JP2012022251A - Speech retrieval interface device and speech input retrieval method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech retrieval interface device capable of efficiently narrowing a retrieval result, when an user retrieves specific information from a database by using speech recognition.SOLUTION: Retrieval means 105 retrieves a word or a word string of a speech recognition result from a database for retrieval 104 and outputs the retrieval result and the number of retrieval candidates. When correction target word selection means 109 selects a correction target word, correction candidate generation means 111 matches the correction target word to a word of read and syllable storage means 110 for generating a correction candidate in a word unit and acquires the number of retrieval candidates to the correction candidate. Correction candidate display means 112 displays the correction candidate and the number of retrieval candidates obtained by the correction candidate generation means 111.

Description

  The present invention relates to a voice search interface device and a voice input search method for searching a specific database using voice recognition, and in particular, a function that allows a user to easily narrow down a search result and a user to correct the voice recognition result. The present invention relates to a voice search interface apparatus and a voice input search method having a function of correcting a voice recognition result easily and quickly when desired.

With the spread of the Internet and the diversification and complexity of functions of electronic devices, there are increasing opportunities to search for specific information from a large database by inputting a search keyword. If you are using a computer, it is common to use a keyboard to enter search keywords. However, when you use a keyboard that cannot be used or when you enter a search keyword with a large number of characters, you can use voice recognition to search by voice. A technique for inputting a keyword is disclosed (for example, see Patent Document 1).
However, in speech recognition, there is an essential problem that the recognition rate varies depending on the usage environment and individual differences of users, and erroneous recognition occurs. As a result, there may be a case where a problem that the user does not intend is output as the database search result. In view of this, a speech recognition apparatus such as that disclosed in Patent Document 1 discloses a technique for selecting a word of a speech recognition result and narrowing down a database search result.

  On the other hand, as a method of increasing the speech recognition rate and suppressing misrecognition, there are a method of tuning a recognition parameter by a user and a method of narrowing a recognition vocabulary depending on a scene, but it is assumed that an unspecified number of users use it. Furthermore, when a large-scale vocabulary must be targeted, such as a car navigation facility name search, the above solution is not an essential solution. Therefore, it is extremely important to provide an interface that can easily and quickly correct a recognition result when a misrecognition occurs in speech recognition. So far, many interfaces for correcting the speech recognition result have been proposed.

  For example, in Patent Literature 2, a list of words that are candidates for correction is displayed together with the recognition result by voice recognition, and the voice recognition that can be corrected simply by the user selecting a desired word from the list of words. An apparatus is disclosed. The technique described in Patent Literature 2 uses a confusion network to divide a word graph based on speech input into a plurality of word segments by acoustic clustering, and a word having a high contention probability for each word segment. Is generated as a correction candidate.

JP 2006-195576 A JP 2006-146008 A

  However, when the user selects a word of the speech recognition result and narrows down the search result, the user cannot grasp which word can be narrowed down efficiently so that the user cannot select the word. The trial and error of selecting and confirming the search result and selecting more words if necessary is repeated. Similarly, when the user wants to correct the speech recognition result, the user cannot grasp how much the search result can be narrowed down as a result of the correction, so the user corrects the word and confirms the search result, If necessary, the trial and error of correcting the word further is repeated.

  The present invention has been made to solve the above-described problems. When a user narrows down specific information from a database using voice recognition, the user can efficiently narrow down a search result. An object of the present invention is to obtain an interface device and a voice input retrieval method.

  The speech search interface device according to the present invention includes a word output means for outputting a word or a word string as a recognition result for speech input, and, when an arbitrary word or word string is given, searches a specific database to search for a word or word Search means for outputting a search result and the number of search candidates for a column, a word dictionary storage means for registering word information, a word output by the word output means, and a word registered in the word dictionary storage means Correction candidate generation means for generating correction candidates in units of words, and correction candidate search candidate number acquisition means for acquiring the number of search candidates for each correction candidate generated by the correction candidate generation means via the search means, The correction candidate generated by the correction candidate generation means and the number of search candidates corresponding to each correction candidate acquired by the correction candidate search candidate number acquisition means are output. Is obtained by a correction candidate output means that.

  The voice search interface device according to the present invention generates correction candidates in units of words for the voice recognition result and outputs the number of search candidates for each correction candidate, so that the user uses voice recognition. When narrowing down specific information from the database, the search results can be narrowed down efficiently.

It is a block diagram which shows the voice search interface apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the search result in the voice search interface apparatus of Embodiment 1 of this invention. It is explanatory drawing at the time of narrowing down a search result with one word in the voice search interface apparatus of Embodiment 1 of this invention. It is explanatory drawing at the time of narrowing down a search result by the some word in the voice search interface apparatus of Embodiment 1 of this invention. It is a flowchart which shows the correction candidate production | generation process in the speech search interface apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the information of the reading and a syllable memory | storage means in the speech search interface apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the display of the correction candidate in the speech search interface apparatus of Embodiment 1 of this invention.

Embodiment 1 FIG.
FIG. 1 shows a block diagram of a voice search interface device according to Embodiment 1 of the present invention, which will be described below.
The voice input unit 101 includes a voice input device such as a microphone and an AD converter. When a user inputs voice, the voice input unit 101 converts an analog voice signal into a digital voice signal that can be processed by a computer. The speech recognition dictionary storage unit 102 is a storage device that stores a recognition dictionary (language model) necessary for speech recognition. The voice recognition means (word output means) 103 recognizes a voice by referring to the voice recognition dictionary storage means 102 with the digital voice signal as an input, and outputs one or a plurality of word strings as a voice recognition result. The search database (specific database) 104 is a storage device in which the name of a facility or person to be searched is stored. The search unit 105 acquires the search result and the number of search result candidates including any one of the word strings of the speech recognition result output from the speech recognition unit 103 from the search database 104. At the same time, the number of search result candidates including only each word of the speech recognition result is acquired.

  The speech recognition result display means 106 simultaneously presents to the user the word sequence of the speech recognition result and the number of search result candidates corresponding to each word output from the search means 105 using a display device such as an LCD display. To do. The narrowed-down word selection unit 107 accepts an operation of selecting a word to be corrected using an input device such as a mouse or a touch panel when the user wants to narrow down the search result by selecting one of the words of the speech recognition result. When the selection operation is performed by the user, the selected word is output. Hereinafter, the word selected for the narrowing is referred to as a narrowed word. The search result narrowing means 108 acquires the search result including only the narrowed word from the search results acquired by the search means 105 and the number of search result candidates when the user selects the narrow word in the narrowed word selecting means 107. . The correction target word selection unit 109 receives an operation of selecting a word to be corrected using an input device such as a mouse or a touch panel when the user wants to correct any word in the speech recognition result, and the user performs a selection operation. When is done, the correction target word is output.

  The reading / syllable storage means (word dictionary storage means) 110 is a storage device that stores notation of words to be recognized, reading information, and syllable information. The correction candidate generation unit 111 generates and outputs a word having a high similarity as a correction candidate with respect to the correction target word output by the correction target word selection unit 109. At this time, correction candidates are generated using word-by-word reading information and syllable information stored in the reading / syllable storage means 110. At the same time, when each correction candidate is selected by the user, the search result candidate number indicating how many search results are narrowed down is acquired and output from the search means 105 (correction candidate search candidate number acquisition means and correction candidates). Function as output means). The correction candidate display means 112 displays the correction candidates and the number of search result candidates output from the correction candidate generation means 111 to the user at the same time using a display device such as an LCD display. The correction candidate selection unit 113 receives an operation of selecting a correction candidate intended by the user using an input device such as a mouse or a touch panel, and outputs the selected correction candidate when the selection operation is performed by the user. The correction execution unit 114 receives the correction candidate selected by the user as an input, updates the recognition result already presented to the user, and re-presents the correction result to the user.

Hereinafter, the processing flow of the voice search interface device configured as described above will be described with a specific example.
It is assumed that the user utters “Mitsubishi Denki Kabushiki Geisha” while trying to input “Mitsubishi Electric Corporation” by voice.
At this time, the voice input unit 101 first converts the uttered analog voice signal into a digital voice signal. Next, the voice recognition unit 103 recognizes the voice by referring to the voice recognition dictionary storage unit 102 using the converted digital voice signal as an input, and outputs a word string of the voice recognition result. The method for recognizing speech is arbitrary, and a known speech recognition method as described in Non-Patent Documents 2, 3, and 4 below can be used. For example, a method of searching for recognition candidates based on an acoustic score for a basic unit of speech recognition, such as phonemes, and a language score based on a language model, can be considered.

Non-Patent Document 2: Kiyohiro Shikano, Katsunobu Ito, Tatsuya Kawahara, Kazuya Takeda, Mikio Yamamoto: “Speech Recognition System” Ohm Co., Ltd., May 15, 2001 Non-Patent Document 3: Kenji Kita, Junichi Sakurai: "Probabilistic language model", The University of Tokyo Press, November 25, 1999 Non-patent document 4; Seiichi Nakagawa: "Speech recognition by probabilistic model", The Institute of Electronics, Information and Communication Engineers, July 1 1988 Day

When the speech recognition unit 103 outputs a word sequence of the speech recognition result (word output step), the search unit 105 searches the search database 104 for a search result including any one of the speech recognition result words, and The number of search result candidates is acquired (OR search). For example, for the above user utterance, if the speech recognition result is “Mitsubishi / Electricity / Stock / Company” (/ represents a word break), the four words “Mitsubishi” “Electricity” “Stock” A search result including any one of “company” and the number of search result candidates are acquired (search step). For example, “Mitsubishi Estate”, “Okuma Electric Co., Ltd.”, “Stock Stock Exchange”, etc. are search results. Further, for each word w _i (i = 1, 2,..., N; n is the number of words in the search result) of the speech recognition result, the search unit 105 calculates the search result including the word and the number of search result candidates n _i . get. In the above example, if there are 52 words including “Mitsubishi” in the search database 104, n ₁ = 52 when w ₁ = “Mitsubishi”.

Next, as shown in FIG. 2, the voice recognition result display means 106 presents the user with a display device such as an LCD display in a layout that allows the user to know the segment of the words of the voice recognition result. At the same time, acquired by the search unit 105, displayed adjacent to search result candidate number n _i corresponding to each word in the representation of words. Although the number of search result candidates corresponding to the word of the speech recognition result is displayed here, the size and color of each word are changed depending on the number of candidates instead of displaying the number of candidates themselves. For example, the display may indicate the number of candidates.

  When the voice recognition result and the number of search result candidates are displayed by the voice recognition result display unit 106, the user can narrow down the search result by selecting a word of the displayed voice recognition result. At this time, the narrowed word selection means 107 accepts an operation for the user to select the narrowed word using an input device such as a mouse or a touch panel. When the user selects a narrowed word, the narrowed word selection unit 107 detects the selection operation and outputs the selected narrowed word.

  Next, the search result narrowing means 108 narrows down the search results to only those containing only the words output by the narrowed word selection means 107, and displays the narrowed search results to the user as shown in FIG. At this time, it is desirable to display the word selected as the narrowed-down word so that the user can see it, for example, by changing the background color. For example, when “Mitsubishi” is selected by the user in the above example, search results including any of “Mitsubishi”, “Electricity”, “Stock”, and “Company” are narrowed down to search results including only “Mitsubishi”. It is possible to select not only one narrowed word but also a plurality of selected words. When a plurality of selected words are selected, the search results are narrowed down to search results including all of the selected narrowed words (AND search). For example, when “Mitsubishi” and “Electric” are selected as narrowing words, the search results including both “Mitsubishi” and “Electric” are narrowed down as shown in FIG.

  Further, the search result narrowing means 108 calculates the number of search results to be narrowed down when a word other than the narrowed word is selected next, and updates the number of search result candidates for each word as shown in FIG. And present it to the user. For example, when only “Mitsubishi” is selected as a refinement word, when “Electric” is selected next, and when the search result is narrowed down to 15 by AND search of “Mitsubishi” and “Electric”, “Electric” The number of search result candidates 15 is displayed adjacent to the display part. The same calculation is performed for the other words “stock” and “company”. In this way, since it is possible to grasp at a glance how many search results are narrowed down by the user's next operation, it is possible to efficiently narrow down the search results.

Here, when the user wants to correct the word of the speech recognition result, the correction target word selection unit 109 accepts an operation for the user to select the correction target word using an input device such as a mouse or a touch panel.
When the user selects a correction target word, the correction target word selection unit 109 senses the selection operation and outputs the selected correction target word. Here, the narrow-down word selection operation and the correction target word selection operation must be operations that can be distinguished from each other. For example, when using a mouse as an input device, select a narrowed word by left-click, select a correction target word by right-click, and when using a touch panel, select a narrow-down word by single tap, select a correction target word by double tap, etc. To assign another operation.

  For example, when a user speaks “Mitsubishi Denki Kabushiki Gaisha” to input “Mitsubishi Electric Corporation” by voice, the voice recognition result is “Mitsui / Electricity / Stock / Company”. At this time, the user selects “Mitsui” as a correction target word to correct “Mitsui” in the speech recognition result to “Mitsubishi”. Next, when a correction target word is selected by the correction target word selection unit 109, a flow of processing (a correction candidate generation step, a correction candidate search candidate number acquisition step, a correction candidate output step) performed by the correction candidate generation unit 111 is performed. This will be described in detail with reference to the flowchart of FIG.

  The correction candidate generation unit 111 receives the correction target word output from the correction target word selection unit 109, and first reads the correction target word from the reading information and syllable information stored in the reading / syllable storage unit 110. Syllable information is acquired (step ST101). Here, in the reading / syllable storage means 110, a notation of a word obtained by dividing the recognition target phrase by a word dividing method such as morphological analysis in the form as shown in FIG. 6 is stored. It is desirable to store the reading and syllable information. The correction candidate generation unit 111 searches the reading / syllable storage unit 110 for a correction target word and acquires corresponding reading and syllable information. For example, when the correction target word is “Mitsui”, “Mitsui” is searched from the notation of the reading / syllable storage means 110, and the reading “Mitsui” and the syllable “mi-cu-i” are acquired.

Next, an arbitrary word is selected from the reading / syllable storage means 110 (step ST102), and the similarity between the reading of the correction target word and the reading of the word selected in step ST102 is calculated (step ST103). Any known calculation method can be used as the similarity calculation method using word reading. For example, an edit distance (Levenstein distance) that defines the minimum number of operation procedures (insertion, deletion, replacement) when editing a word as another word as the distance between words is used. For example, the procedure for editing “Mitsui” to “Mitsubishi” is as follows:
"Mitsui"
"Mitsubi" (Replace I with Bi)
"Mitsubishi" (Insert)
Therefore, at least two operation procedures are required. Therefore, the edit distance between the words “Mitsui” and “Mitsubishi” is 2. Since the similarity of reading may be larger as the editing distance is smaller, the reciprocal of the editing distance can be calculated as the similarity between words. Hereinafter, the similarity using the reading information is referred to as reading similarity.

Next, the correction candidate generating unit 111 calculates the similarity between the syllable of the correction target word and the syllable of the word selected in ST102 (step ST104). A known calculation method can be used as the similarity calculation method using the syllables of words. For example, as described in Non-Patent Document 5 below, the confusion probability between the partial syllable strings is calculated from the statistical recognition error tendency of the partial syllable string, and the product of the confusion probabilities of all the partial syllable strings is calculated. It is possible to use a technique for obtaining the similarity of whole words as a logarithm. Below, the similarity using this syllable information is called acoustic similarity.
Non-Patent Document 5: Abe et al .: “Large-scale continuous speech recognition using a two-stage search method using a probability error probability model”, IEICE Journal, Vol. J83-D-II, no. 12, pp. 2545-2553, 2000.
The above steps ST102 to ST104 are repeated for all the words stored in the reading / syllable storage means 110 (step ST105).

After obtaining the reading similarity r _i and the acoustic similarity a _i with respect to the correction target word for all the words i stored in the reading / syllable storage unit 110, the correction candidate generation unit 111 then determines the correction target word and The reading similarity and the acoustic similarity between all the words stored in the reading / syllable storage means 110 are normalized by dividing by the sum of reading similarities and the sum of acoustic similarities, respectively (however, n is the total number of words), as in the following equation to calculate a weighted sum of both similarity to the word similarity between s _i (step ST 106).

  In the above formula, α is a weight that determines how much importance is given to the reading similarity or the acoustic similarity to calculate the word similarity. α can be arbitrarily set according to the use environment of voice recognition. When α = 0, only the acoustic similarity is used. Conversely, when α = 1, only the reading similarity is used. Will do.

  As described above, the correction candidate generation unit 111 calculates the inter-word similarity between the correction target word selected by the user and all the words stored in the reading / syllable storage unit 110, and then the inter-word similarity. Are arranged in descending order of the top m words as correction candidates (step ST107). The number m is arbitrary. For example, when m = 3 and the correction target word is “Mitsui”, for example, words such as “Mitsuishi”, “Mizui”, “Mitsubishi” with high similarity between words are used. Will be selected as a correction candidate.

Next, the correction candidate generation unit 111 acquires the number of search candidates when a correction candidate is selected for each correction candidate (step ST108). For example, for the correction candidate “Mitsubishi”, the number of search result candidates indicating how many of the search results output by the search means 105 include “Mitsubishi” is acquired. Finally, the correction candidate generation unit 111 outputs the correction candidates and the number of search result candidates corresponding to each correction candidate.
The correction candidate display means 112 displays the correction candidates and the number of search result candidates output from the correction candidate generation means 111 simultaneously to the user as shown in FIG. 7 using a display device such as an LCD display. At this time, it is desirable to lay out so that the higher the similarity of the correction candidates, the closer to the correction target word.

  When the correction candidate display unit 112 displays the correction candidates and the number of search result candidates, the correction candidate selection unit 113 accepts an operation of selecting a correction candidate intended by the user using an input device such as a mouse or a touch panel, and the user When the selection operation is performed by the above, the selected correction candidate is output.

Finally, the correction execution unit 114 executes correction of the already displayed speech recognition result with the correction candidate selected by the user as an input. For example, in FIG. 7, when the correction candidate “Mitsubishi” is selected as the correction candidate for the correction target word “Mitsui”, the correction target word “Mitsui” and the correction candidate “Mitsubishi” are replaced. The display is as shown in FIG. When the correction is executed by the correction execution unit 114, the search unit 105 executes the search again from the corrected word. In addition, when the process of using the speech recognition result continues in the subsequent stage, the correction execution unit 114 notifies the appropriate place that the recognition result has been corrected.
The above processing is repeated until a user's desired search result is obtained.
The above is the processing flow of the voice search interface device according to the present invention.

  In the first embodiment, the case where the speech recognition result is a word string composed of a plurality of words has been described. However, the present invention can be similarly applied even if the speech recognition result is only one word. However, when the word is only the first word, the OR search and the AND search are not performed.

  As described above, according to the speech search interface device of the first embodiment, when a word output unit that outputs a word or a word string as a recognition result for speech input and an arbitrary word or word string are given, Search means for searching a database and outputting a search result of word or word string and the number of search candidates, word dictionary storage means in which word information is registered, word output by word output means, and word dictionary storage means The correction candidate generating means for performing the matching with the words registered in, and generating correction candidates in units of words, and the number of search candidates for the respective correction candidates generated by the correction candidate generating means are acquired via the search means. According to the correction candidate search candidate number acquisition means, the correction candidates generated by the correction candidate generation means, and the respective correction candidates acquired by the correction candidate search candidate number acquisition means Since a correction candidate output means for outputting the retrieval candidate number can be when a user narrow down the specific information from the database by using the voice recognition, Filter efficiently search results.

  Further, according to the voice search interface device of the first embodiment, the search means performs an OR search including any of the plurality of words when the word output means outputs a plurality of words when searching the database. As a result, the user can easily know what search results can be obtained from the speech recognition results.

  Further, according to the voice search interface device of the first embodiment, a plurality of words are output by the word output means, and when any one of the plurality of words is selected as a narrowed word, the selected word is selected. Since the search result narrowing means for narrowing down the search result of the search means is provided only for AND search results including all words, efficient narrowing can be performed.

  Further, according to the speech search interface device of the first embodiment, when the correction target word is designated for any of the words output by the word output means, the correction candidate generation means determines the syllable of the correction target word. Using both the similarity to the correction candidate syllable and the similarity between the reading of the correction target word and the correction candidate, and generating the correction candidate with the sum total weighted to both similarities as the overall similarity Therefore, it is possible to generate an appropriate correction candidate.

  Further, according to the voice search interface device of the first embodiment, a plurality of words are output by the word output means, and when any one of the plurality of words is selected as a narrowed word, the selected word is selected. Since the search result narrowing means for acquiring the number of search candidates when performing an AND search of the word and the unselected word and the voice recognition result display means for displaying the number of search candidates acquired by the search result narrowing means are provided. The user can narrow down the search results efficiently.

  Further, according to the speech input search method of the first embodiment, a word output step for outputting a word or a word string as a recognition result for the speech input, and a specific database is searched when an arbitrary word or word string is given Then, a search step for outputting a search result of word or word string and the number of search candidates, a word output in the word output step, and a word registered in the word dictionary storage means are matched, and word unit correction is performed. The correction candidate generation step for generating candidates, the number of search candidates for each correction candidate generated in the correction candidate generation step, the correction candidate search candidate number acquisition step for acquiring by the search step, and the correction candidate generation step Outputs correction candidates and the number of search candidates corresponding to each correction candidate acquired in the correction candidate search candidate number acquisition step. Since a correction candidate output step, when the user narrow down the specific information from the database by using the voice recognition, the user can narrow down efficiently search results

  101 voice input means, 102 voice recognition dictionary storage means, 103 voice recognition means, 104 search database, 105 search means, 106 voice recognition result display means, 107 narrowed word selection means, 108 search result narrowing means, 109 correction target word selection Means 110 reading / syllable storage means 111 correction candidate generation means 112 correction candidate display means 113 correction candidate selection means 114 correction execution means

Claims (6)

Word output means for outputting a word or a word string as a recognition result for voice input;
Search means for searching a specific database and outputting a search result of the word or word string and the number of search candidates when given an arbitrary word or word string;
Word dictionary storage means in which word information is registered;
A correction candidate generating unit that performs matching between the word output by the word output unit and the word registered in the word dictionary storage unit, and generates a correction candidate in units of words;
Correction candidate search candidate number acquisition means for acquiring the number of search candidates for each correction candidate generated by the correction candidate generation means via the search means;
A speech search interface comprising: a correction candidate generated by the correction candidate generation means; and a correction candidate output means for outputting the number of search candidates corresponding to each correction candidate acquired by the correction candidate search candidate number acquisition means. apparatus.   2. The voice search interface device according to claim 1, wherein the search means performs an OR search including any of the plurality of words when the word output means outputs a plurality of words when searching the database. .   When a plurality of words are output by the word output unit and any one of the plurality of words is selected as a narrowed word, only the AND search result including all the selected words 2. The voice search interface device according to claim 1, further comprising search result narrowing means for narrowing search results.   When a correction target word is specified for any of the words output by the word output unit, the correction candidate generation unit includes a similarity between the syllable of the correction target word and the syllable of the correction candidate, and the correction target word. 2. The method according to claim 1, wherein both the readings of the readings of the correction candidates and the readings of the correction candidates are used, and the sum total weighted to both similarities is used as information for generating correction candidates as the overall similarity. The voice search interface device described. When a plurality of words are output by the word output means and any one of the plurality of words is selected as a narrowed word, and an AND search is performed on the selected word and the unselected word Search result narrowing means to obtain the number of search candidates for,
The voice search interface device according to claim 1, further comprising voice recognition result display means for displaying the number of search candidates acquired by the search result narrowing means. A word output step for outputting a word or word string as a recognition result for voice input;
When an arbitrary word or word string is given, a search step of searching a specific database and outputting a search result of the word or word string and the number of search candidates;
A correction candidate generation step of matching the word output in the word output step with a word registered in the word dictionary storage means to generate a correction candidate in units of words;
A number of correction candidate search candidates for acquiring the number of search candidates for each of the correction candidates generated in the correction candidate generation step by the search step;
A voice input search comprising: a correction candidate generated in the correction candidate generation step; and a correction candidate output step for outputting the number of search candidates corresponding to each correction candidate acquired in the correction candidate search candidate number acquisition step Method.

Images