JP2010072081A - Voice recognition dictionary creating device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide a voice recognition dictionary creating device shortening a period from the start of creation of a tree-structure voice recognition dictionary to the start of voice recognition. <P>SOLUTION: When an instruction shows initial registration (initial registration in S110), input of a music data group is received (S120), and data are sorted according to sort priority and divided into a plurality of sets (S130), and from the set including music data of the higher sorting rank, voice recognition dictionaries are created successively (S140). When it is determined that the instruction shows addition of music data (addition in S110), input of music data is received (S145), and a dictionary only for additional data is created (S150). When it is determined that the instruction shows deletion of music data (deletion in S110), input of a music name is received (S155), and a non-recognition-target flag is raised in correspondence with the music name of a deletion target (S160). When it is determined that the instruction shows change of music data (change in S110), processing of a combination of addition and deletion is carried out (S165-S180). <P>COPYRIGHT: (C)2010,JPO&INPIT

Description

  The present invention relates to an apparatus for creating a dictionary for speech recognition.

A technique for reconstructing a speech recognition dictionary used for speech recognition as the number of words to be recognized increases has already been known (Patent Document 1).
This type of speech recognition dictionary has a data structure in which text data corresponding to words (phrases) to be recognized is associated with speech data representing features of speech corresponding to the text data. The voice data serves as an interface between a voice signal representing a user's voice input through a microphone and text data.

The speech recognition dictionary is generally configured in a tree structure. The reason for this is to improve the efficiency of speech recognition by sharing the overlapping portions of speech data in the manner of creating a tree diagram.
JP 2001-249686 A

  The problem with the above-described technique is that it takes a long time from the start of creating a speech recognition dictionary that employs a tree structure to the start of speech recognition. This is because a dictionary for speech recognition that adopts a tree structure cannot be used for speech recognition unless it is completely completed, but it takes time to construct the tree structure. As a result, voice recognition may not be performed for a long time, causing stress to the user.

  SUMMARY OF THE INVENTION In view of the above-described problems, an object of the present invention is to provide a speech recognition dictionary creation device that can shorten the time from the start of creating a speech recognition dictionary that employs a tree structure until speech recognition can be started.

  The speech recognition dictionary creating apparatus according to claim 1 invented to solve the above-described problem includes an input unit, a dividing unit, and a creating unit. The input means receives a phrase group input. The dividing means divides the word group into a plurality of sets. Then, the creating means creates one speech recognition dictionary having a tree structure for each set.

  According to the speech recognition dictionary creating apparatus of the first aspect, it is possible to shorten the time from the start of newly creating the speech recognition dictionary until the speech recognition can be started. This is because, even if a dictionary is not created for all input words / phrases, dictionaries completed separately for the input word / phrase group can be sequentially used for speech recognition.

  According to a second aspect of the present invention, there is provided a voice recognition dictionary creating apparatus including an updating unit in addition to the configuration of the first aspect. This updating means creates a single speech recognition dictionary by integrating a plurality of speech recognition dictionaries created by the creation means by constructing a new tree structure.

  According to the speech recognition dictionary creating apparatus of the second aspect, the effect of improving the efficiency of speech recognition by the tree structure can be further enhanced by combining the plurality of dictionaries into one. This is because if the tree structure is combined into one, more overlapping parts can be integrated.

  On the other hand, the speech recognition dictionary creating apparatus according to claim 3 is also an invention for solving the above-described problem, and includes an input unit and a creating unit. The input means receives a phrase. Then, the creating means creates a speech recognition dictionary for the phrase separately from the existing speech recognition dictionary having a tree structure.

  According to the speech recognition dictionary creating apparatus of the third aspect, it is possible to shorten the time from the start of adding words / phrases that are not recorded in the existing speech recognition dictionary to the start of speech recognition. This is because the existing speech recognition dictionary is not modified, and speech recognition using this dictionary is possible at any time. Furthermore, if a speech recognition dictionary is constructed only for the input word / phrase, the voice recognition using this dictionary can be started.

  According to a fourth aspect of the present invention, there is provided a speech recognition dictionary creating apparatus including an updating unit in addition to the configuration of the third aspect. The updating means integrates the speech recognition dictionary created by the creation means and the existing speech recognition dictionary by constructing a new tree structure, thereby creating one speech recognition dictionary.

  According to the speech recognition dictionary creating apparatus of the fourth aspect, the effect of improving the efficiency of speech recognition by the tree structure can be further enhanced by combining the plurality of dictionaries into one. This is because if the tree structure is combined into one, more overlapping parts can be integrated.

  On the other hand, the speech recognition dictionary creating apparatus according to claim 5 is also an invention for solving the above-described problem, and includes an input unit and a mark addition unit. The input means receives information indicating a word / phrase that is not subject to voice recognition among words / phrases recorded in an existing voice recognition dictionary having a tree structure. The mark adding means adds a mark corresponding to the word or phrase indicated by the information.

  According to the speech recognition dictionary creating apparatus of the fifth aspect, it is possible to shorten the time from the start of deleting a phrase from the existing speech recognition dictionary until the speech recognition can be started. This is because a mark is only added so as not to output the voice recognition result without correcting the tree structure of the existing voice recognition dictionary. Since the word / phrase to which the mark is added in this way is not output as a speech recognition result, the user is in a state equivalent to being deleted.

  In addition to the configuration of the fifth aspect, the speech recognition dictionary creating device according to the sixth aspect includes an updating unit. The updating means creates a speech recognition dictionary by removing a word / phrase with a mark from an existing speech recognition dictionary by constructing a new tree structure.

  According to the speech recognition dictionary creating apparatus of the sixth aspect, the effect of improving the efficiency of speech recognition by the tree structure can be further enhanced by actually erasing the phrase to which the mark has been added. This is because if a phrase that is not output as a speech recognition result is recorded in the dictionary, a process for eliminating the phrase in the speech recognition process is required.

  On the other hand, the speech recognition dictionary creating apparatus according to claim 7 is also an invention for solving the above-described problem, and includes an input unit, a mark adding unit, and a creating unit. The input means receives input indicating that a predetermined word / phrase recorded in an existing speech recognition dictionary having a tree structure is changed to another word / phrase. The mark adding means adds a mark corresponding to the predetermined word / phrase. Then, the creating means creates a speech recognition dictionary for the other words / phrases separately from the existing speech recognition dictionary.

  According to the speech recognition dictionary creating apparatus of the seventh aspect, it is possible to shorten the time from the start of changing a word / phrase in the existing speech recognition dictionary to the start of speech recognition. This is because the change is divided into deletion and addition, and the configurations of claims 3 and 5 are combined. In other words, for existing speech recognition dictionaries, the tree structure is not modified and only a mark is added to the word before change so that it is not output as the speech recognition result. Furthermore, for the word after change, This is because if a speech recognition dictionary only for words is constructed, speech recognition using this dictionary can be started.

  In addition to the configuration of the seventh aspect, the speech recognition dictionary creating apparatus according to the eighth aspect includes an updating unit. This updating unit integrates the speech recognition dictionary created by the creating unit and the speech recognition dictionary excluding words / phrases added with a mark from the existing speech recognition dictionary by constructing a new tree structure. Create two speech recognition dictionaries.

  According to the speech recognition dictionary creation device of claim 8, the efficiency of speech recognition based on a tree structure can be obtained by erasing the word / phrase to which the mark has been added and combining the divided dictionary into one. The effect of improvement can be further enhanced. Because, if the tree structure is combined into one, more overlapping parts can be integrated.Furthermore, if words and phrases that are not output as speech recognition results are recorded in the dictionary, This is because processing that eliminates the phrase is necessary.

  Hereinafter, it demonstrates with drawing. FIG. 1 is a schematic configuration diagram of a navigation device 10 to which the present invention is applied and a portable digital music player 100 (hereinafter referred to as “portable player 100”) as an external device. As shown in FIG. 1, the navigation device 10 includes a positioning device 11, an operation switch group 12, an audio input unit 13, a display unit 14, an audio output unit 15, an HDD (hard disk drive) 17, an external connection interface 18, and a control unit 20. Prepare.

  The positioning device 11 receives a radio wave from a GPS satellite via a GPS antenna, and measures the magnitude of the rotational motion applied to the vehicle, and a GPS receiver 11a that acquires the orbit information and current date / time information of the satellite. And a distance sensor 11c that measures the travel distance of the vehicle.

  The operation switch group 12 includes a touch panel configured integrally with the display unit 14 and mechanical key switches provided around the display unit 14. The voice input unit 13 is configured with a microphone so as to acquire voice uttered by the user.

  The HDD 17 stores map data, music data, and the like and inputs them to the control unit 20. The external connection interface 18 is for data communication with an external device. For example, it is used for connection with a portable player 100 or the like that can store / reproduce music data for the purpose of transmitting / receiving music data. Music data received through the external connection interface 18 is stored in the HDD 17 via the control unit 20.

  The display unit 14 is a color display device and includes a liquid crystal monitor or the like. The display unit 14 displays a map around the current position of the vehicle, a route to the destination designated by the user, and the like based on a control signal from the control unit 20. The voice output unit 15 is a speaker or the like, and outputs a voice for route guidance based on a control signal from the control unit 20.

  The control unit 20 includes a CPU 20a, a RAM 20b, a ROM 20c, an NVRAM 20d, and the like. The control unit 20 executes various programs stored in the ROM 20c by the CPU 20a, thereby realizing functions such as route guidance, voice recognition, and music reproduction.

  Incidentally, the navigation device 10 executes the following processing as the voice recognition function and the music playback function. The process is as follows. When a song name of music data stored in the HDD 17 is inputted by voice through the voice input unit 13, the song name is voice-recognized by a voice recognition dictionary, and music corresponding to the recognized song name is recorded. This is a process of reproducing data.

  Before explaining the processing, it will be described how the music titles and music data are associated with each other using FIG. FIG. 2 shows a music data group stored in the portable player 100. This music data group is input from the portable player 100 to the navigation device 10 in a dictionary construction process described later.

  The music data group is a collection of music data. The song data is data for each song, and the notation and reading of the song name, the album name as the song information, the hit ranking and the number of times of reproduction, and the music data for reproduction (not shown). Each item is stored. With such a music data structure, music names and music data are associated with each other.

  Further, sort priority information is stored in the music data group. This sort priority is used when the music data group is divided into a plurality of sets in S130 described later. Specifically, the sort priority is information indicating the order in which items are prioritized to sort data, and is determined by user input.

  In the example of FIG. 2, since the number of reproductions is the sort priority number 1, the higher the number of reproductions, the higher the sorting result. Those with the same number of reproductions are arranged in descending order of the second highest hit ranking (ranking such as the number sold in the market). The same applies to the sort priority 3 and below.

  Next, the dictionary construction process will be described with reference to FIG. The dictionary construction process shown in the flowchart in FIG. 3 is a process executed mainly by the control unit 20 in order to construct a speech recognition dictionary. Note that the speech recognition dictionary constructed by this process is for specifying the song name designated by the user's voice. The purpose of specifying the song name is to reproduce music data corresponding to the specified song name, as described above.

  The trigger for starting this process is that the initial registration, addition, deletion or change of music data is instructed through the voice input unit 13 (microphone). It should be noted that a voice recognition dictionary for voice recognition for “initial registration”, “addition”, “deletion”, and “change”, as well as words and phrases necessary for executing the processing (S230) described later, is constructed by dictionary construction processing. It is stored separately from the dictionary.

  When the dictionary construction process is started, it is determined what is the user instruction that triggered the start of the process (S110). When it is determined that the user instruction is initial registration (initial registration in S110), the music data group input described with reference to FIG. 2 is received from the portable player 100 via the external connection interface 18 (S120).

  Note that the portable player 100 specifies which music data is included in the music data group to be added here. Further, the number of music data included in the music data group may be one or any number of songs. These two points are the same when adding later.

  Next, according to the sorting priority, each piece of music data is sorted as described above, and this sort result is used to divide into a set composed of a predetermined number of pieces of music data (S130) (see FIG. 4). For example, if data of 950 songs is input and the upper limit of the number of songs belonging to each set after division is set to 100 songs, the sort results are 1 to 100, 101 to 200,..., 901 to 950. Specifically, the fraction is treated as one set and divided into 10 sets.

  Next, a voice recognition dictionary having a tree structure is created and registered as a temporary dictionary one by one from the set including the music data with the higher sorting results (smaller numbers) (S140), and the process is terminated. In other words, a tree structure is created by replacing each song name reading (text data) with voice data so that it can be used for voice recognition (see FIG. 5).

  This voice data is obtained by digitally expressing the characteristics of a voice waveform corresponding to text data. Note that the term “provisional” dictionary is used here because it is assumed that the dictionary will be updated by a later update process.

  The tree structure is a hierarchical structure in which one element has branch information to a plurality of elements as shown in FIG. That is, the overlapping part is made common to the voice data in the manner of creating a tree diagram. Here, an example is shown in which a tree structure is created based on the Japanese syllabary order, but other orders may naturally be used. Further, as shown in FIG. 5, a non-recognition flag can be set in this temporary dictionary. Briefly describing the non-recognition flag, the corresponding speech data is forbidden from being output as a speech recognition result (details are described in S160 and FIG. 6).

  On the other hand, if it is determined that the instruction from the user is addition of music data (added in S110), input of a music data group is accepted from the portable player 100 via the external connection interface 18 (S145).

  Then, a temporary dictionary for only the song name of the added song data is created and registered (S150), and this process is finished. At this time, as in the initial registration, it may or may not be divided every 100 songs based on the sorting result.

  In other words, adding here means adding new song names that are not included in these existing voice recognition dictionaries, assuming that other voice recognition dictionaries have already been registered, both formal and temporary. It is an instruction to do. However, as described above, in this process, the process for reconstructing the existing speech recognition dictionary is not performed, and a temporary dictionary having a tree structure is created and registered separately from the existing speech recognition dictionary.

  If it is determined that the instruction from the user is to delete the music data stored in the HDD 17 (deleted in S110), the input of the name of the music to be deleted is accepted through the voice input unit 13, and the music name is received. Is recognized (S155). Next, without actually deleting the music data corresponding to the music name, a non-recognition flag is set in correspondence with the music name in the voice recognition dictionary (S160), and the process is terminated.

  As shown in FIG. 5, the speech recognition dictionary regardless of whether it is formal or provisional is configured such that a recognition-in-progress flag can be set for the speech data of each song. The audio data of the music for which this flag is set is handled as a non-target at the time of voice recognition. The non-recognition target flag will be described with reference to FIG.

  FIG. 6 is a flowchart showing music playback processing. This process is a process executed mainly by the control unit 20 and is started when a song name is input through the voice input unit 13.

  First, the degree of coincidence with the input song name is ranked for the song names of each song recorded in all registered speech recognition dictionaries including formal and provisional (S210). Then, among those that have already been read out and those that have not been recognized flag set, those having the highest rank are read out through the voice output unit 15 (S220). Note that “already read” occurs when S220 is executed every time it is determined No in S230.

  Then, it is determined using speech recognition whether or not an input to the effect that the read song name matches that spoken by the user is made through the speech input unit 13 (S230). If it is determined that the input that the read song name does not match the one spoken by the user is made through the voice input unit 13 (No in S230), the process returns to S220.

  On the other hand, if it is determined that the input that the read song name matches the one spoken by the user is made through the voice input unit 13 (Yes in S230), the music data included in the song data specified by the read song name Is reproduced through the audio output unit 15 (S240), and this process ends.

As described above, the music name for which the flag to be excluded from recognition is set is not output as a voice recognition result. Therefore, it is the same as deleted for the user.
Returning to FIG. If it is determined that the instruction from the user is a change of the song name corresponding to the song data stored in the HDD 17 (changed in S110), the input of the song name before and after the change is accepted through the voice input unit 13, Are recognized by voice (S165).

  Then, a temporary dictionary only for the changed music data is created and registered (S170). Then, in the speech recognition dictionary, a flag not to be recognized is set in correspondence with the music name before the change (S180), and this process is finished. In other words, changes are performed separately in addition and deletion.

  Next, the update process will be described with reference to FIG. The update process is a process executed mainly by the control unit 20. This process is intended to create a dictionary in a desirable state for use in speech recognition by organizing and integrating dictionaries that have been divided into a plurality of segments or flagged as non-recognized as a result of the dictionary construction process. is there.

  The “desirable state” will be described in detail. If there are a large number of dictionaries, the effect of the tree structure for improving the efficiency of speech recognition is reduced. In addition, if the song name with the flag not to be recognized comes to the top of the recognition result, an extra process such as “It has come to the top but will not be output as a recognition result” is necessary. It is better not to stand. Therefore, by this update process, it is updated to one official dictionary for which no recognition target flag is set.

  Note that the dictionary that is the target of the update process is only a dictionary constructed by the dictionary construction process (that is, a dictionary for voice recognition of music titles), and a dictionary that is necessary for determining user instructions in the dictionary construction process Etc. is not a target.

  Further, the trigger for starting the process is that the processing load of the CPU 20a included in the control unit 20 is equal to or less than a threshold value. In other words, it is executed when the processing capacity of the control unit 20 is excessive and does not interfere with other processes. However, it is not executed when there is only one dictionary and no recognition target flag is set.

  First, the music data corresponding to the music name for which the non-recognized flag is set in all dictionaries including formal and provisional are deleted (S310). Then, a new dictionary is created by creating a tree structure for all the music names included in the music data remaining without being deleted (S320). Then, in place of the official dictionary registered at the present time, the created new dictionary is registered as an official dictionary (S330). Finally, the temporary dictionary is deleted (S340), and the process ends.

  The effect will be described here. In the initial registration of the music data group, there is an effect that voice recognition can be executed in a short time by the provisional dictionary without creating a formal dictionary that tends to have an enormous amount of data. For example, even if a dictionary for all 1000 songs is not created, when a dictionary containing 100 songs is registered for the time being, speech recognition can be executed for these 100 songs.

  In addition, since these 100 songs have the highest sort result, if the sort priority is selected so that the sort result is higher in the order that the user wants to play it sooner, playback is commanded from these 100 songs. There is a high possibility that the voice to be input will be input.

  As described in the background art, a dictionary that is being created cannot be used for speech recognition. As described above, at the time of voice recognition, the degree of coincidence between the voice data corresponding to each piece of music data and the input voice is ranked. Therefore, it cannot be used for speech recognition in a state where a dictionary is being created, that is, a tree structure is being constructed. Such a feature causes a problem of the prior art. In the present embodiment, this problem is solved by dividing the data for which a dictionary is to be created.

  Also, regarding addition, the tree structure of the formal dictionary does not have to be immediately reconstructed in order to add a small number of music data. Accordingly, speech recognition can be executed at any time for the formal dictionary and as soon as the temporary dictionary is created. Similarly, with regard to deletion, speech recognition can be resumed by performing a process that only sets a flag. Since the change is a process in which addition and deletion are combined, the same effect can be obtained.

As described above, by using the temporary dictionary and the non-recognition flag, the speech recognition can be resumed in a short time, and the time for waiting the user is shortened.
In addition, the update process updates the formal dictionary with no recognition target flag set at an appropriate timing. By this processing, voice recognition can be performed efficiently by using a single tree structure, and further, processing due to the flag not to be recognized being set is unnecessary.

  The embodiment of the present invention is not limited to the above-described example. For example, it may be applied to a mobile phone with a music playback function or a mobile player 100 instead of the navigation device 10. In addition, when applied to a mobile phone, a use of using a speech recognition dictionary to recognize a registered name such as a person name and identifying a telephone number from the recognized registered name is conceivable.

  The music data group may be divided in any way. For example, the number of pieces of music data included in the set after the division may be a natural number regardless of the number of songs. In addition, the number of sets formed by division may be determined. Further, the music data may be sorted in any way when the music data group is divided. For example, the order may be input or random.

  Finally, the correspondence between the claims and the examples will be described. The input means is realized by S120, S145, S155 and S165, the dividing means is S130, the creation means is S140, S150 and S170, the mark adding means is S160 and S180, and the updating means is realized by update processing.

  By the way, at the time of filing, there are a plurality of independent claims. On the other hand, in the embodiment, the invention of each independent term is realized by one apparatus. Accordingly, the invention corresponding to any combination of the independent claims is within the scope of the description of the present application, and even if added by amendment, no new matter is added.

The figure which shows schematic structure of a navigation apparatus. The table which shows the data structure of a music data group. The flowchart showing a dictionary construction process. The table which shows the structure of the divided music data. The figure showing a tree structure. The flowchart showing a music reproduction process. The flowchart showing an update process.

Explanation of symbols

DESCRIPTION OF SYMBOLS 10 ... Navigation apparatus, 11 ... Positioning device, 11a ... GPS receiver, 11b ... Gyroscope, 11c ... Distance sensor, 12 ... Operation switch group, 13 ... Voice input part, 14 ... Display part, 15 ... Voice output part, 17 ... HDD, 18 ... external connection interface, 20 ... control unit, 20a ... CPU, 20b ... RAM, 20c ... ROM, 20d ... NVRAM, 100 ... portable player

Claims (8)

An input means for receiving an input of words and phrases,
Dividing means for dividing the phrase group received by the input means into a plurality of sets;
A speech recognition dictionary creation device comprising: creation means for creating a speech recognition dictionary having a tree structure for each set divided by the dividing means. The update means for creating one speech recognition dictionary by integrating a plurality of speech recognition dictionaries created by the creation means by constructing a new tree structure. Voice recognition dictionary creation device. An input means for receiving an input of a phrase;
A speech recognition dictionary creation device comprising: creation means for creating a speech recognition dictionary for a word received by the input means separately from an existing speech recognition dictionary having a tree structure. The speech recognition dictionary created by the creation means and the existing speech recognition dictionary are integrated by constructing a new tree structure to provide an update means for creating one speech recognition dictionary. The dictionary creation apparatus for speech recognition according to claim 3. An input means for receiving input of information indicating words to be excluded from speech recognition among words recorded in an existing speech recognition dictionary having a tree structure;
A speech recognition dictionary creating apparatus comprising: mark adding means for adding a mark corresponding to a word indicated by information received by the input means. The speech recognition according to claim 5, further comprising: update means for creating a speech recognition dictionary by constructing a new tree structure by removing a word / phrase to which the mark is added from the existing speech recognition dictionary. Dictionary creation device. An input means for receiving input of information indicating that a predetermined phrase recorded in an existing speech recognition dictionary having a tree structure is changed to another phrase;
Mark adding means for adding a mark corresponding to the predetermined word indicated by the information received by the input means;
A speech recognition dictionary creation device, comprising: creation means for creating a speech recognition dictionary for the other words received by the input means separately from the existing speech recognition dictionary. By integrating the speech recognition dictionary created by the creating means and the speech recognition dictionary excluding the phrase with the mark added from the existing speech recognition dictionary by constructing a new tree structure, one speech The speech recognition dictionary creating apparatus according to claim 7, further comprising an updating unit that creates a recognition dictionary.

Images