JP2011215291A - Speech recognition device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a speech recognition device and program, which appropriately changes whether recognition processing is performed in forward or reverse direction, according to speech uttered by a user.SOLUTION: Based on a forward direction dictionary, recognition processing is performed in the forward direction by reproducing speech information in the forward direction (S401), and based on the recognition result of the forward direction recognition processing, a position of a fixed word in speech information is estimated (S403 and S405). When it is estimated that the fixed word is in the latter half of the speech information (S405), recognition processing is performed in reverse direction by reproducing speech information in the reverse direction (S406) based on a reverse direction dictionary. Based on a position of the fixed word of the speech uttered by the user, it is appropriately changed whether recognition processing of the speech information is performed in the forward direction or reverse direction, and thereby, accuracy of the speech recognition is improved without making the user conscious of operation about the speech recognition direction.

Description

  The present invention relates to a speech recognition apparatus and a program.

  2. Description of the Related Art Conventionally, a voice recognition device that recognizes a voice uttered by a user is known. In speech recognition processing, it is common to perform recognition processing in order from the user's voice, but for example, when recognizing addresses in Europe and the United States that are in order of small administrative units to large administrative units, Is detected by a user's switch operation or the like, and a technique for performing backward recognition processing using a backward tree dictionary for address recognition is disclosed only in that case (see, for example, Patent Document 1). .

Japanese Patent No. 4104313

However, Patent Document 1 does not disclose that backward recognition processing is performed in speech recognition other than an address, and when backward recognition processing is performed, a user's switch operation or the like is performed in order to start backward recognition processing. It was necessary to detect. That is, in Patent Document 1, it has not been possible to appropriately switch between performing the recognition process in the forward direction or in the reverse direction according to the voice uttered by the user.
The present invention has been made in view of the above-described problems, and the purpose of the present invention is to recognize a voice that can be appropriately switched between a forward process and a backward process in accordance with a user's voice. To provide an apparatus and a program.

  The speech recognition apparatus according to claim 1, wherein the speech data when the word is reproduced in the forward direction and the dictionary data of the forward tree structure in which the word is stored in association with each other, and the voice when the word is reproduced in the reverse direction It has dictionary data of backward tree structure in which data and words are stored in association with each other. The speech recognition device is based on speech information acquisition means for acquiring speech information consisting of fixed words that form the root of the tree structure, and variable words that correspond to the fixed words and that form branches and leaves of the tree structure, and forward-facing tree structure dictionary data, Forward direction recognition processing means for performing recognition processing in the forward direction by reproducing the voice information in the forward direction; fixed word position estimation means for estimating the position of the fixed word in the speech information based on the recognition result of the forward direction recognition processing; When the fixed word position estimating means estimates that the fixed word is in the latter half of the speech information, the recognition processing is performed in the reverse direction by reproducing the speech information in the reverse direction based on the backward-facing tree structure dictionary data. Direction recognition processing means. Accordingly, the voice information recognition process can be appropriately switched between the forward direction and the reverse direction based on the position of the fixed word of the voice uttered by the user. Improves the accuracy of speech recognition without making you conscious.

  In the second aspect of the present invention, the speech recognition apparatus includes divergence determination means for determining whether or not the first half of the speech information is divergence based on the recognition result of the forward direction recognition processing means. The fixed word position estimating means estimates that the fixed word is in the latter half of the speech information when the divergence determining means determines that the first half of the speech information is diverging. Thereby, it is possible to reduce the calculation load related to the estimation of the position of the fixed word.

  In the invention according to claim 3, the divergence determining means recognizes the partial voice information that is the voice information for a predetermined period from the start of the acquisition of the voice information by the voice information acquisition means, and the recognition result of the partial voice information is If not obtained, it is determined that the first half of the voice information is diverging. Accordingly, it is possible to determine whether or not the first half of the voice information is diverging without performing the entire recognition process of the voice information, so that the calculation load can be further reduced. Further, when it is determined that the first half of the voice information is diverging, it is possible to quickly switch to the backward processing. Thereby, compared with the case where it switches to reverse direction processing after carrying out forward direction processing of all the audio | voice information, the time until a recognition result is acquired after audio | voice information is acquired can be shortened.

The determination as to whether or not the first half of the voice information is diverging can be made specifically as follows.
In the invention according to claim 4, the divergence determining means is the recognition target candidate having the highest score and the recognition target having the lowest score among the plurality of recognition target candidates having a score for specifying the recognition result of the partial speech information. If the difference between the score and the candidate is within a predetermined threshold, it is determined that the recognition result of the partial voice information is not obtained and the first half of the voice information is divergent.

  In the invention according to claim 5, the divergence determining means has a predetermined difference in score from the recognition target candidate having the highest score among a plurality of recognition target candidates having a score for specifying the recognition result of the partial speech information. If there are a predetermined number or more of recognition target candidates within the threshold value, it is determined that the recognition result of the partial speech information is not obtained and the first half of the speech information is divergent.

  As in the invention according to claim 4 or 5, when the variation in the score between the plurality of recognition candidates is small based on the score for specifying the recognition result in the plurality of recognition candidates, the first half of the voice information is diverged. Judge that Thereby, it is possible to appropriately determine whether or not the first half of the voice information is diverging.

  In the invention according to claim 6, the divergence determining means calculates the reliability of each word specified as the recognition result based on the recognition result by the forward direction recognition processing means, and the reliability of the word in the first half of the speech information. If the degree is less than or equal to a predetermined value and the reliability of a word in the latter half of the voice information is greater than or equal to the predetermined value, it is determined that the first half of the voice information is diverging. This makes it possible to appropriately determine whether or not the first half part is diverging based on the reliability of the word specified as the recognition result.

As described above, the invention has been described as the invention of the speech recognition apparatus, but it can also be realized as an invention of the following program.
In other words, the speech information acquisition means for acquiring speech information consisting of fixed words that form the root of the tree structure, and variable words that correspond to the fixed words and that form branches and leaves of the tree structure, and based on dictionary data of the forward tree structure, Forward direction recognition processing means for performing recognition processing in the forward direction by reproducing in the direction, fixed word position estimation means for estimating the position of the fixed word in the speech information based on the recognition result of the forward direction recognition processing, fixed word position estimation means When it is estimated that the fixed word is in the second half of the speech information, the backward recognition processing means performs the recognition processing in the reverse direction by reproducing the speech information in the reverse direction based on the dictionary data of the backward tree structure. A program that causes a computer to function. By executing such a program, an effect similar to that of the navigation device described above can be obtained.

It is a block diagram which shows the structure of the speech recognition apparatus of one Embodiment of this invention. It is a figure explaining the recognition dictionary of one Embodiment of this invention, Comprising: (a) is explanatory drawing explaining the dictionary data of a forward tree structure, (b) is explanatory drawing explaining the dictionary data of a backward tree structure It is. It is a flowchart explaining the speech recognition process of one Embodiment of this invention. It is a flowchart explaining the speech recognition process of one Embodiment of this invention. It is explanatory drawing explaining the recognition process based on forward tree structure data by one Embodiment of this invention. It is explanatory drawing explaining the recognition process based on backward facing tree structure data by one Embodiment of this invention. It is explanatory drawing explaining the divergence determination process by the modification of this invention.

Hereinafter, a speech recognition apparatus according to the present invention will be described with reference to the drawings.
(One embodiment)
FIG. 1 is a block diagram illustrating the overall configuration of a speech recognition apparatus 1 according to an embodiment of the present invention. The voice recognition device 1 is applied to, for example, a navigation device mounted on a vehicle. The voice recognition device 1 is configured around a control unit 10, and includes a voice recognition information storage unit 20, an operation switch group 30, a voice input unit 40, a voice output unit 50, a drawing unit 60, connected to the control unit 10. An information storage unit 70 and the like are provided.

The control unit 10 is configured as a normal computer. The control unit 10 includes a CPU, a ROM, an I / O, a bus line that connects these components, and the like.
The voice recognition information storage unit 20 is a storage device realized as, for example, a hard disk device (HDD). Although the HDD is used in the present embodiment, other media such as a DVD-ROM and a memory card may be used.

  A voice recognition database 21 is stored in the voice recognition information storage unit 20. The speech recognition database 21 includes an acoustic model 22, a recognition dictionary 23, and a language model 24. The acoustic model 22 is data in which a feature amount of speech and a phoneme are associated with each other. In this embodiment, the acoustic model 22 includes a forward model 221 that is an acoustic model for the forward direction when reproduced in the forward direction and a backward model 222 that is an acoustic model for the backward direction when reproduced in the reverse direction. Have. The recognition dictionary 23 stores words associated with phoneme strings. The language model 24 is data that models the probability of being located at the beginning or end of a sentence, the connection probability between consecutive words, the dependency relationship, and the like.

  The recognition dictionary 23 includes a forward dictionary 231 as a forward tree structure dictionary data in which a phoneme string and a word reproduced in the forward direction are associated with each other, and a phoneme string and a word reproduced in the backward direction. Are stored in the backward dictionary 232 as the backward-facing tree-structured dictionary data. For example, the word “reproduction” will be described. In the forward dictionary 231, “saisei” is stored in association with the word “reproduction” as voice data when reproduced in the forward direction. Further, the backward dictionary 232 stores “iesias” in association with the word “reproduced” as voice data when reproduced in the reverse direction. Note that “saisei”, “iesias” and the like are expressed in this way for convenience of explanation, but the main body is stored in units of phonemes.

  Here, the data structures of the forward dictionary 231 and the backward dictionary 232 will be described with reference to FIG. In FIG. 2, examples of variable words “ABC” and “DEF” associated with “reproduction”, which is a fixed word for instructing a predetermined process, will be described. The fixed word “reproduction” is stored in a “fixed word dictionary” described later, and the variable words “ABC” and “DEF” are associated with “reproduction” and stored in a “variable word dictionary” described later. Has been.

  The forward dictionary 231 is a dictionary for reproducing and recognizing the utterance content in the forward direction, and associating with the fixed word “saisei” that precedes the variable word in time series as shown in FIG. The variable words “ABC”, “DEF” and the like are stored. That is, as shown in FIG. 2A, the forward dictionary 231 has a “forward tree structure” in which a fixed word serving as the root of the tree structure is located on the start position side where the acquisition of voice information is started. I can say that. Based on the forward dictionary 231, for example, when recognizing the utterance content uttered as “reproduction ABC”, the waveform data when the acquired voice information is reproduced in the forward direction and “saisei” stored in the forward dictionary 231 are used. ”And“ ABC ”stored in association with“ saisei ”, and the content of the utterance is specified as“ reproduced ABC ”.

  On the other hand, the reverse dictionary 232 is a dictionary for reproducing and recognizing the utterance contents in the reverse direction, and associating with the fixed word “saisei” after the variable word in the time series as shown in FIG. The variable words “ABC”, “DEF” and the like are stored. That is, as shown in FIG. 2B, the backward dictionary 232 has a “backward-facing tree structure” in which a fixed word serving as the root of the tree structure is located on the end position side where the acquisition of the voice information is completed. . For example, when recognizing speech content uttered as “ABC playback” based on the reverse dictionary 232, waveform data when the acquired speech information is reproduced in the reverse direction and “iesias” stored in the reverse dictionary 232 are used. ”And“ cies ”and“ CBA ”stored in association with each other to identify the utterance content as“ ABC playback ”.

  Each of the forward dictionary 231 and the backward dictionary 232 includes fixed word dictionaries relating to fixed words that are words that instruct predetermined processing and form the basis of a tree structure. In addition, each fixed word stored in the fixed word dictionary is stored in association with a variable word dictionary related to a variable word that can be a target word of the fixed word and has branches and leaves of a tree structure. The “fixed word” in the present embodiment is a verb for instructing a predetermined process, for example, music reproduction. It can also be said that the “fixed word” is a command for instructing a predetermined process.

For example, for example, an “artist name dictionary” or “song name dictionary” as a variable word dictionary is associated with a verb “play” stored in the fixed word dictionary. Condition. For example, the word “play” is a noun, but the noun part of the word “noun + sub-variable” such as “play” is a word for instructing predetermined processing. In the form, it is stored in the fixed word dictionary as “fixed word”. Also, for example, those containing particles such as “reproduce” are registered as fixed words in the fixed word dictionary.
The recognition dictionary 23 of this embodiment stores words in a plurality of languages such as Japanese and English.

  The operation switch group 30 includes a touch switch integrated with the display 61, a mechanical switch, a remote control device, or the like, and is used for various inputs. The operation switch group 30 includes a talk switch 31. The talk switch 31 is operated at the time of voice input.

The voice input unit 40 is connected to a microphone 41 for inputting voice. When the talk switch 31 is turned on, voice information issued by the user is acquired via the microphone 41.
A speaker 51 for outputting sound is connected to the sound output unit 50.
A display 61 is connected to the drawing unit 60. The display 61 is a color display using liquid crystal or CRT. Information is displayed via the display 61.

  The information storage unit 70 is for storing the acquired voice information, and is configured by the same HDD as the voice recognition information storage unit 20. Of course, other media such as a memory card may be used.

  In this embodiment, it is appropriately determined whether the utterance content is a forward-facing tree structure or a backward-facing tree structure, and the speech processing is performed in the forward direction by reproducing the speech information in the forward direction, or the speech information is reversed in the backward direction. It is characterized in that it appropriately switches whether to perform the recognition process in the reverse direction by playing back. Therefore, the speech recognition process will be described based on the flowcharts shown in FIGS.

  Hereinafter, performing the recognition process in the forward direction by reproducing the voice information in the forward direction is referred to as “forward process”, and performing the recognition process in the reverse direction by reproducing the voice information in the reverse direction is referred to as “reverse process”. It is called “direction processing”. In this embodiment, in the forward processing, first, recognition processing based on the fixed word dictionary included in the forward dictionary 231 is performed, and if no recognition result is obtained, the fixed word included in the forward dictionary 231. It is assumed that recognition processing based on a dictionary other than the dictionary is performed. Similarly, in the backward processing, first, recognition processing based on the fixed word dictionary included in the backward dictionary 232 is performed, and if a recognition result is not obtained, a dictionary other than the fixed word dictionary included in the backward dictionary 232 is used. It is assumed that recognition processing based on is performed.

First, the main flow of the speech recognition process will be described with reference to FIG. The voice recognition process shown in FIG. 3 is a process performed when the talk switch 31 is turned on.
In the first step S100 (hereinafter, “step” is omitted and simply indicated by the symbol “S”), it is detected that the talk switch 31 is turned on.

In S200, the recognition dictionary 23 is set.
In S300, the voice uttered by the user input via the microphone 41 is A / D converted into waveform data that can be processed, and the waveform data is acquired as voice information. The acquired audio information is sequentially stored in the information storage unit 70.
In S400, it is specified whether the voice information acquired in S300 is to be forward-processed or reverse-processed, and recognition processing is performed.

The recognition process performed here is demonstrated based on the subflow shown in FIG. Note that the recognition process shown in FIG. 4 may be started at the same time as the acquisition of voice information is started in S300, or may be started after the acquisition of voice information is completed.
In S401 in FIG. 4, the voice information acquired in S300 is forward processed based on the forward dictionary 231. Note that the forward processing in S401 is limited to recognition processing based on the fixed word dictionary included in the forward dictionary 231, and recognition processing based on a dictionary other than the fixed word dictionary may not be performed.

  In S402, it is determined whether or not the first half of the acquired voice information is diverging. Details of the method for determining whether or not the first half of the acquired voice information is diverging will be described later. In the present embodiment, partial audio information that is audio information for a predetermined period (for example, 0.5 seconds) after acquisition of audio information is set as “first half portion”, and the remaining portion excluding the first half portion in audio information is “ Let's say the second half. When it is determined that the first half of the acquired voice information is diverging (S402: YES), the process proceeds to S405. When it is determined that the first half of the acquired voice information is not diverging (S402: NO), the process proceeds to S403.

In S403, it is estimated that the fixed word is in the first half of the voice information.
In S404, the speech information is forward processed based on the forward dictionary 231 and the result obtained by the forward processing is used as the recognition result. Then, the process proceeds to S500 in FIG.

When it is determined that the first half of the acquired voice information is diverging (S402: YES), it is estimated that the fixed word is in the second half.
In S <b> 406, the voice information is processed backward based on the backward dictionary 232. When audio information is reproduced in the reverse direction, based on the audio information stored in the information storage unit 70, the position where the amplitude of the waveform data is equal to or less than a predetermined value is a predetermined time or more. To be identified. Then, the audio information is reproduced in the reverse direction with the identified end position as the start position, and recognition processing is performed.

  In step S407, it is determined whether or not the voice information has been identified by the backward processing. When it is determined that the recognition result cannot be specified by the backward process (S407: NO), the process proceeds to S409. When it is determined that the recognition result can be specified by the backward process (S407: YES), the process proceeds to S408.

In S408, the result obtained by the backward process is set as the recognition result. Then, the process proceeds to S500 in FIG.
If it is determined that the recognition result could not be specified by the reverse process (S407: NO), the process proceeds to S500 in FIG. The case where there is no recognition result is exemplified when the volume of the acquired voice is low, or when the utterance content cannot be specified.

  In S500 in FIG. 3 that is shifted to after completion of the recognition process shown in FIG. 4, it is determined whether or not a recognition result has been obtained. If it is determined that the recognition result has not been obtained (S500: NO), that is, if a negative determination is made in S407 in FIG. 4, the processing of S600 is not performed and a voice such as “could not be recognized” is obtained via the speaker 51. Is output to the user, and information indicating that the recognition result has not been obtained is notified to the user, and the process is terminated. When it is determined that the recognition result has been obtained (S500: YES), the process proceeds to S600.

  In S600, the recognition result is output, and this process ends. That is, in S600 that moves after S404 in FIG. 4, the recognition result obtained by the forward processing is output. Further, in S600 that moves after S408 in FIG. 4, the recognition result obtained by the backward process is output.

By the way, in general speech recognition, a beam search that limits a search range is performed. In the beam search, a so-called “hypothesis pruning” is performed in which a predetermined beam width is set and a hypothesis that is outside the beam width range set based on the maximum likelihood hypothesis is rejected. For each hypothesis, a score, which is a value indicating the degree of coincidence with the voice information, is calculated. This score is a value calculated to identify the recognition result, and the higher the score, the higher the probability that it is a recognition result of speech information. The hypothesis corresponds to the “recognition target candidate”.
There are two methods for setting the beam width: a method of setting the number of hypotheses to be held, and a method of setting the score.

Here, two determination methods will be described as to whether or not the search result of the voice information is divergent.
<Divergent judgment method 1>
The divergence determination method 1 is a determination method when the beam width is set by setting the number of hypotheses to be held to n. When the recognition process is performed by reproducing the partial voice information in the forward direction, the hypothesis having the highest score and the hypothesis having the lowest score among the n hypotheses held at a specific time in the first half If the difference between the scores is within a predetermined threshold, that is, if the difference between the scores for each hypothesis is small, it is determined that the search processing for the first half of the speech information is diverging (S402: YES).

<Divergent judgment method 2>
The divergence determination method 2 is a determination method in a case where the beam width is set so as to hold a hypothesis whose score difference with the maximum likelihood hypothesis that is the hypothesis having the highest score is within a predetermined threshold T. When the recognition process is performed by reproducing the partial voice information in the forward direction, if there is a predetermined number or more of hypotheses whose score difference with the maximum likelihood hypothesis is within the threshold T at a certain point in the first half, When there are many hypotheses within a predetermined threshold T and the score difference for each hypothesis is small. It is determined that the search process for the first half of the voice information is diverging (S402: YES).

Here, a specific example of the speech recognition processing according to the present embodiment will be described. Hereinafter, an utterance content related to the reproduction of music by the artist “BBB” will be described as an example.
The utterance content related to music playback includes fixed words (in this example, “play”, “Play”, etc.) for instructing predetermined processing, and variable words (this is an object corresponding to fixed words such as artist name and song name) In the example, “BBB”).

(1) Specific example 1
Specific example 1 will be described with reference to FIG. 5 by taking as an example a case where the user utters “Play BBB” when playing the music of the artist “BBB”. In this case, as shown in FIG. 5A, the fixed word “Play” is uttered earlier in time series than the variable word “BBB”, and is included in the partial voice information. FIG. 5A and FIG. 6A to be described later are diagrams for explaining the structure of the utterance content uttered by the user.

  When it is detected that the talk switch 31 is turned on (S100), the recognition dictionary 23 is set (S200), and the voice information related to the uttered “Play BBB” is acquired (S300). When the voice information is acquired, first, the acquired voice information is forward-processed. In Specific Example 1, since “Play”, which is a fixed word, is included in the partial voice information, “Play” is specified based on the fixed word dictionary included in the forward dictionary 231, so that the first half of the voice information diverges. (S402: NO), it is estimated that the fixed word “Play” is in the first half of the voice information (S403).

  Therefore, as shown in FIG. 5B, by collating the waveform data of the voice information with the forward dictionary 231 in the forward direction along the time axis, that is, from the left to the right in FIG. 5B. Process forward. In specific example 1, since “Play” as a fixed word is specified, forward processing based on “artist name dictionary” and “song name dictionary” which are variable word dictionaries associated with “Play” is performed. Thus, it is relatively easy to identify the variable word “BBB”. Therefore, “Play BBB”, which is the result specified by performing forward processing on the audio information acquired in S300, is set as the recognition result (S404, S500: YES), and the recognition result is output (S600). In the first specific example, based on the recognition result, the process related to the reproduction of the song of the artist “BBB” is performed in a process different from this process.

(2) Specific example 2
Specific example 2 will be described with reference to FIG. 6, taking as an example a case where the user utters “play BBB” when playing the music of the artist “BBB”. In this case, as shown in FIG. 6A, it is assumed that the “reproduction” that is a fixed word is uttered later in time series than the variable word “BBB” and is not included in the partial voice information. In the second specific example, “reproduce” including a particle will be described as a fixed word.

  When it is detected that the talk switch 31 is turned on (S100), the recognition dictionary 23 is set (S200), and the voice information related to the “reproduce BBB” is acquired (S300). When the voice information is acquired, first, the acquired voice information is forward-processed. In the specific example 2, since the fixed word “reproduce” is not included in the partial voice information, the first half of the voice information diverges (S402: YES). At this time, when speech recognition processing is performed with reference to a dictionary other than the fixed word dictionary, the word is similar to “BBB” such as “DDD” corresponding to another fixed word such as “call to”. There is a risk of misunderstanding. If “BBB” is misrecognized as “DDD” and the hypothesis is pruned, “replay” following “BBB” cannot be recognized.

  In the present embodiment, as in the specific example 2, when the first half of the voice information diverges (S402: YES), it is estimated that the fixed word is in the second half of the voice information (S405). Therefore, as shown in FIG. 6B, by collating the waveform data of the voice information and the reverse dictionary 232 in the reverse direction along the time axis, that is, from the right to the left in FIG. 6B. Process in the reverse direction. In the second specific example, the fixed word “replay” is specified (S406). Therefore, the inverse based on the “artist name dictionary” or the “song name dictionary” that is the variable word dictionary associated with “replay”. By performing the direction processing, the variable word “BBB” can be identified relatively easily (S407: YES). Accordingly, “reproduce BBB”, which is the result specified by performing backward processing on the audio information acquired in S300, is set as the recognition result (S408, S500: YES), and the recognition result is output (S600). In the second specific example, as in the first specific example, based on the recognition result, the process related to the reproduction of the song of the artist “BBB” is performed in a process different from the present process.

  As described above in detail, the speech recognition apparatus 1 reproduces the forward dictionary 231 in which the speech data and the word when the word is reproduced in the forward direction are stored in association with each other and the word in the backward direction. A recognition dictionary 23 including a backward dictionary 232 in which voice data and words are stored in association with each other. In the present embodiment, speech information including a fixed word that forms the basis of a tree structure and variable words that correspond to the fixed word and form branches and leaves of the tree structure is acquired (S300), and recognition processing is performed. First, recognition processing is performed in the forward direction by reproducing the speech information in the forward direction based on the forward dictionary 231 (S401), and the position of the fixed word in the speech information is estimated based on the recognition result obtained in the forward direction. (S403, S405). When it is estimated that the fixed word is in the second half of the voice information (S405), the recognition process is performed in the reverse direction by reproducing the voice information in the reverse direction based on the reverse dictionary 232 (S406). Accordingly, the voice information recognition process can be appropriately switched between the forward direction and the reverse direction based on the position of the fixed word of the voice uttered by the user. Improves the accuracy of speech recognition without making you conscious.

  In the present embodiment, based on the recognition result performed in the forward direction, it is determined whether or not the first half of the voice information is diverging, and if it is determined that the first half is diverging (S402: YES), the fixed It is estimated that the word is in the latter half of the voice information (S405). Thereby, it is possible to reduce the calculation load related to the estimation of the position of the fixed word.

  Further, in the present embodiment, when the voice information recognition process is started and the partial voice information that is the voice information for a predetermined period is recognized in the forward direction and the recognition result of the partial voice information cannot be obtained, the voice information It is determined that the first half part is diverging (S402: YES). Accordingly, it is possible to determine whether or not the first half of the voice information is diverging without performing the entire recognition process of the voice information, so that the calculation load can be further reduced. Further, based on the recognition result of the partial voice information, when it is determined that the first half of the voice information is divergent, it is possible to quickly switch to the reverse process in which the voice information is reproduced in the reverse direction and the recognition process is performed. . Thereby, compared with the case where it switches to reverse direction processing after carrying out forward direction processing of all the audio | voice information, the time until a recognition result is acquired after audio | voice information is acquired can be shortened.

The determination as to whether or not the first half of the voice information is diverging can be made as follows.
In a plurality of recognition target candidates having a score for specifying a recognition result of partial speech information, if the difference between the score of the recognition target candidate with the highest score and the recognition target candidate with the lowest score is within a predetermined threshold, It is determined that the voice information recognition result is not obtained and the first half of the voice information is divergent (S402: YES).

  In addition, among a plurality of recognition target candidates having a score for specifying a recognition result of partial speech information, there are a predetermined number or more of recognition target candidates whose difference in score from a recognition target candidate having the highest score is within a predetermined threshold. In this case, the recognition result of the partial voice information may not be obtained, and it may be determined that the first half of the voice information is divergent (S402: YES).

  As described in the divergence determination methods 1 and 2, based on the score for identifying the recognition result of the plurality of recognition candidates, when the variation in the score between the plurality of recognition candidates is small, the recognition result of the voice information is diverged. Judge that Thereby, it can be determined appropriately whether the recognition result of audio | voice information is diverging.

  In this embodiment, the control unit 10 constitutes “voice information acquisition means”, “forward direction recognition processing means”, “fixed word position estimation means”, “reverse direction recognition processing means”, and “divergence determination means”. To do. Also, S300 in FIG. 3 corresponds to processing as a function of “voice information acquisition means”, S401 in FIG. 4 corresponds to processing as a function of “forward recognition processing means”, and S403 and S405 are “ The process corresponds to a process as a function of “fixed word position estimating means”, S406 corresponds to a process as a function of “reverse direction recognition processing means”, and S402 corresponds to a process as a function of “divergence determination means”.

As mentioned above, this invention is not limited to the said embodiment at all, In the range which does not deviate from the meaning of invention, it can implement with a various form.
(A) Fixed word position estimation means In the above embodiment, if the recognition result of the first half of the speech information is not obtained and it is determined that the speech is diverging (S402: YES), the fixed word is in the second half of the speech information. Presumed to be.

  By the way, in the voice recognition apparatus, there is a case where an operation language related to the operation of the apparatus can be selected. For example, when Japanese is selected as the operation language, various information provided via the speaker 51 and the display 61 is output in Japanese. For example, when English is selected as the operation language, various information provided via the speaker 51 or the display 61 is output in English.

Here, the relationship between the positional relationship between the fixed word and the variable word and the operation language will be described with reference to FIGS.
It can be said that the relationship between the fixed word instructing the predetermined processing and the variable word that is the target word of the fixed word has a tree structure with the fixed word as a root and the variable word as a branch leaf. For example, in a language such as English, which is in the order of words of fixed words and variable words, as shown in FIG. 5A, it can be said that the forward tree structure is preceded by a fixed word that is the basis of the tree structure. Further, for example, in a language in which variable words and fixed words are arranged in the order of words such as Japanese, as shown in FIG. 6A, it can be said that the tree has a backward-facing tree structure followed by a fixed word that is the root of the tree structure. .

  When an operation language having a characteristic that a fixed word is spoken before a variable word, for example, English, is selected, the user assumes that the spoken word is spoken in English, and estimates that the fixed word is in the first half of the voice information. Note that although English has been described here as an example of an operation language in which the word order is “fixed word” and “variable word”, if the operation language has the characteristics that the word order is “fixed word” and “variable word”, then the operation language Not exclusively.

  In addition, when an operation language having a characteristic that a fixed word is uttered after a variable word, for example, Japanese, is selected, the user is assumed to speak in Japanese, and the fixed word is estimated to be in the latter half of the voice information. To do. In this example, Japanese is described as an example of an operation language in which the word order is “variable word” or “fixed word”. However, if the operation language has a characteristic in which the word order is “variable word” or “fixed word”, Japan It is not limited to words.

  Thus, the positional relationship between fixed words and variable words is determined to some extent by the characteristics of the language. Therefore, in a modified example, the position of the fixed word may be estimated based on the selected operation language. That is, if it is considered that the user speaks in the language selected as the operation language, it can be said that the user's speech language is estimated based on the selected operation language. Then, based on the set operation language, switching between performing the recognition process in the forward direction or the reverse direction is based on whether the recognition process is performed in the forward direction or based on the user's utterance language indirectly. It can be said that switching is performed.

  Even in this configuration, the language spoken by the user can be estimated based on the set operation language, and the position of the fixed word can be estimated appropriately, so that the recognition process is performed in accordance with the voice uttered by the user. Since it is possible to appropriately switch between performing in the direction and in the reverse direction, the accuracy of voice recognition is improved without making the user aware of the operation related to the recognition processing direction.

(A) Divergence determination means The determination as to whether or not the first half of the voice information is divergent may be performed as follows.
When the voice information is acquired, based on the forward dictionary, the entire acquired voice information is reproduced in the forward direction to perform recognition processing, and a temporary recognition result is acquired. In the temporary recognition result, the reliability is calculated for each identified word. The “reliability” is a scale indicating how much the identified word can be trusted as a recognition result, and can take a value in the range of 0-1. When the reliability value of the identified word is close to 1, it indicates that there is almost no other hypothesis having a score comparable to that word, and it can be said that the probability that the word is reliable as a recognition result is high. On the other hand, when the reliability of the identified word is close to 0, it indicates that there are many other hypotheses having scores similar to that of the word, and it can be said that the probability of being reliable as the recognition result of the word is low.

Here, a specific example of estimating the position of the fixed word based on the reliability will be described with reference to FIG. In FIG. 7, an example in which “Speak to PPP” is spoken will be described.
When the user utters “call PPP”, the acquired voice information is recognized in the forward direction, and the reliability for each word included in the obtained temporary recognition result is calculated. As shown in FIG. 7A, the reliability of “PPP” in the first half of the voice information is as low as 0.13. On the other hand, the reliability of “calling” in the second half of the voice information is 0.82, which is higher than the reliability of “PPP” in the first half. Here, if it is sufficient to trust a word having a reliability of 0.5 or more as a recognition result, the word “PPP” in the first half of the voice information has a reliability of a predetermined value of 0.5 or less. The reliability of the word “call phone” in the latter half of the voice information is a predetermined value of 0.5 or more, and it can be determined that the first half of the voice information is diverging (S402: YES). At this time, since the reliability of “PPP” in the first half of the voice information is low, “PPP” may be erroneously recognized. If “PPP” is misrecognized and the hypothesis is pruned, it becomes impossible to recognize “call to” following “PPP”.

  Therefore, as shown in FIG. 7B, the voice information is reproduced in the reverse direction and the recognition process is performed based on the reverse dictionary, whereby the fixed word “call” can be specified. . Then, by setting a dictionary stored in association with “call to” and performing “PPP” recognition processing, the accuracy of the recognition processing is improved.

  Thus, based on the recognition result obtained by performing the recognition process in the forward direction, the reliability for each word specified as the recognition result is calculated, and the reliability of the word in the first half of the voice information is equal to or less than a predetermined value, If the reliability of the word in the second half of the speech information is greater than or equal to a predetermined value, it may be determined that the first half of the speech information is divergent and the fixed word is in the second half of the speech information. This makes it possible to appropriately determine whether or not the first half part is diverging based on the reliability of the word specified as the recognition result.

  In the forward recognition process, if the reliability of “call to” the fixed word in the second half is greater than or equal to a predetermined value, the “call to” fixed word is in the second half of the voice information. It may be estimated. In other words, “the fixed word position specifying means calculates the reliability of each word specified as the recognition result based on the recognition result obtained by performing the recognition process in the forward direction, and the reliability of the word in the latter half of the speech information. Is greater than or equal to a predetermined value, it is estimated that the fixed word is in the second half of the speech information ”. In this case, the backward processing of the voice information of the portion corresponding to “call to” based on the fixed word dictionary included in the backward dictionary 232 may be omitted. Then, a dictionary corresponding to “call to” included in the reverse dictionary 232 is set, and the waveform data corresponding to “call to” is reproduced in the reverse direction from the start point in the forward direction to perform recognition processing. You may go.

(C) Audio information In the above embodiment, partial audio information, which is audio information for a predetermined period (for example, 0.5 seconds) from the start of acquisition of audio information, is defined as “first half part”, and the first half part of the audio information is excluded. The remaining part was designated as the “second half part”. In the modification, the start position and the end position of the voice information are specified, and the predetermined ratio (for example, 1/2) of the first half of the total recognition time of the voice information is set as the “first half part” of the voice information, and the remaining part is the “second half part”. It is good. Further, a predetermined ratio (for example, ½) of the latter half of the total recognition time of the voice information may be set as the “second half” of the voice information, and the remaining may be set as the “first half”.

  Further, in the above embodiment, in the speech information, there is one fixed word that forms the basis of the tree structure and one variable word that corresponds to the fixed word and forms the branches and leaves of the tree structure. There may be a plurality of each. When the acquired voice information has a tree structure of n layers, the words from the root to the (n-1) layer may be fixed words. Similarly, when the acquired voice information has a tree structure of n layers, variable words from the end to the (n−1) layer may be used.

DESCRIPTION OF SYMBOLS 1 ... Voice recognition apparatus 10 ... Control part (Voice information acquisition means, forward direction recognition processing means, fixed word position estimation means, reverse direction recognition processing means, divergence judgment means)
DESCRIPTION OF SYMBOLS 20 ... Audio | voice information storage part 21 ... Audio | voice information database 22 ... Acoustic model 221 ... Forward model 222 ... Reverse model 23 ... Recognition dictionary 231 ... Forward dictionary (forward) Dictionary data of tree structure)
232 ... Reverse dictionary (dictionary data of backward-facing tree structure)
24 ... Language model 30 ... Operation switch group 31 ... Talk switch 40 ... Audio input unit 41 ... Microphone 50 ... Audio output unit 51 ... Speaker 60 ... Drawing unit 61 ... Display 70 ... Information storage

Claims (7)

Speech data when a word is reproduced in the forward direction and the word are stored in association with the dictionary data of the forward tree structure, and voice data when the word is reproduced in the reverse direction and the word are stored in association with each other A speech recognition apparatus having dictionary data of a backward-facing tree structure,
Speech information acquisition means for acquiring speech information consisting of fixed words that form the basis of a tree structure, and variable words that correspond to the fixed words and form branches and leaves of the tree structure;
Forward recognition processing means for performing recognition processing in the forward direction by reproducing the voice information in the forward direction based on the dictionary data of the forward tree structure;
Fixed word position estimating means for estimating the position of the fixed word in the speech information based on the recognition result of the forward direction recognition process;
When the fixed word position estimating unit estimates that the fixed word is in the latter half of the speech information, the speech information is reproduced in the reverse direction based on the backward tree structure dictionary data to reverse the recognition process. Reverse direction recognition processing means for performing the direction;
A speech recognition apparatus comprising: Based on the recognition result of the forward direction recognition processing means, comprising divergence determination means for determining whether or not the first half of the voice information is divergence,
The fixed word position estimating means estimates that the fixed word is in the latter half of the voice information when the divergence judging means determines that the first half of the voice information is diverging. The speech recognition apparatus according to claim 1.   The divergence determining means recognizes the partial voice information as the voice information for a predetermined period after the voice information acquisition means starts to acquire the voice information, and the recognition result of the partial voice information cannot be obtained. The speech recognition apparatus according to claim 2, wherein it is determined that the first half of the speech information is diverging.   In the plurality of recognition target candidates having a score for specifying the recognition result of the partial speech information, the divergence determination unit is configured to determine the recognition target candidate having the highest score and the recognition target candidate having the lowest score. 4. The voice according to claim 3, wherein when the difference in scores is within a predetermined threshold, the recognition result of the partial voice information is not obtained, and it is determined that the first half of the voice information is diverging. Recognition device.   The divergence determination means includes a plurality of recognition target candidates having a score for specifying a recognition result of the partial speech information, and the difference in the score from the recognition target candidate having the highest score is within a predetermined threshold. 4. The voice according to claim 3, wherein when a plurality of recognition target candidates exist, a recognition result of the partial voice information is not obtained, and it is determined that the first half of the voice information is divergent. Recognition device.   The divergence determining means calculates the reliability for each word specified as the recognition result based on the recognition result by the forward direction recognition processing means, and the reliability of the word in the first half of the speech information is less than a predetermined value 3. The method according to claim 2, wherein if the reliability of a word in the second half of the voice information is equal to or greater than the predetermined value, it is determined that the first half of the voice information is divergent. Voice recognition device. Speech information acquisition means for acquiring speech information consisting of fixed words forming the basis of a tree structure, and variable words corresponding to the fixed words and forming branches and leaves of the tree structure;
Forward direction recognition processing means for performing recognition processing in the forward direction by reproducing the voice information in the forward direction based on dictionary data of a forward tree structure;
Fixed word position estimating means for estimating the position of the fixed word in the speech information based on the recognition result of the forward direction recognition process;
When the fixed word position estimating means estimates that the fixed word is in the latter half of the speech information, the speech information is reproduced in the reverse direction based on the backward-facing tree structure dictionary data. Reverse direction recognition processing means for performing in the reverse direction;
As a program that allows the computer to function.

Images