JP2010175708A - System and method for retrieval of speech recognition - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a speech recognition retrieval system for suppressing the number of vocabularies in a range that guarantees recognition accuracy for practically use and improving the covering ratio for utterance. <P>SOLUTION: The speech recognition retrieval system includes an EPG DB 29 for accumulating data of program information on a broadcast program, a speech input section 11, a speech recognition section 21 for recognizing input speech and converting it into a text, a speech recognition dictionary 22 referred to by the speech recognition section 21, a retrieval section 28 for retrieving a range to be retrieved in the EPG DB 29 using the text recognized by the speech recognition section 21 as a retrieval keyword, a display section 27 for displaying the speech recognition result, the retrieval result or the like, an operation section 12 for switching selection and display of a speech recognition candidate to be displayed, an utterance habit extracting section 23 for extracting user's utterance habit from the text recognized by the speech recognition section 21, an utterance habit DB 24 for accumulating the utterance habit as data, and a speech recognition dictionary creating section 25 for creating the speech recognition dictionary 22 based on the data of the range to be retrieved and the contents of the utterance habit DB 24. <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to a voice recognition search system and a voice recognition search method for searching for desired information from a target search range by voice recognition input.

  Efforts are being made to search for and operate desired information by voice recognition input in situations where manual operation is not possible, such as car navigation systems. In the case of isolated word speech recognition in which the beginning and end of speech are detected only once during a certain period of time, the number of vocabularies and the recognition rate are in a trade-off relationship. Therefore, a method of ensuring the voice recognition accuracy by appropriately switching the voice recognition dictionary according to the input voice attribute (for example, genre, etc.) can be considered. For example, the input attribute is designated first and the appropriate voice recognition is performed. A method of inputting voice after selecting a dictionary has been proposed (for example, see Patent Document 1).

  In addition, when speech recognition is performed for all vocabularies and there are many speech search key candidates, a search key confirmation related question is presented and related information is uttered, and speech is obtained from the search key recognition likelihood and the related information recognition likelihood. A method for identifying search key candidates has been proposed (see, for example, Patent Document 2).

  However, in applications where manual operation is possible (for example, TV program (contents) reservation), when using voice recognition input to reduce the burden of operating the remote control device, etc., all input is performed by voice recognition input. Rather, it is considered that the overall usability is improved by appropriately combining with key operations. From such a viewpoint, there has been proposed an approach for making a program reservation by voice recognition using an EPG (Electronic Program Guide) (see, for example, Patent Document 3).

JP 2007-264198 A Japanese Patent No. 3420965 JP 2000-316128 A

  When voice recognition input is used for an application that can be operated by hand, a dictionary for voice recognition prepared in advance is fixedly used in the related art. However, in searching for information that changes from day to day, such as program information and information on the Internet, it is necessary to maintain the voice recognition accuracy to update the dictionary for voice recognition according to the contents of the search target range. In addition, the relationship between the number of words in the dictionary for speech recognition and the recognition accuracy shows that the recognition accuracy decreases when the number of vocabularies is increased in order to increase the coverage rate for utterances, while the recognition accuracy improves when the number of vocabularies is reduced. The rate drops.

  Therefore, the present invention has been made in view of the above-described problems, and the speech recognition search system and the speech recognition search that improve the coverage rate for utterances within the range that guarantees the recognition accuracy that can withstand the number of vocabulary in practical use. It aims to provide a method.

  According to one aspect of the present invention, a content information database that accumulates data of content information, a voice input unit for inputting voice, and voice recognition that recognizes voice input from the voice input unit and converts it into text. A speech recognition dictionary that is referred to by the speech recognition unit for recognition processing, a search unit that searches a search target range in the content information database using text recognized by the speech recognition unit as a search keyword, and a search result A display unit that displays a voice recognition candidate displayed on the display unit, an utterance tendency extraction unit that extracts a user's utterance tendency from text recognized by the voice recognition unit, and the utterance Based on the utterance tendency database that accumulates the utterance tendency extracted by the trend extraction unit, the data of the search target range, and the contents of the utterance tendency database Speech recognition retrieval system characterized by comprising a speech recognition dictionary generating unit which generates the speech recognition dictionary is provided.

  According to another aspect of the present invention, there is provided a speech recognition search method in a speech recognition search system for recognizing a speech search key to be searched that is input from a user using speech, wherein the search keyword is spoken. Input, perform speech recognition on the input speech signal, convert it to text, present a plurality of speech recognition candidates in descending order of likelihood, and use the character string selected from the presented candidates as content information Search the database, select the content from the content candidate list, and when the content is selected, extract the utterance tendency, compare the official name of the selected content with the uttered content, and extract the features of the utterance The speech recognition retrieval method is characterized in that it is stored in an utterance tendency database.

  According to the present invention, the coverage rate for utterances can be improved even within the range that guarantees the recognition accuracy that can withstand practical use of the vocabulary number.

It is a block diagram which shows an example of the speech recognition search system which concerns on embodiment of this invention. It is the schematic which shows the example of mounting of the remote control apparatus of the speech recognition search system which concerns on embodiment of this invention. It is a block diagram which shows an example which mounts the radio | wireless communication function in the speech recognition search system which concerns on embodiment of this invention. It is a flowchart which shows the processing operation of the speech recognition search system which concerns on embodiment of this invention. It is a figure explaining the example of presentation of a speech recognition candidate. It is a figure explaining the example of a display of a search result. It is a figure explaining the example of giving of reading of a program name. It is a figure explaining the example of EPG data. It is a flowchart which shows the processing operation at the time of a speech recognition dictionary production | generation part producing | generating a speech recognition dictionary.

  Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the drawings, the same portions are denoted by the same reference numerals, and redundant description is omitted.

  FIG. 1 is a block diagram illustrating a configuration example of a speech recognition / retrieval system 100 according to an embodiment of the present invention, and FIG. 2 is a schematic diagram illustrating an implementation example of a remote control device 10.

  As shown in FIG. 1, the speech recognition / retrieval system 100 includes a remote control device 10 and a recording device 20. As shown in FIG. 2, the remote control device 10 includes a voice input unit 11 and an operation unit 12. The voice input unit 11 can be provided internally or externally at an arbitrary position of the remote control device 10, and is, for example, a microphone. FIG. 2 shows an example in which the voice input unit 11 is incorporated. The operation unit 12 is a part that receives an operation of the operator. In the example illustrated in FIG. 2, two cross keys and two push buttons for moving a recognized voice candidate are provided at an arbitrary position of the remote control device 10. The configuration of the operation unit 12 is not limited to this, and the operation unit 12 may be configured so that the pointer can be operated with a pointing device.

  The recording device 20 includes a voice recognition unit 21, a voice recognition dictionary 22, an utterance tendency extraction unit 23, an utterance tendency database (hereinafter referred to as an utterance tendency DB) 24, a voice recognition dictionary generation unit 25, a voice recognition candidate display unit 26, and a display unit. 27, a search unit 28, an EPG database (hereinafter referred to as EPG DB) 29, and a recording / control unit (not shown). The recording / control unit is a device having a recording function such as a video hard disk recorder, a TV with a recording function, or a personal computer. When the recording / control unit is composed of a personal computer with a recording function, the voice input unit 11 may be connected to the personal computer, and the operation unit 12 may use a personal computer operating device such as a mouse.

  The voice recognition unit 21 recognizes the voice acquired through the voice input unit 11 of the remote control device 10. The speech recognition unit 21 is equipped with a speech recognition engine including morphological analysis, for example.

  The voice recognition dictionary 22 stores a large amount of data for voice recognition, and accepts access from the voice recognition unit 21. The voice recognition dictionary 22 can be constructed by, for example, an HDD or an EEPROM.

  The utterance tendency extraction unit 23 extracts the utterance tendency (described later) of the utterer captured via the voice input unit 11. The utterance tendency extraction unit 23 can be constructed by software, for example.

  The utterance tendency DB 24 accumulates utterance tendency data extracted by the utterance tendency extraction unit 23. The utterance tendency DB 24 can be constructed with, for example, a volatile memory.

  The voice recognition dictionary generation unit 25 takes in data from the utterance tendency DB 24 and the EPG DB 29 and generates a voice recognition dictionary. The speech recognition dictionary generated by the speech recognition dictionary generation unit 25 is sent to the speech recognition dictionary 22.

  The speech recognition candidate display unit 26 lists speech recognition candidates according to likelihood as a result of speech recognition by the speech recognition unit 21. The speech recognition candidate display unit 26 can be constructed with, for example, a volatile memory.

  The display unit 27 manages various displays in voice recognition and search operations. The display unit 27 is preferably a display such as an LCD.

  The search unit 28 searches the EPG DB 29 using a character string selected from the speech recognition candidates as a search keyword. The search unit 28 can be constructed by software, for example.

  The EPG DB 29 accumulates EPG (Electric Program Guide) data, and can be constructed with, for example, an HDD or an EEPROM. The EPG data is a system for providing a TV program guide as digital data and displaying it on a screen or making a recording reservation easily. EPG data can be provided by "ADAMS-EPG" on the terrestrial wave, ARIB (Association of Radio Industries and Businesses of Japan) standard "EPG" on BS digital, "ADAMS" provided via the Internet. -EPG + "and" iEPG ". FIG. 8 shows an example of EPG data. As shown in FIG. 8, <TITLE> includes a program name, and <TEXT> includes information about performers on an XML (Extensible Markup Language) basis.

  As shown in FIG. 1, the remote control device 10 and the recording device 20 can be connected by wire, but the present invention is not limited to this. As shown in FIG. 3, the remote control device 10 may include a communication unit 13 and the recording device 20 may include a communication unit 30 so that wireless communication is possible. In addition, since it can implement with the same structure except the structure which concerns on radio | wireless communication, hereafter, it demonstrates according to the structure shown in FIG.

  Next, speech recognition and search processing operations in the speech recognition and retrieval system configured as described above will be described with reference to the flowchart of FIG.

  First, the voice recognition search system waits for a voice recognition start instruction from the user (step S10). For the voice recognition start instruction, a button assigned to the voice recognition start instruction function in the operation unit 12 of the remote control device 10 is pressed. Alternatively, a button on the display arranged on the display unit 27 of the recording device 20 may be pressed using the operation unit 12. The instruction waits until there is a voice recognition start instruction (step S11). After the voice recognition start instruction is received, the user inputs a desired search keyword into the voice input unit 11 (step S12). Also, a voice recognition start instruction is given, and the voice recognition unit 21 detects a silent section after voice input and automatically ends voice recognition (step S13). As for when to end speech recognition after starting speech recognition, speech recognition may be performed while the button is pressed.

  Next, the speech recognition unit 21 performs speech recognition on the speech signal input to the speech input unit 11 using the speech recognition dictionary 22, and presents a plurality of speech recognition candidates in descending order of likelihood (step). S14).

  FIG. 5 is a diagram illustrating an example of presenting speech recognition candidates. In the example shown in FIG. 5, three candidates are listed as speech recognition candidates as a result of speech recognition for the input speech “Ebisu-Kokoukouza”. This candidate is presented on the speech recognition candidate display unit 26. Note that it is easy for the user to list the speech recognition candidates as narrowed down as possible. Therefore, for example, it is possible to set so that candidates with a voice recognition score of 80 or more are listed. The part enclosed in parentheses is “reading” of each candidate, and the part before the parenthesis is each character string. In this way, if both candidates and their readings are displayed, the user can confirm why the candidates are listed and it is easy to understand. The number and order of the voice recognition candidates presented depends solely on the recognition performance of the voice recognition engine.

  When there is a desired character string, for example, a program name (partial word) or a performer name among the presented candidates, the user operates the operation unit 12 to select a desired candidate and there is no desired candidate. In this case, voice recognition input is performed again (step S15).

  When a desired character string is selected, the search unit 28 searches the EPG DB 29 using the character string as a first search keyword (step S16). If the character string is the name of a performer, the <TEXT> tag after <ITEM> performer </ ITEM> is searched from the EPG data whose example is shown in FIG. If the character string is a program name (or a part thereof), a <TITLE> tag is searched. As a result of the search, program broadcast date / time, channel, program name, etc. are extracted from the EPG data of the hit program and listed on the display unit 27 (step S17).

  A display example of the search result is shown in FIG. In this example, in response to the user selecting “ABC High School Course” from the character strings presented in the speech recognition candidate list shown in FIG. 5, the EPG DB 29 is searched and corresponds to “ABC High School Course”. A list of 10 programs is displayed as a list of program candidates.

  As described above, the voice recognition candidates shown in FIG. 5 can be set to list, for example, candidates having a voice recognition score of 80 or more, but the likelihood of the first candidate is the second candidate. If it is clearly higher than the following candidates, the program candidates shown in FIG. 6 may be listed without performing a direct search and presenting the list of voice recognition candidates shown in FIG. .

  As a result of the search, a desired program in the program candidate list shown in FIG. 6 is selected (step S18). When a program is selected, details of the EPG data as shown in FIG. 8 can be referred to (step S19).

  Furthermore, after selecting a desired program, the utterance tendency is also extracted. As will be described later, in reading the program name, not only the official program name but also a partial program name divided based on a predetermined rule is generated when the voice recognition dictionary 22 is generated. As the predetermined rule, for example, in FIG. 7, the program name can be divided before and after the parenthesis of the program name, divided depending on the presence or absence of a space, the inside of the parenthesis can be divided, and the division before and after “no”. However, it is not appropriate to divide into dark clouds because the search keywords increase more than necessary.

  The utterance tendency extraction unit 23 compares the official name of the selected program with the content actually uttered by the user, and extracts the characteristics of the utterance by, for example, one of the following two examples or a combination of both ( Step S20), and stored in the utterance tendency DB 24 (Step S21).

<Speech feature extraction based on utterance length>
FIG. 7 is a diagram for explaining an example of giving program name reading. In the example of FIG. 7, the official program name is read as “Ebisu-Kikokuri-zukuri-zukuri-buri-no-bushi-buratsu-buri” (30 characters). On the other hand, the length of the reading actually spoken is evaluated. 19 characters are used for speaking "Ebisu-Kikuzakusoka", 11 characters are used for speaking "Rakubata no Utsuri", and 3 characters are used for "Butsuri". The lengths of these actually uttered readings are normalized by the reading length of the official program name, and the values 19/30, 11/30, and 3/30 are stored in the utterance tendency DB 24, respectively. Or, for the length of reading of the official program name, several threshold values are set in advance such as less than 1/3, 1/3 or more, less than 2/3, 2/3 or more, and the frequency of each range is set. You may make it count.

<Speech feature extraction based on utterance location>
In the example of FIG. 7, the official program name is read as “Ebisu-Kikokuri-zukuri-zukuri-buri-no-bushi-buratsu-buri” (30 characters). On the other hand, the position of the actually spoken reading, that is, which part of the official program name is read is evaluated. When uttering from the beginning of the official program name such as `` Ebisu Kokoukouza '', when uttering not from the beginning to the end of the official program name such as `` first half '', `` Rikasogo '' When the utterance is made to the end of the official program name, such as “intermediate” and “hit”, the latter half is stored and stored in the utterance tendency DB 24.

  Next, a method for generating the speech recognition dictionary 22 will be described. EPG data distributed by terrestrial digital broadcasting or the like is transmitted together with a broadcast signal and stored in the EPG DB 29 by a well-known method. The voice recognition dictionary generation unit 25 analyzes the EPG data in the EPG DB 29 at a frequency of once a day, for example, and generates the voice recognition dictionary 22 according to the content of the utterance tendency data in the utterance tendency DB 24. Hereinafter, an example of the generation method will be described with reference to the flowchart of FIG. FIG. 9 is a flowchart showing a processing operation when the voice recognition dictionary generation unit 25 generates the voice recognition dictionary 22.

  FIG. 8 shows an example of EPG data for one program. In this example, it is understood that the data is in the XML format, but data other than the XML format such as iEPG may be used. These EPG data are stored in the EPG DB 29. The EPG DB 29 is preferably constructed by an XML database in the case of XML format data, but may be constructed by another database (RDB or the like). From the EPG data stored in the EPG DB 29, the program name enclosed in the <TITLE> tag and the performer name enclosed in the <TEXT> tag next to the <ITEM> performer </ ITEM> are extracted (step) S30).

  Next, there are some program titles that are fairly long and include subtitles as they are, so for example, using the space, parentheses, and particles (for example, “no”) extracted by morphological analysis included in the program name. Split a string. At this time, the past utterance tendency stored in the utterance tendency DB 24 is read out and reflected, the program name is divided, and the reading for generating and registering the voice recognition dictionary 22 is changed (steps S31 and S32). For example, if the utterance length tends to be short, a short reading from the official program name is preferentially registered in the dictionary. Conversely, if the utterance length tends to be long, the official program Preferentially register long readings from names. Or, if there is a tendency to utter the first half of the official program name, the reading of the first half of the official program name should be preferentially registered in the dictionary, and conversely, the middle or the second half of the official program name will be uttered a lot. If there is a tendency, the middle or latter half of the official program name is preferentially registered in the dictionary.

  By doing so, a plurality of readings are given to the same program name in order to improve the coverage rate for the utterance. If the number of partial words is increased in the dark clouds, the number of vocabularies stored in the speech recognition dictionary 22 will increase, but the speech that is likely to be spoken according to the past user's utterance tendency is preferentially spoken. By storing it in the recognition dictionary 22, it is possible to improve the coverage rate while ensuring the voice recognition accuracy.

  A general-purpose kana-kanji conversion dictionary may be used for giving a reading, or another reading such as a nickname may be generated based on information on the Internet, and any means is available (step S33). . In addition, it is convenient for searching if an identifier is prepared for whether each reading relates to a program name or a performer name.

  It is preferable that the update processing of the voice recognition dictionary 22 is performed periodically, for example, once a day or late at night.

  According to this embodiment, when searching for information that changes daily such as program information and information on the Internet by voice recognition input, a voice recognition dictionary suitable for the search target range is dynamically generated and the user's utterance Since the generated vocabulary can be appropriately changed according to the tendency, the coverage rate for the utterance can be improved even within the range in which the number of vocabularies is assured for the recognition accuracy to be practically used.

  Note that the present invention is not limited to the above-described embodiments as they are, and can be embodied by modifying the constituent elements without departing from the scope of the invention in the implementation stage. Moreover, various inventions can be formed by appropriately combining a plurality of constituent elements disclosed in the above embodiments. For example, some components may be deleted from all the components shown in the embodiment. Furthermore, constituent elements over different embodiments may be appropriately combined.

  DESCRIPTION OF SYMBOLS 100 ... Voice recognition search system, 10 ... Remote control device, 11 ... Voice input part, 12 ... Operation part, 20 ... Recording apparatus, 21 ... Voice recognition part, 22 ... Speech recognition dictionary, 23 ... utterance tendency extraction unit, 24 ... utterance tendency DB, 25 ... speech recognition dictionary generation unit, 26 ... speech recognition candidate display unit, 27 ... display unit, ..Search part, 29 ... EPG DB.

Claims (7)

A content information database for storing content information data;
A voice input unit for inputting voice;
A voice recognition unit that recognizes the voice input from the voice input unit and converts it into text;
A speech recognition dictionary referred to by the speech recognition unit for recognition processing;
A search unit for searching a search target range in the content information database using the text recognized by the voice recognition unit as a search keyword;
A display for displaying the search results;
An operation unit for selecting a voice recognition candidate displayed on the display unit;
An utterance tendency extraction unit that extracts a user's utterance tendency from the text recognized by the voice recognition unit;
An utterance tendency database for accumulating the utterance tendency extracted by the utterance tendency extraction unit;
A speech recognition search system, comprising: a speech recognition dictionary generation unit that generates the speech recognition dictionary based on the data of the search target range and the contents of the utterance tendency database. The utterance tendency extraction unit comprises utterance length extraction means for extracting the utterance length represented by the number of characters of the recognized text,
While storing the utterance length in the utterance tendency database,
The speech recognition search system according to claim 1, wherein the speech recognition dictionary generation unit changes a speech recognition dictionary that is generated based on a tendency of the utterance length. The utterance tendency extracting unit includes a partial word position extracting unit that extracts a partial word position feature with respect to a formal content name among a plurality of partial words with respect to the same keyword,
In the utterance tendency database, the characteristics of the partial word positions are accumulated,
The speech recognition search system according to claim 1, wherein the speech recognition dictionary generation unit changes a speech recognition dictionary that is generated based on a tendency of the partial word position.   The speech recognition search system according to claim 2, wherein the tendency of the utterance length is determined by normalizing the utterance length with a reading length of a formal content name.   The feature of the partial word position is characterized in that it is determined by uttering the formal content name from the beginning, uttering the formal content name halfway, or uttering the formal content name to the end. The speech recognition search system according to claim 3.   The speech recognition search system according to claim 1, wherein the speech recognition result or the search result is displayed on the display unit in a descending order of likelihood. A speech recognition search method in a speech recognition search system that recognizes a speech search key to be searched that is input from a user using speech,
Spoken search terms,
Speech recognition is performed on the input speech signal to convert it into text, and multiple speech recognition candidates are presented in descending order of likelihood,
Search the content information database using the character string selected from the presented candidates as a search keyword,
Select content from the content candidate list,
When selecting content, extract the utterance tendency,
Compare the official name of the selected content with the content of the utterance, extract the features of the utterance, and store in the utterance tendency database.
A speech recognition search method characterized by the above.