JP2008134979A - Moving picture retrieval method and moving picture retrieval system - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To prevent a retrieval keyword from being commonplaced due to the lapse of time or the like and to retrieve moving pictures by using a keyword adapted to time in keyword retrieval of many moving pictures registered in a moving picture site of the Internet or the like. <P>SOLUTION: The moving picture retrieval system comprises a moving picture registering interface for inputting reading Kana characters in the retrieval of many moving pictures registered in the moving picture site of the Internet or the like, a voice recognition part 7 for performing voice recognition for a moving picture file in a moving picture database on the basis of a phoneme string generated from the reading Kana characters, an index database 5 for adding a new vocabulary obtained from the voice recognition result, and a data processing part 4 for controlling the whole system by issuing an SQL sentence to respective databases. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a search method and a search system for moving images registered on a site.

  Due to advances in speech recognition technology, recognition accuracy of 90% or more for word speech recognition and 70% or more for continuous speech recognition can be obtained in a no-noise environment. Using such a voice recognition technique, a technique for performing voice recognition on a voice signal accompanying a video signal, detecting a vocabulary, and associating it with the video signal becomes possible. Along with this, it has become possible to extract and index keywords by performing voice recognition on moving images on a moving image site.

  In Patent Document 1, an audio signal included in a received video signal is converted into character information by voice recognition, and the character string is used as a keyword and associated with the video signal or the audio signal. The search user can extract a video signal associated with the keyword by using the keyword. However, vocabularies that can be keyword candidates remain within the range of a static vocabulary database in advance, and there is a problem that it is practically difficult to use new words, coined words, and proper nouns that change every day as keywords.

  JP 2006-54517 A

  Due to advances in speech recognition technology, recognition accuracy of 90% or more for word speech recognition and 70% or more for continuous speech recognition can be obtained in a no-noise environment. Using such a voice recognition technique, a technique for performing voice recognition on a voice signal accompanying a video signal, detecting a vocabulary, and associating it with the video signal becomes possible. However, the number of vocabulary that can be recognized by speech depends on the number of vocabulary in the large vocabulary database. If the number of vocabulary is small, the recognition accuracy improves, but most recognition vocabularies are recognized as unknown words. If there are many, the vocabulary recognized as an unknown word will decrease, but there exists a possibility that the vocabulary with a phonological distance may be misrecognized. For this reason, it is desirable to improve the quality while keeping the number of vocabularies in the large vocabulary database within a certain range, that is, to accommodate vocabularies with contents that always match the times. In a static vocabulary database, speech recognition of new words, coined words, proper nouns, etc. that change from day to day cannot be performed, and all are processed as unknown words. For example, names of politicians and talents who have suddenly improved their names are not usually stored in the vocabulary database, but since their names are suddenly improved, there are many users who want to search for these keywords. It is. However, since it is actually difficult to update the contents of the vocabulary database according to the times, there is a problem that the vocabulary that the user wants to use most as a keyword cannot be used.

  The present invention prompts a registered user to input a keyword suitable for the moving image and its reading kana when the moving image is newly registered, adds the input keyword to the large vocabulary database, and reads the kana reading of the vocabulary The present invention is characterized in that speech recognition is performed on a moving image already registered using a phoneme sequence obtained from the above, and the vocabulary is added to an index database when the phoneme sequence is detected. It is characterized by being effective on many interactive media such as sites.

  Further, the present invention is characterized by categorizing the vocabulary having a high appearance frequency in the index database and responding to the search user's category search needs.

  In addition, in order to prevent the enlargement of the large vocabulary database due to the addition of the above new vocabulary, the present invention deletes keywords that have become obsolete from the statistical information of the search keyword of the search user from the large vocabulary database. It is characterized by optimizing the contents of a large vocabulary database while always keeping the number of vocabularies within a certain range.

  According to the invention, the search user can always obtain a search result of an appropriate moving image.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. FIG. 1 is a basic configuration diagram for realizing an embodiment of the moving image search system of the present invention.

  A video file uploader and a keyword input field are installed on the video registration interface 2, and the keyword input field includes a character string input field and a reading input field.

  The URI of the registered moving image file is inserted into the moving image database 3.

  For the registered moving image file, the data processing unit 4 instructs the voice recognition unit 7 to perform keyword spotting based on the vocabulary of the large vocabulary database 6, and keyword spotting together with the moving image ID of the moving image file in the moving image database 3. The vocabulary ID on the large vocabulary database 6 for the vocabulary detected by is inserted into the index database 5. At this time, for the functional words such as particles and those that are not suitable as search keywords, the value of the importance field in the large vocabulary database 6 is set to False to avoid being inserted into the index database 5. .

  In speech recognition, the amount of vocabulary information obtained from large vocabulary continuous speech recognition is more effective than keyword spotting, but the recognition accuracy of continuous speech recognition for unspecified speakers is actually low, especially for spoken language Since the grammar is not clear, keyword spotting is adopted in the present invention. In the future, it is desirable to use large vocabulary continuous speech recognition.

  Along with the moving image ID of the moving image file in the moving image database 3, a vocabulary ID returned as a result of referring to the large vocabulary database 6 for the character string input in the character string input field is inserted into the index database 5.

  When the character string input in the character string input field does not exist in the large vocabulary database 6, the character string input in the reading kana input field by the data processing unit 4 is converted into a phoneme string form usable by the speech recognition unit 7. Then, it is inserted into the large vocabulary database 6 together with the character string input in the character string input field.

  Further, the data processing unit 4 instructs the voice recognition unit 7 to perform keyword spotting on the existing moving image file inserted into the moving image database 3 for the phoneme string.

  When the phoneme string is detected as a result of keyword spotting on the existing movie file inserted in the movie database 3, the movie ID on the movie database 3 and the vocabulary ID for the phoneme sequence on the large vocabulary database 6 are index database. 5 is inserted. Here, since the data processing unit 4 holds the vocabulary IDs that have not been subjected to keyword spotting, the homonyms are not erroneously inserted into the index database 5.

  The large vocabulary database 6 holds statistical information indicating the number of searches for each vocabulary of the search user 9, and when the number of records in the large vocabulary database 6 exceeds a certain value due to the addition of a new vocabulary, the number of searches is based on the statistical information. Delete less vocabulary. By optimizing the vocabulary, it is possible to obtain effects such as preventing the search keyword group from becoming obsolete, speeding up speech recognition processing, and preventing erroneous recognition of speech recognition.

  The data processing unit 4 categorizes the vocabulary having a high appearance frequency on the index database 5 on the moving image search interface 8, thereby performing not only the keyword search of the search user 9. Corresponds to the needs of category search.

  When the number of video files subject to speech recognition is large and the load of speech recognition processing is high, voice recognition is performed by keyword spotting for multiple new vocabularies when a certain number of new vocabularies are reached, or once every week, etc. The load of recognition processing can be reduced.

  The moving image registration interface 2 and the moving image search interface 8 exist for the convenience of the user, and when the registered user or the search user of the present invention is not assumed as a human, these interfaces are not necessarily required for realizing the system. For example, these interfaces are not necessarily required for a search robot that patrols on the WEB.

  The present invention can be applied to an industrial field that provides a moving image search service on a moving image site or the like.

  1 is a basic configuration diagram for realizing an embodiment of a moving image search system of the present invention.   It is a basic composition figure of the table of each database in the present invention.

Explanation of symbols

DESCRIPTION OF SYMBOLS 1 Registered user 2 Movie registration interface 3 Movie database 4 Data processing part 5 Index database 6 Large vocabulary database 7 Voice recognition part 8 Movie search interface 9 Search user

Claims (3)

  A keyword entered by a registered user of a video on a video site is added to a large vocabulary database, and speech recognition is performed on a video already registered using a phoneme string obtained from the reading of the vocabulary. When a column is detected, by adding the vocabulary to the index database, the keyword group that can be used for the search is always kept in accordance with the age and purpose, and when the video search user performs a search, a new word, A moving picture search method and a moving picture search system equipped with a dynamic large vocabulary database characterized by avoiding, as much as possible, a situation in which an appropriate search result cannot be obtained due to unknown words such as coined words and proper nouns.   The vocabulary with high appearance frequency is categorized on the index database of claim 1 to categorize the vocabulary and respond to not only the keyword search shown in claim 1 but also the category search needs. Method and video search system.   In order to prevent the enlargement of the large vocabulary database due to the addition of a new vocabulary, the large vocabulary is deleted by deleting from the large vocabulary database keywords that have become obsolete from the statistical information of the search keywords of the search user. A moving picture search method and a moving picture search system characterized by optimizing the contents of a database while always keeping the number of vocabularies within a certain range.

Images