JP2012014042A - Audio input interface device and audio input method - Google Patents

Abstract

PROBLEM TO BE SOLVED: To obtain an audio input interface device capable of presenting an accurate correction candidate without explicitly correcting a section by a user when the user corrects a false recognition result in voice recognition.SOLUTION: Voice recognition means 103 outputs a word string as a recognition result for audio input. Linking correction candidate generating means 109 performs matching between a phrase obtained by linking all words outputted from the voice recognition means 103 and a phrase in linking reading and syllable storage means 108 for registering therein information on a phrase obtained by linking a plurality of words, and generates a word unit of correction candidate. Difference extraction means 110 extracts a difference in a matching result. Correction candidate display means 111 displays the difference extracted by the difference extraction means 110 as the correction candidate.

Description

  The present invention relates to a voice input interface device and a voice input method, and in particular, a correction unit that allows a user to easily correct a voice recognition result when a voice recognition result different from the user's intention is output in voice recognition. The present invention relates to a voice input interface device and a voice input method provided.

  In speech recognition, the recognition rate varies depending on the usage environment and individual differences of users, and there is an essential problem that erroneous recognition occurs. As a method of increasing the recognition rate and suppressing erroneous recognition, there are a method of tuning a recognition parameter by a user and a method of narrowing a recognition vocabulary depending on a scene. However, when it is assumed that a large number of unspecified users will use it, and it is necessary to target a large vocabulary such as a facility name search of a car navigation system, the above solution is essential. It is not a 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 Document 1, 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 easily corrected by the user simply selecting a desired word from the word list. An apparatus is disclosed. The technique described in Patent Literature 1 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 with a high contention probability for each word segment. Is generated as a correction candidate. However, when word segmentation fails, a phenomenon occurs in which a desired word does not appear in the list of correction candidate words.
On the other hand, when word segmentation fails in speech recognition, a method in which a user corrects a word segment to redivide the segment into correct segments and re-present candidates to be corrected is disclosed in Non-Patent Document 1, for example. Although disclosed, in this case, it is necessary for the user to manually correct the section, which increases labor.

JP 2006-146008 A

Endo, Terada: “Interactive candidate selection method for speech input”, Interaction 2003, p. p. 195-196, 2003.

  In conventional speech recognition, as shown in Non-Patent Documents 2 to 4 to be described later, a word n-gram model can be used as a language model to recognize words with a high degree of freedom in word concatenation. Because of the limitation of learning data, it is difficult to implement n equal to or greater than 3, and there is a high possibility of misrecognizing long word chains. In addition, there is no interface that can easily and quickly correct the recognition result when erroneous recognition occurs including the divided section between each word.

  The present invention has been made to solve the above-described problems. When a user corrects a misrecognition result in speech recognition, a correct correction candidate is presented without an explicit correction of a section by the user. An object of the present invention is to obtain a voice input interface device and a voice input method that can be used.

  The voice input interface device according to the present invention includes a word string output unit that outputs a word string of a recognition result for voice input, a connection information storage unit in which information on a phrase that connects a plurality of words is registered, and a word string output unit Generated by the linked correction candidate generating means, which generates a correction candidate for each phrase by matching the phrase obtained by concatenating all the words output in step 1 and the phrase registered in the linked information storage means, and generated by the linked correction candidate generating means Correction candidate output means for outputting the corrected correction candidates.

  The speech input interface device according to the present invention performs matching between a phrase obtained by concatenating all the words output by the word string output unit and a phrase registered in the concatenation information storage unit, and generates a correction candidate for each phrase. Therefore, when the user corrects the misrecognition result in the speech recognition, the correct correction candidate can be presented without correcting the explicit section by the user.

BRIEF DESCRIPTION OF THE DRAWINGS It is a block diagram which shows the audio | voice input interface apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the speech recognition result in the audio | voice input interface device of Embodiment 1 of this invention. It is a flowchart which shows the process which the word correction candidate production | generation means in the speech input interface device of Embodiment 1 of this invention performs. It is explanatory drawing which shows the information of the word reading and syllable storage means in the audio | voice input interface apparatus of Embodiment 1 of this invention. It is a flowchart which shows the process which the connection correction candidate production | generation means in the audio | voice input interface apparatus of Embodiment 1 of this invention performs. It is explanatory drawing which shows the information of the joint reading and syllable storage means in the audio | voice input interface apparatus of Embodiment 1 of this invention. It is explanatory drawing which shows the speech recognition result in the audio | voice input interface apparatus of Embodiment 1 of this invention, a word correction candidate, and a connection correction candidate. It is explanatory drawing at the time of replacing the correction object word in the audio | voice input interface apparatus of Embodiment 1 of this invention with the word correction candidate. It is explanatory drawing when the connection correction candidate in the audio | voice input interface apparatus of Embodiment 1 of this invention is selected. It is a block diagram which shows the audio | voice input interface apparatus of Embodiment 2 of this invention. It is a flowchart which shows the narrowing-down process in the audio | voice input interface apparatus of Embodiment 2 of this invention.

Embodiment 1 FIG.
FIG. 1 is a configuration diagram of the voice input interface device according to the first embodiment of the present invention.
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 string output means) 103 receives the digital voice signal converted by the voice input means 101, recognizes the voice with reference to the voice recognition dictionary storage means 102, and outputs the voice recognition result. The voice recognition result display means 104 presents the voice recognition result to the user using a display device such as an LCD display. The correction target word selection unit 105 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 the voice recognition result, and when the user performs a selection operation. , The correction target word is output.

  The word reading / syllable storage means (word information storage means) 106 is a storage device that stores notation of words to be recognized, reading information, and syllable information. The word correction candidate generation unit 107 generates and outputs a word having a high similarity as a word correction candidate with respect to the correction target word selected by the user. At this time, word correction candidates are generated using word-by-word reading information and syllable information stored in the word reading / syllable storage means 106.

  The connected reading / syllable storage means (connected information storage means) 108 is a storage device that stores reading information and syllable information of a phrase to which one or more words to be recognized are connected. The link correction candidate generation unit 109 generates and outputs a word having a high degree of similarity as a link correction candidate for a word obtained by connecting all the words in the speech recognition result. At this time, linked correction candidates are generated using the reading information and syllable information stored in the linked reading / syllable storage means 108. The difference extraction unit 110 extracts a difference phrase between the generated connection correction candidate and the recognition result already presented to the user.

  The correction candidate display means (correction candidate output means) 111 presents both the generated word correction candidate and the difference phrase simultaneously to the user as a correction candidate using a display device such as an LCD display. The correction candidate selection unit 112 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 a selection operation is performed by the user. The correction execution unit 113 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 flow of processing of the voice input 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 speech recognition unit 103 recognizes the speech by referring to the speech recognition dictionary stored in the speech recognition dictionary storage unit 102 using the converted digital speech signal as an input, and outputs a speech recognition result (word Column output step). The method for recognizing speech is arbitrary, and a known speech recognition method as described in Non-Patent Documents 2 to 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

  The voice recognition result display means 104 presents the result of the voice recognition and the word segmentation section to the user using a display device such as an LCD display in a layout that allows the user to know the word segmentation section. For example, for the utterance “Mitsubishi Denki Kabushiki Gaisha”, it is assumed that the result of speech recognition is “Mitsui / City Electric / Equipment / Stock / Company” (/ represents a word break). At this time, as shown in FIG. 2, it is desirable to display the divided words so that the words are displayed sequentially from left to right according to the utterance order.

  If the result of speech recognition displayed by the speech recognition result display means 104 is the result as intended by the user, the user can use the speech recognition result as it is, but if the result is not as intended, the recognition result is corrected. There is a need to. For example, for the utterance “Mitsubishi Denki Kabushiki Gaisha”, if the result of speech recognition is “Mitsui / City / Equipment / Stock / Company” as described above, a result not intended by the user is output. In addition, the word segmentation and number of divisions are also incorrect. Specifically, the word “Mitsubishi / Denki / Kabushiki / Gaisha” should be separated from the word “Mitsubishi / Denki / Kabushiki / Gaisha” (number of divisions 4). (Division number 5). Therefore, it is necessary for the user to correct the recognition result including not only the word but also the divided section of the word.

  Therefore, the correction target word selection unit 105 accepts an operation for the user to select a word to be corrected using an input device such as a mouse or a touch panel. When the user selects a word to be corrected, the correction target word selecting unit 105 detects the selection operation and outputs the selected correction target word. Here, it is assumed that “Mitsui” is selected as the correction target word among the five words “Mitsui / city tram / equipment / stock / company”.

Next, the flow of processing performed by the word correction candidate generation unit 107 will be described in detail with reference to the flowchart of FIG.
The word correction candidate generation unit 107 receives the correction target word output from the correction target word selection unit 105 as an input, and firstly reads the correction target word from the reading information and syllable information stored in the word reading / syllable storage unit 106. Reading and syllable information are acquired (step ST101). Here, the word reading / syllable storage means 106 stores 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. 4, and further corresponds to the notation. The reading and syllable information are preferably stored. The word correction candidate generation unit 107 searches for a correction target word from the word reading / syllable storage unit 106 and acquires corresponding reading and syllable information. For example, when the correction target word is “Mitsui”, “Mitsui” is searched from the notation of the word reading / syllable storage means 106, and the reading “Mitsui” and the syllable “mi-cu-i” are acquired.

Next, an arbitrary word is selected from the word reading / syllable storage means 106 (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:
1. "Mitsui"
2. "Mitsubi" (Replace I with Bi)
3. "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 word correction candidate generation means 107 calculates the similarity between the syllable of the correction target word and the syllable of the word selected in step 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-step 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 word reading / syllable storage means 106 (step ST105).

When the reading similarity r _i and the acoustic similarity a _i are obtained for all the words i, the word correction candidate generation unit 107 then stores the correction target word and all of the words stored in the word reading / syllable storage unit 106. Normalization is performed by dividing the reading similarity and acoustic similarity between two words by the sum of reading similarities and the sum of acoustic similarities (where n is the total number of words). The weighted sum of degrees is calculated and set as the similarity between words s _i (step ST106).

  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.

  In this way, the word correction candidate generation unit 107 calculates the inter-word similarity between the correction target word selected by the user and all the words stored in the word reading / syllable storage unit 106, and finally the word Arranged in descending order of the similarity between words, the top n words are generated as word correction candidates (step ST107). The number n is arbitrary. For example, when n = 3 and the correction target word is “Mitsui”, for example, a word such as “Mitsui”, “Mitsu”, “Mizui” or the like having a high degree of similarity between words. Will be selected as a word correction candidate.

  Now, if the divided section between words is correct, the correction candidate intended by the user is likely to appear at the top of this word correction candidate, but if the divided section between words is incorrect, the user's The possibility that the intended correction candidate appears at the top of this word correction candidate is reduced. This is because, when the divided section between words is incorrect, both the reading similarity and the acoustic similarity of the correction target word and the word intended by the user are reduced.

In order to solve this problem, a method in which a user first corrects a divided section between words and then generates a word correction candidate is the method described in Non-Patent Document 1 described above. However, this method increases the time and effort for the user to correct the divided section between words.
Therefore, in the present embodiment, in order to generate a correction candidate intended by the user without causing the user to correct the divided section between words even when the divided section between words is incorrect. Then, all the words appearing in the speech recognition result are connected, and the degree of similarity is obtained for the connected phrases (concatenated correction candidate generation step, correction candidate output step). Specifically, the link correction candidate generation unit 109 uses the words stored in the linked reading / syllable storage unit 108 with respect to a phrase obtained by connecting a plurality of words output as a recognition result by the speech recognition unit 103. A word with high similarity is generated as a concatenation correction candidate. The processing of the link correction candidate generation unit 109 will be described in detail according to the flowchart of FIG.

  First, with reference to the word reading / syllable storage means 106, readings and syllables of all words appearing in the speech recognition result are acquired (step ST201). Next, the phrase which connected those readings and syllables is produced | generated (step ST202). For example, if the voice recognition result is “Mitsui / City / Equipment / Stock / Company”, the reading “Mitsui / Siden / Kiki / Kabushiki / Gaisha” is acquired, and the phrase “Mitsui Shidenki Kabushiki Gaisha” concatenated all the readings Is generated. Also, the syllable “mi-cu-i / si-de-n / ki-ki / ka-bu-si-ki / ga-i-sja” is acquired and the syllable “mi−” is obtained by connecting all the syllables. cu-i-si-de-n-ki-ki-ka-bu-si-ki-ga-i-sja ". Next, an arbitrary word / phrase is selected from the concatenated reading / syllable storage means 108 (step ST203), and the similarity between the reading and syllable of the word concatenated in step ST202 and the reading and syllable of the word selected in step ST203 is determined. Calculate (step ST204, step ST205). Here, the connected reading / syllable storage means 108 stores the notation of the recognition target phrase in the form as shown in FIG. 6, and further stores the reading and syllable information corresponding to the notation. desirable. In calculating the similarity between words, the weighted sum of reading similarity and acoustic similarity is used (step ST206), as in the case of obtaining the similarity between words.

  The above steps ST203 to ST206 are repeated for all the words stored in the linked reading / syllable storage means 108. In this way, the connection correction candidate generation unit 109 calculates the similarity between the words obtained by connecting all the words in the recognition result and all the words stored in the connection reading / syllable storage unit 108 (step (ST207) Finally, one or a plurality of higher-order words are arranged as concatenated correction candidates arranged in descending order of similarity between words (step ST208). For example, if the phrase that concatenates all the words in the recognition result is “Mitsubishi Denki Kabushiki Geisha”, “Mitsubishi Electric Corporation (Mitsubishi Denki Kabushiki Geisha)” or “Mitsui Electric Co., Ltd.” that has a high degree of similarity between the linked words. A phrase such as “Company (Mitsui Denki Kabushiki Gaisha)” is generated as a candidate for concatenation correction.

  Next, the difference extraction unit 110 extracts a difference between the link correction candidate generated by the link correction candidate generation unit 109 and the already-explained phrase as a recognition result. For example, as in the above example, if the recognition result is “Mitsui / City Electric / Equipment / Stock / Company” and the consolidation candidate is “Mitsubishi Electric Corporation”, the word parts of “Stock” and “Company” are Since they are the same, the difference phrase is the “Mitsubishi Electric” part. In the present embodiment, since the final display form is used as a notation, the extraction of difference words is not performed between readings but between notations.

As described above, when the word correction candidate generation unit 107 generates a word correction candidate and the difference extraction unit 110 extracts a difference phrase from the linked correction candidates, the correction candidate display unit 111 displays the generated word correction candidate and the difference phrase. Both are simultaneously presented to the user as correction candidates using a display device such as an LCD display. Although the layout at the time of display is arbitrary, as shown in FIG. 7, the word correction candidates are displayed in a list form under the correction target word selected by the user, and the difference word / phrase is displayed at a glance on the speech recognition result. It is desirable to lay out and display as you can see.
Next, the correction candidate selection unit 112 receives an operation for the user to select a correction candidate using an input device such as a mouse or a touch panel. When the user selects a correction candidate, the correction candidate selection unit 112 senses the selection operation and outputs the selected correction target word.

Finally, the correction execution means 113 executes correction of the already displayed speech recognition result with the correction candidate selected by the user as an input. For example, in the example of FIG. 7, when the word correction candidate “Mitsuki” is selected as the correction candidate for the correction target word “Mitsui”, the correction target word “Mitsui” and the word correction candidate “Mitsuki” "Is replaced and displayed as shown in FIG. On the other hand, when “Mitsubishi Electric” is selected as the candidate for correction of the difference word with the linked correction candidate, the word string “Mitsui / city / device” corresponding to the difference word is replaced with “Mitsubishi Electric”. It is displayed as 9. In addition, when the process of using the speech recognition result continues in the subsequent stage, the correction execution unit 113 notifies the appropriate place that the recognition result has been corrected.
The above is the processing flow of the voice input interface device according to the present embodiment.

In the above example, because the word segmentation section is incorrect in speech recognition, the correction candidate “Mitsubishi” intended by the user cannot be presented to the correction target word “Mitsui”, but the recognition result “Mitsui / By calculating the similarity by concatenating the words “city / device / stock / company”, the intended correction candidate “Mitsubishi Electric” can be generated.
Further, in the present embodiment, for the sake of simplicity, the similarity is calculated by connecting words, but the same effect can be obtained even by connecting characters.

  As described above, according to the first embodiment, even if a result unintended by the user is output in speech recognition and the word segmentation section is also incorrect, the correction candidate for each word of the recognition result and the recognition result The correction interface for presenting correct correction candidates can be presented at the same time without requiring the user to correct the divided sections of the words. Can be provided.

  As described above, according to the voice input interface device of the first embodiment, the word string output means for outputting the word string of the recognition result for the voice input, and the link information in which the information on the phrase connecting the plurality of words is registered. A storage means, a phrase obtained by concatenating all the words output by the word string output means, and a phrase registered in the linkage information storage means, and a linked correction candidate generating means for generating a correction candidate in phrase units; And a correction candidate output means for outputting the correction candidates generated by the linked correction candidate generation means, so that when the user corrects the erroneous recognition result in the speech recognition, the user does not correct the explicit section. , Correct correction candidates can be presented.

  In addition, according to the voice input interface device of the first embodiment, a word string output unit that outputs a word string as a recognition result for voice input, and a link information storage unit in which information on a phrase that links a plurality of words is registered. A connected correction candidate generating means for matching a phrase obtained by connecting all the words output by the word string output means and a word registered in the connection information storage means, and generating a correction candidate for each phrase; Difference extraction means for extracting a difference between the correction candidate generated by the candidate generation means and a phrase obtained by connecting all the words output by the word string output means, and the difference extracted by the difference extraction means is output as a correction candidate. Since the correction candidate output means is provided, the correction candidate can be confirmed more easily.

  Further, according to the voice input interface device of the first embodiment, the word information storage means for the word information storage means in which information in units of words is registered and the words in the word string output by the word string output means. Since the word correction candidate generating means for generating the word-by-word correction candidates from the words stored in is provided, more accurate correction candidates can be presented.

  In addition, according to the voice input interface device of the first embodiment, the correction target word selecting unit that receives the specification of the correction target word for the word string output by the word string output unit is provided, and the word correction candidate generation unit includes the correction target word. Using both the similarity between the syllable and the correction candidate syllable, and the similarity between the reading of the correction target word and the correction candidate, the correction candidate is determined with the sum total weighted on both similarities as the overall similarity. Since it is used for the information at the time of generation, it is possible to present an accurate correction candidate word.

  In addition, according to the speech input interface device of the first embodiment, the linking correction candidate generating unit, when a phrase obtained by concatenating all the words output by the word string output unit is a correction target phrase, the syllable of the correction target phrase And the correction candidate syllable similarity, and both the reading of the correction target phrase and the correction candidate reading are used, and a correction candidate is generated with the sum total weighting both similarities as the overall similarity. Since it is used for the information at the time, it is possible to present the correct correction candidate words.

  According to the voice input interface device of the first embodiment, the recognition result display means for displaying the output result of the word string output means, and the correction for accepting the designation of the correction target word for the recognition result displayed by the recognition result display means. Target word selection means, correction candidate display means for displaying the linked correction candidates generated by the linked correction candidate generation means, correction candidate selection means for receiving the user's selection results for the correction candidates displayed by the correction candidate display means, Since the correction execution means for correcting the output result of the word string output means based on the selection result received by the correction candidate selection means is provided, it is possible to provide an interface for easy voice input and correction candidate selection.

  Further, according to the speech input method of the first embodiment, a word sequence output step for outputting a word sequence as a recognition result for speech input, a phrase obtained by connecting all the words output in the word sequence output step, and connection information Matching with a phrase obtained by concatenating a plurality of words registered in the storage means and generating a correction candidate generated in the linked correction candidate generation step, and a correction candidate generated in the linked correction candidate generation step Since the candidate output step is provided, when the user corrects the misrecognition result in the speech recognition, the correct correction candidate can be presented without correcting the explicit section by the user.

Embodiment 2. FIG.
In the first embodiment, in order to present a correction candidate intended by the user even when the divided section is incorrect, a linked correction candidate for a phrase obtained by connecting words output as a speech recognition result is generated. At this time, the linked correction candidates can be further narrowed down according to the temporal position of the correction target word selected by the user. That is, as shown in FIG. 2, when the speech recognition result is displayed laid out from left to right according to the utterance order, when the user corrects the speech recognition result, the correction is performed in order from the leftmost word. Natural to think. Therefore, when the user selects the word appearing second and later from the left as the correction target word, the recognition result of the word appearing before the correction target word is assumed to be correct, and the linked correction candidates can be narrowed down. For example, in FIG. 2, when the third “equipment” from the left is selected, it is possible to connect to “Mitsui” and “city tram” on the assumption that the word unit recognition results of “Mitsui” and “city tram” are correct. Correction candidates can be narrowed down only to simple word correction candidates and linked correction candidates.

  Therefore, in the second embodiment, in addition to the configuration of the voice input interface device in the first embodiment, when the user selects a word appearing second or later in the voice recognition result as a correction target word, it appears before the correction target word. The speech input interface device having the function of narrowing down the word correction candidates and the connection correction candidates is assumed as the correct word recognition result.

  FIG. 10 shows a configuration diagram of the voice input interface device according to the second embodiment of the present invention. As shown in FIG. 10, only a portion in which the word correction candidate narrowing means 114 is inserted in the subsequent stage of the word correction candidate generating means 107 and the connection correction candidate narrowing means 115 is inserted in the subsequent stage of the connection correction candidate generating means 109. However, this is different from the first embodiment. That is, the word correction candidate narrowing means 114 is a means for narrowing down the word correction candidates generated by the word correction candidate generating means 107 according to the position in the word string of the correction target word received by the correction target word selecting means 105. Further, the linked correction candidate narrowing means 115 is a means for narrowing down the linked correction candidates generated by the linked correction candidate generating means 109 according to the position of the correction target word received by the correction target word selecting means 105 in the word string. Since the configuration other than these is the same as that of the first embodiment, the description of the configuration and operation of other portions is omitted.

  First, the flow of processing performed by the word correction candidate narrowing means 114 will be described with reference to the flowchart of FIG. As shown in FIG. 11, when the correction target word selected by the user in the correction target word selection means 105 is the first word in the speech recognition result, the word correction candidate narrowing means 114 does not perform any processing ( Step ST301; NO). On the other hand, when the correction target word selected by the user is the second and subsequent words in the speech recognition result (step ST301; YES), the word correction candidate narrowing means 114 acquires the recognition result before the correction target word (step ST302). . For example, it is assumed that the user tries to input “Mitsubishi Electric Corporation” by voice and speaks “Mitsubishi Denki Kabushiki Geisha” and the voice recognition result is “Mitsubishi / Electronic / Equipment / Co., Ltd.”. At this time, when the user selects “electronic” as the correction target word, the speech recognition result “Mitsubishi” before the correction target word is acquired.

  Next, the word correction candidate narrowing means 114 can connect to the speech recognition result before the correction target word from the word correction candidates generated by the word correction candidate generation means 107 with reference to the linked reading / syllable storage means 108. By selecting only correct word correction candidates, processing for narrowing down word correction candidates is performed (step ST303). For example, in the above example, only the word correction candidates connectable to the word “Mitsubishi” before the correction target word are selected from the word correction candidates for the correction target word “electronic” with reference to the linked reading / syllable storage unit 108. Select to narrow down word correction candidates. In this way, by narrowing down the word correction candidates using the word information before the correction target word, there is an effect that the possibility that the correction result intended by the user appears at the top of the word correction candidates can be improved.

  On the other hand, the flow of processing performed by the linked correction candidate narrowing means 115 is the same as the flow of processing in which the word correction candidate is replaced with the linked correction candidate in the processing flow performed by the word correction candidate narrowing means 114 of FIG.

  As described above, according to the voice input interface device of the second embodiment, the correction target word selection unit that receives the specification of the correction target word for the word string output by the word string output unit, and the position of the correction target word in the word string Accordingly, since the word correction candidate narrowing means for narrowing down the word correction candidates generated by the word correction candidate generating means is provided, the possibility that the correction result intended by the user appears at the top of the correction candidates can be improved. .

  In addition, according to the voice input interface device of the second embodiment, the correction target word selecting unit that receives the specification of the correction target word for the word string output by the word string output unit, and the correction target word received by the correction target word selecting unit Since there is a link correction candidate narrowing means for narrowing down the link correction candidates generated by the link correction candidate generation means according to the position in the word string, the correction result intended by the user may appear at the top of the link correction candidates Can be improved.

  101 speech input means, 102 speech recognition dictionary storage means, 103 speech recognition means, 104 speech recognition result display means, 105 correction target word selection means, 106 word reading / syllable storage means, 107 word correction candidate generation means, 108 concatenated reading / Syllable storage means, 109 linkage correction candidate generation means, 110 difference extraction means, 111 correction candidate display means, 112 correction candidate selection means, 113 correction execution means, 114 word correction candidate narrowing means, 115 linked correction candidate narrowing means.

Claims (9)

A word string output means for outputting a word string of a recognition result for voice input;
A concatenation information storage means in which information on a phrase obtained by concatenating a plurality of words is registered;
A combined correction candidate generating unit that performs matching between a phrase obtained by connecting all the words output by the word string output unit and a word registered in the connection information storage unit, and generates a correction candidate in units of words;
A speech input interface device comprising correction candidate output means for outputting the correction candidates generated by the linked correction candidate generation means. A word string output means for outputting a word string of a recognition result for voice input;
A concatenation information storage means in which information on a phrase that concatenates a plurality of words is registered;
A combined correction candidate generating unit that performs matching between a phrase obtained by connecting all the words output by the word string output unit and a word registered in the connection information storage unit, and generates a correction candidate in units of words;
A difference extraction means for extracting a difference between the correction candidate generated by the connection correction candidate generation means and a phrase obtained by connecting all the words output by the word string output means;
A voice input interface device comprising correction candidate output means for outputting the difference extracted by the difference extraction means as a correction candidate. Word information storage means in which word unit information is registered;
A word correction candidate generating unit that generates a word-by-word correction candidate from words stored in the word information storage unit with respect to the words in the word string output by the word string output unit. The voice input interface device according to claim 1 or 2. A correction target word selection means for accepting designation of a correction target word for the word string output by the word string output means;
The word correction candidate generation means uses both the similarity between the syllable of the correction target word and the syllable of the correction candidate, and the similarity between the reading of the correction target word and the correction candidate, and weights both similarities. 4. The voice input interface apparatus according to claim 3, wherein the added sum is used as information for generating a correction candidate as an overall similarity.   The link correction candidate generating means, when a phrase obtained by connecting all the words output by the word string output means is a correction target phrase, the similarity between the syllable of the correction target phrase and the syllable of the correction candidate, and the correction target The use of both the word reading and the correction candidate reading similarity, and the sum total weighted by both similarities is used as information for generating a correction candidate as an overall similarity. The voice input interface device according to claim 1 or 2. Correction target word selection means for accepting designation of a correction target word for the word string output by the word string output means;
4. The voice input interface device according to claim 3, further comprising word correction candidate narrowing means for narrowing down the word correction candidates generated by the word correction candidate generating means according to the position of the correction target word in the word string. . Correction target word selection means for accepting designation of a correction target word for the word string output by the word string output means;
And a connected correction candidate narrowing means for narrowing down the connected correction candidates generated by the connected correction candidate generating means according to the position of the correction target word received by the correction target word selecting means in the word string. The voice input interface device according to any one of claims 1, 2, and 5. Recognition result display means for displaying the output result of the word string output means;
Correction target word selection means for accepting designation of a correction target word for the recognition result displayed by the recognition result display means;
Correction candidate display means for displaying the connection correction candidates generated by the connection correction candidate generation means;
Correction candidate selection means for receiving a user's selection result for the correction candidates displayed by the correction candidate display means;
3. The voice input interface device according to claim 1, further comprising a correction execution unit that corrects an output result of the word string output unit based on a selection result received by the correction candidate selection unit. A word string output step for outputting a word string of a recognition result for voice input;
Concatenated correction that matches a phrase obtained by concatenating all the words output in the word string output step with a phrase obtained by concatenating a plurality of words registered in the concatenated information storage unit, and generates a correction candidate for each phrase. A candidate generation step;
A speech input method comprising: a correction candidate output step for outputting the correction candidates generated in the linked correction candidate generation step.

Images