JP2005189287A - Speech recognition device and speech recognition method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide an efficient speech recognition device which starts with a dictionary of a small capacity at the beginning and dynamically adds contents for learning to the dictionary according to a use environment, because conventionally, when operating speech recognition equipment on a means of transportation such as a vehicle, there has been only a choice between the two alternatives to limit a task to make recognition accuracy high or to ease a task limit to use memory in large quantities. <P>SOLUTION: The speech recognition device is so configured that it recognizes an input speech signal by using a network language dictionary; an extracting means extracts the other vocabularies having much frequency information on connection or appearance concerning a vocabulary as a result of the recognition; and an adding means adds the above other vocabularies together with the frequency information on the connection and appearance to the network language dictionary. <P>COPYRIGHT: (C)2005,JPO&NCIPI

Description

  In particular, the present invention relates to a voice recognition apparatus and a recognition method applied to a limited use such as control of equipment mounted on a vehicle such as a vehicle or a navigation apparatus.

  Conventionally, as a dictionary used in a speech recognition apparatus, as described in “Non-patent document 1” below, a network grammar language dictionary in which a grammar is described by a finite automaton and a statistical connection between words or words There was a statistical language dictionary expressing probabilities. The former describes the connection pattern of words in advance, and has the advantage that the structure is simple and the dictionary capacity is small, but the user is required to follow the preset connection grammar of the word. High recognition accuracy was obtained only in limited applications.

  On the other hand, a statistical language dictionary stores a large amount of sample data, and estimates the transition probability of a vocabulary (word, idiom or morpheme) and vocabulary (word, idiom or morpheme) using a statistical method. There is a problem that the application target is flexible but requires a large amount of memory, the memory is wasted for a relatively limited use, and the recognition rate is not as high as the network grammar language dictionary. was there.

  There are reported examples such as “Non-Patent Document 2” and “Non-Patent Document 3” as methods for solving such problems, but they have not reached the practical level.

"Voice Recognition System" Ohm Publishing Office Tsurumi, Lee, Saruwatari, Shikano “Examination of Algorithm Using Vocabulary N-gram and Network Grammar” Acoustical Society of Japan 2002 Autumn Meeting Konashi et al., "A Study on Automatic Grammar Generation for Spoken Dialogue System Using Dialogue Corpus" The Acoustical Society of Japan Autumn 2003 Sponsored Papers

  An object of the present invention is to provide a network language dictionary that can be flexibly adapted to the pronunciation grammar used by the user in consideration of the use of the network language dictionary.

  In order to achieve the above object, the present invention has a storage means for storing a network language dictionary and a statistical language dictionary, and an arbitrary vocabulary A in the recognition vocabulary A from the recognition means output for recognizing the input speech. Is present in the statistical language dictionary, the vocabulary B having a lot of connection or appearance frequency information is extracted from the vocabulary A, and the frequency information and the vocabulary B are added to the network language dictionary. (Here, frequency information refers to the appearance frequency and probability of each vocabulary, what vocabulary to transition to, what vocabulary to transition from, what kind of appearance frequency and appearance probability a given connected vocabulary has. This refers to information on the appearance and connection of the vocabulary such as “Isaka”.

  According to the configuration of the present invention, the recorded contents of the network language dictionary can be added at any time even when the speech recognition apparatus is in operation, and a dictionary adapted to the usage state can be dynamically constructed. As a result, it is only necessary to describe the basic grammar at the time of design, and it is possible to obtain high recognition accuracy by effectively using the memory capacity. A language dictionary can be created.

  In speech recognition, input speech is AD-converted, and a language expression ω that best matches the discrete sequence x is estimated. In order to realize these, in order to predict the language expression ω, a dictionary in which language expressions are described in advance (hereinafter referred to as language dictionary) is required. As a technique proposed as a conventional technique, there are a dictionary in which a grammar describes a grammar using a finite automaton (hereinafter referred to as a network grammar language dictionary), a statistical language dictionary that expresses a statistical connection probability between vocabularies, Has been proposed.

  In the network grammar language dictionary, a system designer predicts a pattern of a sentence that a user can speak in the task and describes in advance a connection pattern of allowed vocabulary. This has the advantage that the structure is simple and the dictionary capacity is small, and it is known that the user has extremely high recognition accuracy as long as the user performs the input intended by the system designer. However, it is vulnerable to user utterances not included in the pattern. Generally speaking, human utterance patterns fluctuate, and even if the same content is uttered, the utterance patterns are not always the same due to omission and inversion of the particles and omission of the subject. For example, even if all utterance patterns are investigated and predicted in advance and the grammar is created accordingly, the user's utterances are highly individual and diverse for each user, and it is difficult to cover all of them. Also, when the scope of tasks such as voice commands is limited to some extent, it is not impossible to construct a grammar that covers all of them, but grammar will be added to increase the coverage of such ambiguity and fluctuations. This results in a problem that the range of sentences accepted by the grammar is inappropriately widened, and the recognition rate deteriorates due to an increase in unnecessary hypotheses.

  On the other hand, statistical language dictionaries estimate vocabulary-to-vocabulary transition probabilities from a large amount of sample data using statistical methods, which requires a large amount of memory and only increases redundancy for specific applications. In addition, there are problems regarding recognition accuracy and processing time.

As already mentioned, in applications such as navigation systems where the scope of application is relatively limited, such as aircraft, ships, vehicles, work vehicles, and other applications that use voice recognition in various vehicle environments, address input and operation command input Language dictionaries using network grammar have been widely used because it is sufficient to accept voice utterances limited to specific tasks. However, in speech recognition using a network grammar language dictionary, it is necessary to determine in advance a grammar that can be input as described above. Recognize in advance the grammar used by the user,
2. Describe all the grammar that the user can speak,
It is necessary to satisfy one of the conditions. On the other hand, in a language dictionary called an N-gram language dictionary (hereinafter abbreviated as N-gram dictionary), although the degree of freedom of grammar that can be input is high, the speech recognition accuracy is low compared to the network grammar language dictionary, and the computation time is short. Therefore, it has not been used for the purpose of accepting task-limited voice utterances. Here, the N-gram dictionary is formed in a statistical language dictionary, and the N-gram dictionary stores an N-gram model (that is, a connection probability based on the N-gram) and a vocabulary dictionary. Yes. Here, the N-gram dictionary is a statistical language dictionary that is the simplest and most widely used method for estimating vocabulary and vocabulary transition probabilities from a large amount of sample data using a statistical method. This is when estimating the input language expression.

のような近似を行うモデルである。このとき、Ｐ(ω_１ω_２…ω_ｎ)は入力された語彙列ω_１ω_２…ω_ｎに対する出現確率を表す。

It is a model that performs an approximation like this. In this _{case, P (ω 1 ω 2 ...} ω n) represents the probability of occurrence for the vocabulary columns ω ₁ ω _{2 ...} ω _n that has been input.

  However, among the above conditions required for the network grammar language dictionary, 2 is difficult due to problems such as design cost because a large number of various grammars are considered. For this reason, the recognition performance is similar to that of a network grammar language dictionary (hereinafter abbreviated as “network language dictionary”) under a specific condition while maintaining a speech receiving function with a high degree of freedom like an N-gram dictionary. In particular, a speech recognition device dictionary creation method that can be acquired dynamically is desired. Here, the N-gram language dictionary can automatically include many vocabulary connection patterns if there is a large amount of sample data to be learned. It is also possible to accept input grammar. However, speech recognition using a statistical language dictionary requires a large amount of memory and a large amount of calculation, and using it for speech recognition limited to a specific task is redundant in terms of cost. In addition, the degree of freedom is higher than that of the network language dictionary, but there is a problem that the recognition rate is low (Non-Patent Document 1).

  In the method described in “Non-Patent Document 2” (GA method), the network grammar is determined in advance, and the logarithmic likelihood of the connected vocabulary whose category matches the network grammar in the N-gram dictionary based on the information. Multiplying the coefficient every time and adjusting the final recognition score. At this time, the larger the vocabulary contained in the network grammar, the more connected vocabulary that is modified at the time of output, and the output result approaches a dictionary that uses only the network grammar language dictionary. Even if it is simply applied, it is expected that it is less effective than the network language dictionary.

For this reason, in the present invention, information on the frequency of vocabulary and connected vocabulary existing in the statistical language dictionary is extracted, and this extracted vocabulary is first added to the existing network language dictionary manually created by the designer. This is a technique for automatically generating a network language dictionary even if the designer does not describe all the grammars that the user can speak in advance. By using the method according to the present invention, it is possible to automatically add an accurate vocabulary for a target task to the network language dictionary. That is, (1) only necessary vocabulary (vocabulary that can be spoken by the user) can be added to the network language dictionary when necessary, and (2) the size of the network language dictionary can be changed sequentially.
In the present invention, the basic network grammar language dictionary is described as using a vocabulary connection pattern that can be predicted by the system designer, but is not limited thereto. Others are those that are created by acquiring a basic language according to the situation by setting certain conditions according to the usage and user's preference from a huge language database, etc., or vocabulary with a high frequency of appearance Various basic network dictionaries can be considered, such as those that extract only such as to create a basic network.

  As described above, in “Non-Patent Document 3”, the transition probability is assigned to a vocabulary that already exists in the grammar to improve the narrowing performance of the grammar. The difference is that it dynamically expands (adds new vocabulary) based on statistical information.

  The configuration of the present invention will be described below with reference to the drawings.

(First embodiment)
FIG. 1 shows a system of each means for performing speech recognition processing according to the first embodiment of the present invention, and FIG. 2 shows an apparatus configuration relating to this system. In the system diagram of FIG. 1, the voice input means 110 collects voices spoken by the user and converts them into easy-to-handle digital voice signals. This corresponds to the microphone 335 and the AD converter 330 in FIG. 2, and specifically includes an audio input device represented by a microphone, an AD converter that is a real-time signal discretizer, and the like. Here, the audio signal is collected and converted into a discrete audio signal by performing AD conversion.

  The voice recognition means 120 in FIG. 1 recognizes the input voice signal and sends a recognition result signal W1. The recognition result signal W1 is converted into an information form such as text. This is realized by the arithmetic device 320 and the storage device 310 of FIG. As the arithmetic unit 320, for example, a single or a plurality of CPUs, MPUs, and DSPs constituting a system having an arithmetic function such as a general personal computer, a microcomputer, and a signal processing unit may be combined, and real-time processing can be performed. It is desirable to have possible computing power. The storage device may be a device having information storage capability used for general information processing devices such as a cache memory, a main memory, a disk memory, a flash memory, and a ROM. The recognition result signal W1 sent from the voice recognition means is converted into information to be presented to the user or converted into an operation signal of another device.

  In the storage unit 130 of FIG. 1, a network language dictionary 131 used for speech recognition pattern matching, a statistical language dictionary 132 having information for automatically generating the network language dictionary 131, a speech recognition unit 120, and a network Phonemes and vocabulary dictionaries 133 are stored as mediating means for data exchange with the language dictionary. This corresponds to the storage device 310 of FIG. 2, and specifically, a device having a rewritable information storage capability used for general information processing devices such as a cache memory, a main memory, a disk memory, and a flash memory. May be used.

  1 extracts one or more vocabulary B having a lot of frequency information related to connection from the vocabulary A that recognizes a user's utterance and is character information, frequency information about the connection between vocabulary A and vocabulary B, and vocabulary. Frequency information regarding the appearance of B is extracted from the statistical dictionary. This can be realized by the arithmetic device 320 and the storage device 310 of FIG.

  1 adds to the network language dictionary 131 the vocabulary B extracted by the extraction means 150 of FIG. 1, frequency information regarding the connection between the vocabulary A and vocabulary B, and frequency information regarding the appearance of the vocabulary B. It is means to do.

  The flow of recognition processing in the configuration shown in FIGS. 1 and 2 will be described with reference to FIG.

  In FIG. 3, when the voice recognition system in the voice recognition apparatus starts to operate (START), first, in step S110, the voice recognition system is initialized. At this time, the network language dictionary 131 is expanded on the memory, and the voice recognition system including the voice recognition unit 120 is activated to wait for voice input by the user. At this time, if the statistical language dictionary 132 is also expanded on the memory, the processing becomes efficient although it costs high.

  In step S120, the vocabulary W (i) recognized by the speech recognition means 120 is checked for the presence of a vocabulary that matches the vocabulary W (i) from the statistical language dictionary 132. If a vocabulary that matches the recognized vocabulary is found in the statistical language dictionary 132 (Yes in step S120), the vocabulary W (i) is sent to step S130. If no vocabulary is found (No in step S120), the process returns to the state immediately after initialization, and again enters the wait state for the speech recognition process.

In step S130, there is a recognized vocabulary W (i) in the vocabulary group included in the statistical language dictionary 132, and further, for this vocabulary W (i), a vocabulary group Wc (i) that exceeds a threshold value J regarding frequency. Is present (Yes in step 130), this vocabulary group Wc (i), frequency information regarding the connection between the vocabulary W (i) and the vocabulary group Wc (i), the vocabulary W (i) and this vocabulary group Wc ( The frequency information regarding the appearance of i) is sent to step S140. If the vocabulary group Wc (i) does not exist (No in step 130), the state returns to the state immediately after the initialization, and again enters the waiting state for the speech recognition process.

  In step S140, the vocabulary group Wc (i) extracted from the statistical language dictionary 132, frequency information regarding the connection between the vocabulary W (i) and the vocabulary group Wc (i), the vocabulary W (i) and the vocabulary group Wc (i) After deleting unnecessary vocabulary and frequency information regarding the frequency information related to the appearance of the vocabulary, the vocabulary group Wc (i), the frequency information regarding the connection from the vocabulary W (i) to the vocabulary group Wc (i), and the vocabulary group Wc (i) And frequency information regarding the appearance of the vocabulary W (i) is added to the network language dictionary. Here, the unnecessary vocabulary indicates a vocabulary that does not correspond to any task in the completed network language dictionary 131.

  In step S160, the i + 1th unprocessed vocabulary W (i + 1) is searched. If there is an unprocessed vocabulary W (i + 1), the process proceeds to step S130 via S150. If there is no vocabulary W (i + 1), the process returns to the state immediately after initialization and again enters the wait state for the speech recognition process. Here, the state returns to the state immediately after the unprocessed vocabulary initialization, and again enters the waiting state for the speech recognition processing. Here, the unprocessed vocabulary W (i + 1) is the connected vocabulary group Wc (i) added in step S140.

(Second Embodiment)
FIG. 4 shows a system diagram of the processing means according to the second embodiment of the present invention. In the history means 160, an appearance history is stored for each type of recognized vocabulary for the user utterance, and a specific vocabulary is further stored. , The vocabulary input frequency is calculated, and if the vocabulary input frequency exceeds a certain threshold, the vocabulary is sent to the statistical language dictionary 132 of FIG. In this case, the type of vocabulary may be a single vocabulary, or a plurality of vocabularies may form a class. This can be realized by the arithmetic device 320 and the storage device 310 of FIG. The above threshold value may be set in advance at the time of design, or may be set dynamically according to the input frequency of all vocabularies. Furthermore, when this threshold value is set from the outside, the input device 340 of FIG. 2 may be used. Specifically, the input device 340 can be realized if there is a device that allows a user to directly input individual words such as a switch, a keyboard, and a voice input device.

Hereinafter, the flow of the speech recognition process in the configuration of FIG. 4 will be described with reference to FIG.
In FIG. 5, in step S120, as in the case of the first embodiment, the speech recognition process ends, and the recognized vocabulary W (i) is searched from the statistical language dictionary 132. When the vocabulary is found (Yes in step S120), the vocabulary W (i) is sent to step S121.
In step S121, if the information on the vocabulary W (i) received from step S120 already exists in the history means 160, the number of times the vocabulary W (i) is input is counted. To initialize the counter that counts the input frequency. Further, when the number of times the vocabulary W (i) is input exceeds the threshold (Yes in step S121), the vocabulary W (i) is sent to step S130. If the threshold value is not exceeded (No in step S121), the process returns to the state immediately after initialization and again enters the wait state for the voice recognition process. The vocabulary once sent out may be managed using a flag or the like in order to avoid useless calculation due to the vocabulary being sent to S130 again.

  In the second embodiment, the case where the two-step vocabulary of the vocabulary A and the vocabulary B is connected has been described. However, as in the case of the first embodiment, the vocabulary C is connected 3 In the case of steps, there may be a multi-step connection such as vocabulary D and vocabulary E. Even in such a case, the frequency information regarding connection of each vocabulary, the frequency information regarding appearance, the vocabulary C, and the vocabulary D may be sequentially added to the network language dictionary in the same manner as described above.

(Third embodiment)
6 predicts the user's next utterance vocabulary based on the recognition vocabulary for the user's utterance, and sends the next utterance vocabulary to the statistical language dictionary 132 shown in FIG. This can be realized by the arithmetic device 320 and the storage device 310 of FIG. Furthermore, when the next utterance to be transmitted is additionally input from the outside, the input device 340 of FIG. 2 may be used.

The flow of the network language dictionary automatic generation process shown in FIG. 6 will be described below with reference to FIG.
In step S120, the speech recognition process ends, and the recognized vocabulary W (i) is searched from the statistical language dictionary 132. When the recognized vocabulary is found in the statistical language dictionary 132 (Yes in step S120), the vocabulary W (i) is sent to step S122. If not found (No in step S120), the process returns to the state immediately after the system initialization, and again enters the wait state for the speech recognition process.
In step S122, the next utterance vocabulary group Wn (i) is searched for the vocabulary W (i) received from step S120. If the next utterance vocabulary group Wn (i) exists (Yes in step S122), the process proceeds to step S130. If it does not exist (No in step S122), the process returns to the state immediately after initialization, and again waits for the voice recognition process. Become. The following is the same as in FIG.

  In the third embodiment, the case where the two-step vocabulary of the vocabulary A and the vocabulary B is connected has been described. However, in the case of the three steps where the vocabulary C is connected, the vocabulary D, the vocabulary E, etc. There can be a connection of steps. Even in such a case, the frequency information regarding connection of each vocabulary, the frequency information regarding appearance, the vocabulary C, and the vocabulary D may be sequentially added to the network language dictionary in the same manner as described above.

(Fourth embodiment)
In the following, the process of adding the recognized vocabulary to the network language dictionary with reference to FIG. 8 will be described using the specific vocabulary with respect to the processing flow shown in FIG.

  The “language dictionary at time t-1” in FIG. 8 indicates the network language dictionary 210 before the change, and the “language dictionary at time t” indicates the network language dictionary 211 after the change by adding a vocabulary that has been voice-recognized. Is shown. When the speech-recognized vocabulary “air conditioner” 220 is input, it is sent to the statistical language dictionary 230. From the statistical language dictionary 230, two words, “Put” and “Put”, have been found as vocabularies that are easily shifted from the recognized vocabulary “Air Conditioner” 220. Both these vocabularies are extracted by the extracting means (150 in FIG. 1), and further added to the network language dictionary 211 by the adding means (140 in FIG. 1). In the network language dictionary 211 of “language dictionary at time t”, the newly added vocabulary “Put” and “Pick” are added as connection vocabulary from “Air conditioner” to one lower hierarchy. . Furthermore, frequency information related to the connection from the vocabulary “air conditioner” to “put in” and “put on” and frequency information related to appearance are also added.

  As for the frequency information related to the appearance, it is sufficient to add only information related to the corresponding vocabulary, or the frequency information related to the appearance of “air conditioner”, “CD” which is another vocabulary included in the first layer of the network language dictionary 211, and All “radio” may have the same appearance probability. In addition, regarding other vocabularies included in the first level of the network language dictionary 211, frequency information regarding appearance in the statistical language dictionary 230 is checked, and the three vocabularies “air conditioner”, “CD”, “radio”, The balance ratio regarding the connection and appearance with the vocabulary that has not been changed may be calculated again and added.

As for the frequency information related to the connection, it is only necessary to add information related to the corresponding vocabulary, but the balance ratio is calculated again for the frequency information related to the two connections from “air conditioner” to “put in” and “turn on”. However, this may be added. For example, the probability of transitioning to “put in”, “put on”, “ON”, and “OFF” may be assigned a probability of 1/4 for any vocabulary.
Here, the balance ratio corresponds to the transition probability for each vocabulary that is a candidate connected to a certain vocabulary. For example, the vocabulary connected to “air conditioner” is a vocabulary before being changed. If there are only two types of “ON” and vocabulary “OFF”, the vocabulary is “½”, whereas the vocabulary “put” is added to the network language dictionary 210 to “language dictionary at time t”. In this case, the vocabulary that can be connected when the network language dictionary 211 is in the state of the network language has three vocabularies that allow the air conditioner to perform the same operation: “put in”, “attach”, and “ON”. The ratio is 1/3.

In this way, the adjustment of the balance ratio related to the connection between two vocabularies is, for example, when considering three vocabularies of vocabulary A, vocabulary B, and vocabulary C.
1. If vocabulary B and vocabulary C are in the same hierarchy and one is connected after vocabulary A,
2. If vocabulary A and vocabulary C are in the same hierarchy and either is connected before vocabulary B,
3. If vocabulary A and vocabulary C are in the same hierarchy and one is connected after vocabulary B,
4). If vocabulary B and vocabulary C are in the same hierarchy and either is connected before vocabulary A,
It is necessary to adjust the balance ratio for each of these connections. Similar adjustment of the balance ratio is necessary for adjustment of the balance ratio related to the appearance of two vocabularies in the three vocabularies.

(Fifth embodiment)
Hereinafter, the process of adding a vocabulary to the network language dictionary will be described with reference to FIG.

  “Language dictionary at time t-1” 410 in FIG. 9 shows the network language dictionary before the change, and “Language dictionary at time t” 411 shows the network language dictionary after the change by adding the vocabulary recognized by the voice. Is shown. When the vocabulary “air conditioner” 420 whose input frequency of the speech-recognized vocabulary exceeds the threshold is input, it is sent to the statistical language dictionary 430.

  From the statistical language dictionary 430, two words, “Put” and “Put”, have been found as vocabularies that are easily changed from “Air conditioner”. Both these vocabularies are extracted by the extracting means (150 in FIG. 4), and further added to the network language dictionary 411 by the adding means (140 in FIG. 4). In the network language dictionary 411 of “language dictionary at time t”, newly added vocabularies “put” and “put” are added to the hierarchy one level lower than “air conditioner”. Furthermore, frequency information related to the connection from the vocabulary “air conditioner” to “put in” and “turn on” and frequency information related to appearance are also added.

  As for the frequency information related to the appearance, it is sufficient to add information related to the corresponding vocabulary, or the frequency information related to the appearance of “air conditioner” and other vocabularies included in the first layer of the network language dictionary “CD” and “ “Radio” may all have the same appearance probability. In addition, regarding other vocabularies included in the first level of the network language dictionary, frequency information regarding appearance in the statistical language dictionary 430 is checked, and the balance ratio of the three vocabularies “air conditioner”, “CD”, and “radio” is checked. May be recalculated.

  For the frequency information related to the connection, it is only necessary to add information about the corresponding vocabulary. However, for the frequency information related to the two connections from “air conditioner” to “put in” and “turn on”, the balance ratio is calculated again. For example, the probability of transition to “put in”, “put on”, “ON”, and “OFF” may be assigned a probability of 0.5 for any vocabulary.

(Sixth embodiment)
Hereinafter, the process of adding a vocabulary to the network language dictionary will be specifically described with reference to FIG. The “language dictionary at time t−1” in FIG. 10 indicates the network language dictionary 510 before the change, and the “language dictionary at time t” indicates the network language dictionary 511 after the change by adding a vocabulary that has been voice-recognized. Is shown. When the predicted vocabulary “air conditioning” 520 is input as the next utterance of the speech-recognized vocabulary, it is sent to the statistical language dictionary 530. FIG. 10 shows an example in which “ON” and “OFF” are found from the statistical language dictionary 530 as vocabularies that are easy to transition from “air conditioning” to the input vocabulary “air conditioning”. Here, since “ON” and “OFF” in the statistical language dictionary are also included in the “language dictionary at time t−1”, they are deleted by the extracting means or the adding means, and only “air conditioning” is the network language. Sent to a dictionary. Therefore, the vocabulary added here is only “air conditioning” which is the vocabulary input and recognized. On the other hand, frequency information regarding the connection from “air conditioning” to “ON” and “OFF” and frequency information regarding the appearance of “air conditioning”, “ON”, and “OFF” are added. Regarding the frequency information related to the connection, only the information related to the corresponding vocabulary may be added, or the balance ratio may be calculated again for the frequency information related to the connection from “air conditioning” to “ON” and “OFF”.

(Seventh embodiment)
A specific example of the next utterance prediction in FIGS. 6 and 7 will be described below.
The vocabulary B inferred from the recognized vocabulary A is predicted to be used in the near future by the user, and the corresponding vocabulary B and the vocabulary C connectable from the vocabulary B are added to the network language dictionary. In order to generate the vocabulary B from the vocabulary A, the next utterance prediction means (170 in FIG. 6) is used. The operation of the next utterance prediction means 170 is to send i) a synonym of vocabulary A, ii) an antonym of vocabulary A, iii) a vocabulary associated with vocabulary A, and the like.

Synonyms: For example, “air conditioner” → “air conditioner”, “cooling”, “heating” etc. Antonym: ON → OFF, turn on, turn on, turn off,
Associative words: Nissan → Nissan Prince, Nissan Satio, XX Nissan Sales, Skyline Museum, etc. The next utterance prediction means 170 needs to have a data group that groups the predicted next utterances against the vocabulary entered in advance. is there.

(Eighth embodiment)
In a speech recognition system whose purpose is to perform device operation, in order to dynamically change a language dictionary, it is necessary to set a changed vocabulary and a link to device operation corresponding to the connection of the vocabulary. . Various methods are possible, but some methods are briefly described below.
(Method 1) The vocabulary extracted from the statistical language dictionary 630 in advance is put into the network language dictionary 610 (FIG. 11).

  In contrast to the network language dictionary 610 designed by the designer (FIG. 11: state 1), the network language dictionary 611 has a vocabulary added by a statistical language dictionary at the time of factory shipment (FIG. 11: state 2). Although many vocabularies are added at this time, the parts other than the vocabulary (italic characters and broken lines) included in the network language dictionary intentionally described by the designer are insensitive. Must not.

The next utterance vocabulary inputted by the next utterance prediction means and the vocabulary connected from the next utterance vocabulary are sequentially made sensitive (FIG. 11: state 3, “air conditioning”, “air conditioning” to “ON”, “OFF” Connection).
(Method 2) Assigning a class to a vocabulary included in a statistical language dictionary Add part-of-speech information to vocabulary contained in statistical language dictionaries (nouns, verbs, adjectives, particles, etc.)
2. For nouns and verbs that are directly related to operations, class information related to operation tasks is assigned. Number, input / output (ON / OFF)
For the vocabulary connected from the input vocabulary, the operation task is applied based on the class information. As a general rule, no nouns and verbs are added when there is no operation task, that is, when there is no content related to device operation. However, even if the part of speech is not related to the operation, for example, in the case of a vocabulary that is added before the name of the operation target device, such as “a little”, “radio”, or “a little”, the recognition device is “a little radio. "" As a single vocabulary, and irrelevant vocabulary may be extracted. To avoid this, vocabulary such as "a little" or "um" may be added.
(Method 3) Prompt the user for input When adding a vocabulary, a dialog is displayed to prompt the user to input operation information.

  The outline and embodiments of the present invention have been briefly described above. The above-described examples are only for easy understanding of the contents of the invention, and do not limit the scope of the invention. Further, it is possible to easily realize a single example or a combination of a plurality of examples.

The systematic diagram of the process in 1st Embodiment. The block diagram of the apparatus structure in 1st Embodiment. The flowchart of the addition process with respect to the network language dictionary in 1st Embodiment. The systematic diagram of the process in 2nd Embodiment. The flowchart of the addition process with respect to the network language dictionary in 2nd Embodiment. The systematic diagram of the process in 3rd Embodiment. The flowchart of the addition process with respect to the network language dictionary in 3rd Embodiment. The flowchart of the process using the specific vocabulary in 4th Embodiment. The flowchart of the process using the specific vocabulary in 5th Embodiment. The flowchart of the process using the specific vocabulary in 6th Embodiment. The flowchart of the process using the specific vocabulary in 7th Embodiment.

Explanation of symbols

110: voice input means 120: voice recognition means 130: storage means 131: network language dictionary 132: statistical language dictionary 140: addition means 150: extraction means 160: history means 170: next utterance prediction means 210: time t-1 Language dictionary 211: Language dictionary at time t 220: Recognition vocabulary at time t-1 230: Statistical language dictionary 310: Storage device 320: Arithmetic device 330: AD converter 335: Microphone 340: Input device 410: Time t-1 411: Language dictionary at time t 420: Recognition vocabulary at time t-1 430: Statistical language dictionary 510: Language dictionary at time t-1 511: Language dictionary at time t 520: Next utterance at time t-1 Predictive vocabulary 610: Network language dictionary in state 1 611: Network language dictionary in state 2 612: Network language dictionary in state 3 620: Next time t-1 Talk prediction vocabulary 630: Statistical language dictionary

Claims (17)

In a speech recognition device used in a navigation system,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
When an arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary, one or more vocabulary B having a lot of frequency information regarding connection to the vocabulary A in the statistical language dictionary Extracting means for extracting
Add to the network language dictionary frequency information related to the connection between the vocabulary B and the vocabulary B to either the vocabulary A or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B And a voice recognition device. In a speech recognition device used in a navigation system,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
When an arbitrary vocabulary A included in the recognized vocabulary is present in the statistical language dictionary, the vocabulary A in the statistical language dictionary has a lot of frequency information regarding connection and is related to appearance. Extraction means for extracting one or more vocabulary B having a lot of frequency information;
Add to the network language dictionary frequency information related to the connection between the vocabulary B and the vocabulary B to either the vocabulary A or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B And a voice recognition device. In a speech recognition device used in a navigation system,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
A history is stored for each type of word for the recognized vocabulary,
A history in which the newly recognized vocabulary of the vocabulary of the vocabulary is newly registered in the history, and if it is the same as the already registered vocabulary, the history is added to the number of already entered vocabularies Means,
When an arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary and the input amount exceeds a preset threshold value, a large amount of connection frequency information is given to the vocabulary A. Extraction means for extracting one or more vocabulary B having;
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In a speech recognition method used in a navigation system,
Input the voice into the voice input means,
The input speech is recognized by speech recognition means and converted into vocabulary,
Furthermore, a network language dictionary and a statistical language dictionary are set on the storage means,
A history is stored for each type of vocabulary or vocabulary group by the history means for the recognized vocabulary, and the frequency of input of the vocabulary is counted,
When the arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary and its input frequency exceeds a preset threshold, it has a lot of frequency information on connection, and , One or more vocabulary B having a lot of frequency information about appearance is extracted by the extraction means,
The vocabulary B, frequency information related to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B are added to the network language dictionary by an adding unit. A speech recognition method characterized by: In a speech recognition device used in a navigation system,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
A next utterance prediction means for predicting a user's next utterance;
When one or more vocabulary A predicted as the next utterance of the recognized vocabulary exists in the statistical language dictionary, one or more vocabulary B having a lot of connection frequency information is extracted from the vocabulary A Extraction means to
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In a speech recognition method used in a navigation system,
Input voice by voice input means,
The input speech is recognized by speech recognition means and converted into vocabulary,
Furthermore, the network language dictionary is stored in the storage means 1,
Storing the statistical language dictionary in the storage means 2;
Predict the next utterance of the user with the next utterance prediction means,
When one or more vocabulary A predicted as the next utterance of the recognized vocabulary exists in the statistical language dictionary and the input frequency exceeds a preset threshold, One or more vocabulary B having a large amount and having a lot of frequency information regarding appearance is extracted by the extraction means,
The vocabulary B, frequency information related to connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B are added to the network language dictionary by an adding unit. A speech recognition method comprising: In a speech recognition device used in a vehicle,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
When an arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary, one or more vocabularies B having a lot of frequency information related to connection to the vocabulary A in the statistical language dictionary. Extracting means for extracting;
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In a speech recognition device used in a vehicle,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
When an arbitrary vocabulary A included in the recognized vocabulary is present in the statistical language dictionary, the vocabulary A in the statistical language dictionary has a lot of frequency information regarding connection and is related to appearance. Extraction means for extracting one or more vocabulary B having a lot of frequency information;
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In a speech recognition device used for a vehicle,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
A history is stored for each type of word for the recognized vocabulary,
A history in which the newly recognized vocabulary of the vocabulary of the vocabulary is newly registered in the history, and if it is the same as the already registered vocabulary, the history is added to the number of already entered vocabularies Means,
When an arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary and the input amount exceeds a preset threshold value, a large amount of connection frequency information is given to the vocabulary A. Extraction means for extracting one or more vocabulary B having;
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In a method for creating a dictionary for a speech recognition device used in a vehicle,
Input the voice into the voice input means,
The input speech is recognized by speech recognition means and converted into vocabulary,
Furthermore, a network language dictionary and a statistical language dictionary are set on the storage means,
A history is stored for each type of vocabulary or vocabulary group by the history means for the recognized vocabulary, and the frequency of input of the vocabulary is counted,
When the arbitrary vocabulary A included in the recognized vocabulary exists in the statistical language dictionary and its input frequency exceeds a preset threshold, it has a lot of frequency information on connection, and , One or more vocabulary B having a lot of frequency information about appearance is extracted by the extraction means,
The vocabulary B, frequency information related to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B are added to the network language dictionary by an adding unit. A speech recognition method characterized by: In a speech recognition device used in a vehicle,
Voice input means for inputting voice;
Speech recognition means for recognizing input speech and converting it into vocabulary,
Storage means 1 for storing a network language dictionary;
Storage means 2 for storing a statistical language dictionary;
A next utterance prediction means for predicting a user's next utterance;
When one or more vocabulary A predicted as the next utterance of the recognized vocabulary exists in the statistical language dictionary, one or more vocabulary B having a lot of connection frequency information is extracted from the vocabulary A Extraction means to
Additional means for adding to the network language dictionary the vocabulary B, frequency information relating to the connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information relating to the appearance of the vocabulary B A speech recognition apparatus characterized by comprising: In speech recognition methods used in vehicles,
Input voice by voice input means,
The input speech is recognized by speech recognition means and converted into vocabulary,
Furthermore, the network language dictionary is stored in the storage means 1,
Storing the statistical language dictionary in the storage means 2;
Predict the next utterance of the user with the next utterance prediction means,
When one or more vocabulary A predicted as the next utterance of the recognized vocabulary exists in the statistical language dictionary and the input frequency exceeds a preset threshold, One or more vocabulary B having a large amount and having a lot of frequency information regarding appearance is extracted by the extraction means,
The vocabulary B, frequency information related to connection of the vocabulary A to the vocabulary B or another vocabulary C included in the network language dictionary, and frequency information related to the appearance of the vocabulary B are added to the network language dictionary by an adding unit. A speech recognition method comprising: In the speech recognition device according to any one of claims 1 to 3, claim 5, claim 7 to claim 9 or claim 11,
Extraction means for extracting frequency information relating to the connection between the vocabulary A and the vocabulary B from the statistical language dictionary;
The balance ratio between the frequency information related to the connection between the vocabulary A and the vocabulary B and the frequency information related to the connection between two vocabulary C, the vocabulary A, and the vocabulary B included in the network language dictionary is adjusted. A speech recognition apparatus comprising: the vocabulary B; frequency information relating to the connection between the vocabulary A and the vocabulary B; and the adding means for adding frequency information relating to the appearance of the vocabulary B. In the speech recognition device according to any one of claims 1 to 3, claim 5, claim 7 to claim 9 or claim 11,
Extraction means for extracting frequency information about the occurrence of the vocabulary A and the vocabulary B from the statistical language dictionary;
The balance ratio between the frequency information related to the connection between the vocabulary A and the vocabulary B and the frequency information related to the connection between two vocabulary C, the vocabulary A, and the vocabulary B included in the network language dictionary is adjusted. A speech recognition apparatus comprising: the vocabulary B; frequency information relating to the connection between the vocabulary A and the vocabulary B; and the adding means for adding frequency information relating to the appearance of the vocabulary B. The additional means according to any one of claims 1 to 3, 5, 7, 7 to 9, 11 or 14 is used by a new vocabulary added to the network language dictionary. A speech recognition apparatus, which is an additional means having a function of enabling device operation by linking to the input vocabulary when input by a person. The additional means in the speech recognition apparatus according to any one of claims 1 to 3, 5, 7, 7 to 9, 11, 11, 14 or 15 is provided in the network language dictionary. A speech recognition apparatus characterized by being an adding means having a function of not executing processing for adding a candidate vocabulary to the network language dictionary when there is no content related to device operation as a candidate vocabulary to be added . The extraction means in the speech recognition apparatus according to any one of claims 1 to 3, claim 5, claim 7, to claim 9, claim 11, claim 14 or claim 15, the network language dictionary. A speech recognition apparatus, comprising: extraction means having a function of deleting a candidate vocabulary from the network language dictionary when there is no content related to device operation as a candidate vocabulary to be added.

Images