JP2002278579A - Voice data retrieving device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice data retrieving device accurately searching a part, including a desired speech at a high speed from voice data stored in a large volume. SOLUTION: This voice data retrieving device is constituted of a voice data registration part 1 converting digitized voice waveform data 4 to a preset voice symbol sequence and recording it, a candidate voice section detection part 2 for converting a retrieval word to a voice symbol and retrieving a matching part from a registered symbol sequence 7 and a retrieval word voice determining part 3 for determining whether or not a candidate section detected in the candidate voice section detection part matches with the retrieval word. Thus, the candidate section is narrowed at a high speed at a symbol level and accurate detection by matching at a voice waveform level is conducted.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an audio data retrieval apparatus for quickly and accurately retrieving an utterance portion containing desired contents from a large amount of accumulated audio waveform data.

[0002]

2. Description of the Related Art Conventionally, there has been proposed an audio data retrieval apparatus for retrieving desired audio data from audio data storage means in which audio data is stored (Japanese Patent Laid-Open No. 2000-02).
0551). In this voice data search device, at the time of registration, voice data is divided into sections, the existence probabilities of words that are vocabulary selected in advance in each voice section are calculated and stored, and at the time of search, synonyms of the input search word are included. Into a search word group, and a speech section having the highest existence probability is output.

[0003]

However, in this method, there is a limitation that only search words registered in a dictionary in advance can be handled, and if the vocabulary is increased in order to alleviate this limitation, the storage capacity is greatly increased. There is a problem of being connected. Furthermore, since the calculation of the existence probability of the search word group performs a full search in all voice sections, there is a problem that as the volume of voice data increases, the search time also increases in proportion thereto.

Accordingly, the present invention has been made in view of such a problem, and an audio data search apparatus capable of quickly and accurately finding a portion including a desired utterance from a large amount of accumulated audio data. It is intended to provide.

[0005]

The above object is achieved by the following means of the present invention.

An audio data search device according to the first aspect of the present invention includes an audio data registration unit that converts digitized audio waveform data into a predetermined audio symbol sequence and records the converted data.
A candidate voice section detection unit that converts a search word into a voice symbol and searches for a matching part from a registered symbol sequence; and determines whether a candidate section detected by the candidate voice section detection unit matches the search word. And a search word / voice determination unit.

According to a second aspect of the present invention, there is provided the voice data search apparatus according to the first aspect, wherein the voice data registration unit uses a voice parameter string used when extracting voice symbols from voice data. Is stored, and the voice parameter string is shared by the search word voice determination unit.

According to a third aspect of the present invention, there is provided the voice data search apparatus according to the first aspect, wherein the candidate voice section detection unit performs a search using a full-text search system to search for a voice symbol. It is characterized by the following.

According to a fourth aspect of the present invention, there is provided the voice data searching apparatus according to the first aspect, wherein the search word voice determining unit is capable of treating a voice at a beginning and an end of the voice freely. The apparatus is used for determination.

According to a fifth aspect of the present invention, there is provided the voice data search apparatus according to the first aspect, wherein the voice data registration unit stores only a voiced vowel, a long vowel, a vowel sound, and a silent voice section. It is characterized by being converted into a voice symbol string for use.

According to a sixth aspect of the present invention, in the voice data search apparatus according to the first aspect, the voice data registration unit is based on a single syllable.
The present invention is characterized in that consonants that easily cause confusion are grouped and treated as one syllable, and a symbol group is used as a voice symbol sequence.

According to a seventh aspect of the present invention, in the voice data search apparatus according to the fifth or sixth aspect, the voice data registration unit preliminarily registers words having a high appearance frequency in a dictionary. It is characterized in that words are added as one voice symbol.

According to the present invention, first, voice data is converted into a symbol capable of expressing characters, and a candidate voice section is extracted by matching at the character level. Next, pattern matching of the voice waveform data is performed in each detected voice section, and it is determined whether or not a search word is actually included. High speed and accuracy can be realized by the two-stage processing.

[0014]

Embodiments of the present invention will be specifically described below with reference to the drawings.

FIG. 1 shows a configuration diagram of a voice data search device according to an embodiment of the present invention. As shown in FIG. 1, the voice data search device includes a voice data registration unit 1, a candidate voice section detection unit 2, and a search word voice determination unit 3. The audio data registration unit 1 converts the digitized audio waveform data 4 into a preset audio symbol sequence 7 and records it. The candidate voice section detection unit 2 converts the input search word into a voice symbol and searches for a matching part from the registered voice symbol sequence 7. The search word voice determination unit 3 determines whether or not the target search word is included in all the voice sections detected by the character search 10 using the original voice data waveform 4.

First, the audio data registration unit 1 shown in FIG.
Will be described. It is assumed that the target audio waveform data 4 is digitized in advance and stored in a storage device such as a hard disk. The audio waveform data 4 is converted into feature amounts such as spectrum information and power in short time units of about 5 to 10 msec by the acoustic analysis 5 and output as audio parameters. In phoneme recognition 4, the type of phoneme is specified from the time-series data of the voice parameter. Each extracted phoneme is mapped to a preset symbol group, and stored in a storage device in association with a voice section (voice symbol sequence 7). Since the conversion accuracy to the voice symbol sequence 7 becomes the detection accuracy at the time of search,
Speech symbols do not represent readings accurately,
It ignores phonemes that are easy to make mistakes and handles several similar phonemes at once.

Next, the candidate voice section detection unit 2 at the time of voice data search will be described with reference to FIG. When a search word is input by an input device such as a keyboard (search word input 8) and passed to the candidate voice section detection unit 6, first, the search word is converted into a series of symbol groups having the same specifications as those used at the time of voice data registration. Mapping (conversion to speech symbol 9). The search word represents the content of the utterance to be searched for, and is a reading symbol written in any of Hiragana / Katakana / Romaji. If you want to handle ordinary Kanji kana mixed notation when entering a search word, it is possible to prepare a separate word dictionary.

In the character string search 10, a part that matches the voice symbol sequence of the search word is extracted from the voice symbol row 7 registered in the voice data registration unit 1. If there are a plurality of matching parts, the detection positions of all of them are output. As a result, high-speed candidate sections can be narrowed down at the symbol level.

Next, a description will be given of the search word voice determination unit 3 at the time of voice data search with reference to FIG. The search word speech determination unit 3 uses the original speech waveform data 4 to determine whether or not all of the speech sections detected by the character string search 10 actually include the target search word. Verification (word speech recognition 12). In general word speech recognition, one of a plurality of registered words is specified. Here, only a search word is input, the similarity is calculated, and the similarity is determined based on a threshold. I have. This enables detection at the audio waveform data level. When it is determined by the word speech recognition 12 that a search word exists, the position is output as a search result (search result output 13).

Next, an audio data search apparatus using audio parameters will be described with reference to FIG. FIG. 2 shows another configuration diagram of the voice data search device according to the embodiment of the present invention. As shown in FIG. 2, a speech parameter 26 obtained in the processing of the acoustic analysis 25 performed by the speech data registration unit 21 is stored in association with a speech waveform, and is stored in the search word speech determination unit 23 during a search. It is also characterized in that it can be used for search word speech determination. Although the storage capacity will increase slightly, the amount of computation during searching can be reduced,
Further higher speed is possible. Since the processing procedure is the same as the above-described processing procedure, the description is omitted here.

Next, a case in which a full-text search system is introduced for voice symbol search in the character string search 9 at the time of search will be described. Candidate voice section detection unit 2 shown in FIG.
In the above, by introducing a full-text search system for the voice symbol search in the character string search 9 at the time of the search, it is possible to perform a high-speed search for narrowing down the candidate voice sections, as compared to searching the voice symbol sequence 11 sequentially from the beginning each time. Become. In particular, the effect is great when the amount of audio data is large. However, since the full-text search system uses a large-scale index file, the storage capacity required for the system increases.

Next, a case where a word spotting speech recognition technique is used for the search word speech determination performed by the search word speech determination unit 3 at the time of search will be described. Here, the word spotting is a recognition method for detecting a target word by referencing a voice pattern to a standard pattern without limiting the section and searching for a portion having a high likelihood of matching.

By using a word spotting speech recognition technique in which the start and end of the speech can be handled freely for the search word speech determination performed by the search word speech determination unit 3 in FIG. 1, both ends of the candidate speech section are slightly widened. By passing the speech waveform data to the word spotting speech recognition, erroneous recognition due to dropout of the phoneme at the beginning of the word can be reduced, and more accurate judgment can be made. In addition, since word spotting speech recognition does not need to provide speech in sections, the internal mechanism also determines whether or not the specified word exists, and there is no need to perform threshold processing again.

Next, a method for setting a voice symbol group in the voice data registration unit 1 will be described.

Here, all consonants that are likely to cause misrecognition are ignored, and only voiced vowels, long vowels, vowels, and silence speech sections are converted into speech symbol strings and used. This allows
Errors generated when narrowing down the candidate sections can be reduced, and a highly accurate detection result can be obtained. For example, if there is an utterance of “voice search ...”, the vowel sequence “On-eo-en-a-ui-wi-e ...
"Is extracted as a voice symbol sequence. In this case, at the time of the search, the search word input in the search word input 8 also needs to be symbolized in the same manner. In the search word input 8 of FIG. 1, when "onsay" is input as a search word, a consonant is deleted from its pronunciation, converted to "on-a", and a search is performed from the registered voice symbol string 11 (character string). Search 10). Here, the character string search 10 detects all words having the same vowel sequence, such as “respect”, in addition to “voice”. Finally, all the detected speech sections are checked by the word speech recognition 12, and only the target section is output as a search result (search result output 13).

Next, a case will be described in which consonants that are likely to cause confusion are treated as one syllable in the speech symbol group. Although it is based on a single syllable, the consonants that are likely to cause confusion are grouped and a symbol group treated as one syllable is used as the voice symbol sequence 7, thereby suppressing errors in narrowing down candidate sections.
The number of candidate sections can be reduced, and higher-speed search can be performed. For example, when it is difficult to distinguish between “P” and “T” of the unvoiced consonant, the syllables in the “P” line and the “T” line are treated as the same. The notation may be set freely, such as unifying the "pa" line, unifying the "ta" line, or assigning a completely different symbol. Since the processing procedure is the same as the above-described processing procedure, the description is omitted here.

Next, a case will be described in which not only a short unit such as a vowel or a single syllable but also a word is assigned to a voice symbol group as one symbol. By preliminarily registering words having a high frequency of occurrence in a dictionary and assigning each word as one voice symbol, the number of candidate sections detected at the same time as the voice symbol string 7 is reduced is reduced. High-speed search becomes possible. When registering voice data in the voice data registration unit 1 in FIG. 1, the target word portion may be extracted from the detected phoneme sequence or a separate word voice recognition device may be prepared.

The present invention is also applied to a system for retrieving / reproducing a part where a desired discussion has been made from a tape recording a conference, a system for extracting a desired scene from a recorded videotape by using sound as a key, and the like. be able to.

Although the preferred embodiment of the present invention has been described in detail, the present invention is not limited to the specific embodiment, and various modifications may be made within the scope of the present invention described in the appended claims. Can be modified and changed.

[0030]

As is apparent from the above description, the first aspect of the present invention provides an audio data registration unit for converting digitized audio waveform data into a predetermined audio symbol sequence and recording the converted data. A candidate speech section detection unit that converts a search term into a speech symbol and searches for a matching part from a registered symbol sequence, and a search term speech determination unit that determines whether the candidate section matches the search term By doing so, high-speed narrowing of candidate sections at the symbol level and accurate detection by matching at the audio waveform level can be performed.

According to a second aspect of the present invention, in the voice data registration unit, a voice parameter sequence used when extracting voice symbols from voice data is stored, and the voice parameter sequence is stored in the search word voice determination unit. By sharing, it is possible to reduce the amount of calculation at the time of retrieval, and further increase the speed.

According to the third aspect of the present invention, the candidate voice section detection unit uses a full-text search system to search for voice symbols, so that the speed of narrowing down the candidate sections can be further increased. This is particularly effective when dealing with a large amount of audio data.

According to a fourth aspect of the present invention, the search word speech determination unit uses a word spotting speech recognition device capable of freely handling the start and end of the speech for determination, thereby reducing erroneous word speech recognition. And a highly accurate detection result can be obtained.

According to a fifth aspect of the present invention, a consonant which is apt to be erroneously recognized is ignored, and only voiced vowels, long vowels, lyophotic and non-voiced voice sections are converted into voice symbol strings and used.
Errors generated when narrowing down the candidate sections can be reduced, and a highly accurate detection result can be obtained.

The invention according to claim 6 is based on a single syllable, but a consonant which is likely to cause confusion is grouped and a symbol group treated as one syllable is used as a voice symbol sequence, thereby making it possible to select a candidate section. The number of candidate sections can be reduced while suppressing an error in narrowing down, and a higher-speed search can be performed.

According to the present invention, words having a high frequency of appearance are registered in a dictionary in advance, and each word is assigned as one voice symbol. Thus, the number of candidate sections can be reduced, and higher-speed search can be performed.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a voice data search device according to an embodiment of the present invention.

FIG. 2 is another configuration diagram of the voice data search device according to the embodiment of the present invention.

[Explanation of symbols]

 1 voice data registration unit 2 candidate voice section detection unit 3 search word voice determination unit

──────────────────────────────────────────────────続 き Continued on the front page (51) Int.Cl. ⁷ Identification symbol FI Theme coat ゛ (Reference) G10L 15/00 G10L 3/00 551P 15/28 571E 5/06 F

Claims (7)

[Claims] 1. An audio data registration unit for converting digitized audio waveform data into a predetermined audio symbol sequence and recording it, and converting a search word into an audio symbol to search for a matching part from a registered symbol sequence. Voice data search device, comprising: a candidate voice section detection unit that performs a search; and a search word voice determination unit that determines whether a candidate section detected by the candidate voice section detection unit matches a search word. . 2. The voice data search device according to claim 1, wherein the voice data registration unit stores a voice parameter sequence used when extracting a voice symbol from voice data, and searches the voice parameter sequence. A voice data search device characterized by being shared by a word voice determination unit. 3. The voice data search device according to claim 1, wherein the candidate voice section detection unit searches for a voice symbol using a full-text search system. 4. The voice data search device according to claim 1, wherein the search word voice determination unit uses a word spotting voice recognition device capable of handling the start and end of the voice freely. apparatus. 5. The voice data search device according to claim 1, wherein the voice data registration unit converts only voiced vowels, long vowels, vocal sounds, and non-voiced voice sections into voice symbol strings and uses them. Voice data search device. 6. The symbol data group according to claim 1, wherein the voice data registration unit is based on a single syllable, but groups consonants that are likely to cause confusion into one syllable. A voice data search device characterized by using a character string as a voice symbol sequence. 7. The voice data search device according to claim 5, wherein the voice data registration unit registers in advance a word having a high frequency of appearance in a dictionary, and adds the word as one voice symbol. Voice data search device.