device according to claim 1,
The phoneme / syllable recognition unit has a function of assuming a replacement error and an insertion error from at least one recognition result candidate, assigning the recognition as an index, and presenting a detection candidate. apparatus. The voice search device according to claim 1 or 2,
The phoneme / syllable recognition unit has a function of assuming a drop error from at least one recognition result candidate, assuming that the drop error is recognized as a search word, giving it as an index, and presenting a detection candidate. Voice search device characterized by the above. The voice search device according to claim 3,
The phoneme / syllable recognition unit assumes a replacement error, an insertion error, and / or a drop error from at least one recognition result candidate, assigns the error recognition as an index, and uses the divided search information. A speech search apparatus comprising a function of selecting detection candidates from a set of detection candidates based on a preset threshold value. The voice search device according to claim 1, wherein:
The phoneme / syllable recognition unit assigns an index using at least one recognition result candidate using a virtual distance between phonemes and between syllables,
A speech search apparatus comprising a function of presenting detection candidates based on a distance from the second recognition result candidate or the third recognition result candidate based on the acoustic similarity with the first recognition result candidate .
The voice search device according to claim 1, wherein:
The phoneme / syllable recognition unit assigns an index to at least one or more detection candidates using a log likelihood defined by Equation 1,

A voice search device comprising a function of presenting detection candidates based on log likelihood with recognition result candidates.
A search word input step by voice or text, a continuous voice database storage step to be searched,
A large vocabulary continuous speech recognition step for recognizing speech data from the input unit and the database storage unit;
A step of storing a continuous speech data recognition result for storing a recognition result of the large vocabulary continuous speech recognition unit;
An unknown word index for indexing unknown words in the continuous speech database;
A phoneme / syllable recognition step for recognizing the speech data of unknown words by dividing them into phonemes or syllables which are the basic units of speech;
Storing a phoneme / syllable recognition result storing step for storing a recognition result of the phoneme / syllable recognition unit;
A phoneme / syllable search step for presenting at least one search candidate,
The phoneme / syllable recognition step has a function of generating a plurality of detection candidates for an unknown word to which the index is assigned. The voice search method according to claim 7,
The phoneme / syllable recognition step has a function of assuming a substitution error and an insertion error from at least one recognition result candidate, assigning the recognition as an index, and presenting a detection candidate. . The voice search method according to claim 7 or 8,
The phoneme / syllable recognition step has a function of assuming a drop error from at least one recognition result candidate, assuming that the drop error is recognized as a search word, giving it as an index, and presenting a detection candidate. A featured voice search method. The voice search method according to claim 9,
The phoneme / syllable recognition step recognizes a substitution error, an insertion error, and / or a drop error from at least one recognition result candidate, assigns the error recognition as an index, and uses information of divided search words A speech search method characterized by having a function of selecting detection candidates from detection candidates obtained by using a preset threshold as a reference. The voice search method according to claim 7, wherein:
In the phoneme / syllable recognition step, at least one recognition result candidate is indexed using a virtual distance between phonemes and between syllables,
A speech search method characterized by having a function of presenting a detection candidate based on a distance from the second recognition result candidate or the third recognition result candidate based on the acoustic similarity with the first recognition result candidate. 12. The voice search method according to claim 7, wherein:
In the phoneme / syllable recognition step, at least one recognition result candidate is indexed using a log likelihood defined by Equation 1,

A speech search method characterized by having a function of presenting detection candidates based on a log likelihood of a recognition result.

Images