JP3478171B2 - Voice recognition device and voice recognition method - Google Patents

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and a voice recognition method, and more particularly to a voice recognition device and a voice recognition method suitable when it is not necessary to re-register a dictionary corresponding to a service scenario.

[0002]

2. Description of the Related Art Conventionally, a voice recognition device has been developed in consideration of use by an unspecified speaker. An example of this kind of voice recognition device is described in a modification of the second embodiment of Japanese Patent Laid-Open No. 11-7292. The publication is intended to maintain high recognition performance for various types of additional words, and includes a speech analysis unit that extracts a feature amount of an input speech for each constant time (frame), and a recognition target word. Alternatively, a standard pattern in which additional word models are connected before or after or before or after a word string, and pattern matching between the standard pattern and the feature amount is performed for each frame, and an optimal word sequence in the word or word string ( An optimal column), and a likelihood calculating unit that calculates the likelihood thereof; and an output unit that determines and outputs the optimal recognition result from the optimal column and the likelihood, and the additional word model has background noise and arbitrary noise. A voice recognition device is disclosed, which is a model that accepts both voices.

That is, the grammar for continuous syllable recognition is a previously known additional word ("Eh" added before and after the word to be recognized).
And words such as "is" and recognition target words sandwiched between them, and using this grammar, additional words are removed and desired recognition is completed. For example, as shown in FIG. 12, this conventional speech recognition apparatus using word spotting connects and recognizes a model in which models for recognizing additional words are connected before and after a continuous grammar to be recognized. The conventional speech recognition apparatus using word spotting having such a configuration operates as follows. That is, the prefix unnecessary words are recognized according to the specified continuous grammar, the best keyword word is recognized from the specified keyword word group, the trailing unnecessary words are recognized, and the recognized keyword words are used as the recognition result. Output and end recognition.

Further, as another conventional example relating to the above speech recognition, for example, a technique described in Japanese Patent No. 2818362 has been proposed. The patent is directed to speaker-independent continuous speech recognition for multiple users, a memory coupled to the data processing means, the voice input means, and the character string using output device, and a memory on the memory. , A first context section including a first plurality of words including a second plurality of phonemes and a second context section on the memory including a third plurality of words including a fourth plurality of phonemes. And a pattern matching section on the memory including a fifth plurality of phoneme pattern matching data devices and each of the second plurality of phonemes in the fifth plurality of pattern matching data devices. Associating a first pointer map including a second plurality of pointers with each other and each of the fourth plurality of phonemes with each of the fifth plurality of pattern matching data devices; First A second pointer map including a plurality of pointers, and selecting the first context partition and the first pointer map or the second context partition and the second pointer map, Selecting means coupled to the memory for converting voice input information received from the input means to character string output information to the character string using device, wherein the context of the voice recognition application is: An instantaneous context switching system on a speech recognizer is disclosed that can be switched without loading a new pattern matching data device on memory.

[0005]

However, the above-mentioned conventional example has the following problems.

The first problem in the conventional example described in Japanese Patent Laid-Open No. 11-7292 is that the necessary or unnecessary recognition target words and additional words may differ depending on the scene of the service. If it's a shameful service, the recognition control grammar must be replaced. Therefore, it has to be loaded again into the memory (see FIG. 2, which will be described later in the embodiment of the present invention, at the time of speech recognition, a standard pattern is loaded, a grammar file is loaded, and speech recognition processing is performed).

The second problem is that in the case of a multi-line type system such as a station, the I / O access is increased by reloading the dictionary according to the service for each recognition unit of each connected call. However, since the load on the system increases, in the case of a multi-line type speech recognition device that provides a continuous operation type service, if the dictionary information is frequently reloaded for each service, the system load increases.

That is, when spotting of place-name words is completed by the conventional means, the grammar file on the memory must be replaced in order to activate spotting of person-name words next time, so a command flows from the host device or the like. In such a large system, there is a risk of hindering smooth service progress.

Further, in the conventional example described in the above-mentioned Japanese Patent No. 2,818,362, as the "context switching system and method of the voice recognition device", it is stated that the service is switched without newly loading the pattern matching data. It is said that the network has individual network information for each service, and that network has a pointer to the standard pattern to avoid the overlapping load of the standard pattern. The purpose is to avoid the overlapping load of the recognition control grammar in the present invention described later. Are essentially different. Also, if control information is registered in advance as in the present invention described later, switching can be performed without sending a command from the outside such as a keyboard selector, but this cannot be realized in the conventional example.

An object of the present invention is to provide a speech recognition apparatus and a speech recognition method which do not stop the system for changing the grammar in speech recognition using word spotting so that continuous operation is required. To do.

[0011]

SUMMARY OF THE INVENTION According to the present invention, a standard pattern registration in which a standard pattern of voice information is registered in a voice recognition device for performing voice recognition by removing unnecessary words added before and after a voice recognition target word. Means, a non-word registration means in which a continuous grammar holding all unnecessary words that are supposed to be used are registered, and a voice recognition by referring to the standard pattern registration means and the non-word registration means. The voice recognition means to be performed and the search space of the non-word registration means by the voice recognition means are used as a service syn
And a control means for appropriately reducing the size according to the Rio .

Further, the speech recognition apparatus of the present invention will be described with reference to FIG. 1. In the speech recognition apparatus which removes unnecessary words added before and after a speech recognition target word to perform speech recognition, Standard pattern registration means (131 in FIG. 1) in which the standard pattern of FIG. 1 is registered, and non-word registration means (FIG. 1 in FIG. 1) in which continuous grammars holding all unnecessary unnecessary words are registered. and 132), said standard pattern and the registration means and the reference speech recognition means for performing speech recognition non term registration means (120 in FIG. 1), the service search space of the non term registration means by speech recognition means Shi A control means (140 in FIG. 1) for appropriately reducing the size according to Nario is provided.

[Operation] The speech recognition apparatus of the present invention is controlled such that possible unnecessary word candidates are previously registered at one time and divided into small sets for use in a service scene. Therefore, if the dictionary corresponding to the expected service scenario is registered in the voice recognition device only once, the service can proceed without re-registering. Also, if all possible unnecessary words are registered, it is expected that there will be frequent errors caused by other words, but unnecessary recognition will be realized by switching the set of unnecessary words in time according to the service scene. It is controlled so that candidates can be eliminated. Therefore, it is expected that the recognition rate will be improved in the assumed service scenario. Also, by registering frequently used grammar control information as a pattern, recognition using word spotting can be efficiently switched.

[0014]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS [First Embodiment] Next, embodiments of the present invention will be described in detail with reference to the drawings.

(1) Description of Configuration FIG. 1 is a block diagram showing an example of the configuration of a speech recognition apparatus according to the first embodiment of the present invention. In FIG. 1, the voice recognition device according to the first exemplary embodiment of the present invention includes a voice input unit 100, a voice analysis unit 110, a distance calculation unit 120, and a standard pattern dictionary 1.
31 and a general-purpose word spotting grammar dictionary 132, and a grammar control unit 140 having a selection unit 141 and a specialization unit 142.

Explaining the configuration of each of the above units, the voice input unit 100 inputs the voice to be recognized. Speech analysis unit 11
0 analyzes the voice input from the voice input unit 100 and sends it to the distance calculation unit 120. The distance calculation unit 120
The parameterized voice is recognized by referring to the standard pattern dictionary 131 and the general-purpose word spotting grammar dictionary 132 of the data section 130. The standard pattern dictionary 131 of the data section 130 stores standard patterns of voice information in advance. In the general-purpose word spotting grammar dictionary 132 of the data section 130, all unnecessary words that are supposed to be used from the start to the end of the service scenario are registered in advance.

Here, the service scenario indicates a flow of a voice recognition service, in which a contracting subscriber connects to this service system and uses recognition processing to enjoy the service until disconnection. This is the flow of a series of voice recognition services. In this embodiment, the service scenario is a preloaded grammar file. However, not only the name of a person or a place, but depending on the service, it may be possible to enter a name of a person or a place, or another noun, so all the utterances that are input before the disconnection of a series of calls are covered It also becomes a grammar file like this.

Further, the grammar control unit 140 appropriately reduces the search space of the general-purpose word spotting grammar dictionary 132 according to the service scenario. The selection unit 141 of the grammar control unit 140 selects the general-purpose word spotting grammar dictionary 132 of the data unit 130 for each service.
The specialization means 142 of the grammar control unit 140 specifically specializes the general-purpose word spotting grammar dictionary 132.

In the voice recognition device having the above structure,
The voice input from the voice input unit 100 is the voice analysis unit 1
It is analyzed at 10 and sent to the distance calculation unit 120. The distance calculation unit 120 recognizes the parameterized voice using the standard pattern dictionary 131 and the general-purpose word spotting grammar dictionary 132 loaded in the data unit 130 in advance. In the general-purpose word spotting grammar dictionary 132, all unnecessary words that are supposed to be used from the start to the end of the service scenario are registered, and the distance calculation unit 120
Only necessary unnecessary words are selected according to the progress of the service scenario, and the voice of the keyword voice is recognized. As a result, the service can be completed without reloading the grammar dictionary 132 including unnecessary words.

Further, the structure of the main part will be described in detail. The standard pattern dictionary 131 stores standard patterns of voice information, and the general-purpose word spotting grammar dictionary 132 stores a desired service scenario completion. It is a continuous grammar that holds a prefix and a suffix including all unnecessary words required up to, and controls the calculation method in the distance calculation unit 120. As described above, the grammar control unit 140 includes a selection unit 141 that selects the general-purpose word spotting grammar dictionary 132 for each service, and a specific specialization unit 142. , General Word Spotting Grammar Dictionary 1
The 32 search spaces are appropriately reduced according to the service scenario.

The selection means 141 of the grammar control section 140 operates as follows. That is, the selecting unit 141 extracts, by the specializing unit 142, a path from a necessary unnecessary word from a set of unnecessary words registered in the general-purpose word spotting grammar dictionary 132 in response to a service scenario request,
The general-purpose word spotting grammar dictionary 132 so that other unnecessary words are not calculated by the distance calculation unit 120.
Remove from.

The specializing means 142 of the grammar control section 140 operates as follows. That is, the specializing means 142 may, for example, in order to reduce unnecessary word candidates, give an initial score of an unnecessary candidate a penalty score such that a transition from the initial score becomes invalid. In order to reduce the amount of calculation, it is possible to prepare a necessary / unnecessary flag in advance and remove it from the calculation candidate loop.

Here, the initial score is the distance calculation unit 120.
In, the initial value of the distance score when starting the distance calculation for each word, and by setting a penalty value described later to this initial value, the distance value becomes enormous and becomes a high-ranked candidate for recognition. You can hold down. Further, the penalty refers to a sufficiently large value so as not to be listed as a recognition candidate, for example, infinity in B2 (when calculated with a positive value).

(2) Description of Operation Next, the operation of the first embodiment of the present invention will be described with reference to FIGS.
Will be described in detail with reference to. 2 and 3 are flowcharts showing the voice recognition processing of the first embodiment, FIG. 4 is a flowchart showing the place name / personal name input processing of the first embodiment, and FIGS. 5 and 6 are cumulative distance calculation processing of the first embodiment. FIG. 7 to FIG. 9 are explanatory diagrams showing a specific example of the voice recognition according to the first embodiment.

In the voice recognition apparatus, as described above, the standard pattern dictionary 131 is loaded into the data section 130 (step S21 in FIG. 2), and the general-purpose word spotting grammar dictionary 132 is loaded (step S2 in FIG. 2).
2) Then, the distance calculation unit 120 performs voice recognition using the standard pattern dictionary 131 and the general-purpose word spotting grammar dictionary 132 (step S23 in FIG. 2).

The above process will be described in more detail. First, the standard pattern dictionary 131 is loaded into the address designated by the voice recognition device (step A1 in FIG. 3). Next, the general-purpose word spotting grammar dictionary 132 is loaded (step A2 in FIG. 3). Further, the grammatical control unit 140 selects unnecessary words necessary for the service by using the selection means 141 (step A3 in FIG. 3). Based on the information in step A3, the specialized means 142 creates a restricted grammar from the general-purpose word spotting grammar dictionary 132 (step A4 in FIG. 3).
After that, the distance calculation unit 120 performs recognition processing (step A5 in FIG. 3).

The selection means 141 of the grammar control unit 140 confirms unnecessary words necessary for the service and sets information (FIG. 5).
Step B1, Step B2, or Step B1, Step C1 of FIG. 6). Specialization means 1 of the grammar control unit 140
42 uses the information set by the selection means 141 to delete or invalidate unnecessary word candidates not required by the service from the grammar (step B2 in FIG. 5, or step C2, step C3, step C4 in FIG. 6). .

Next, the operation of the first embodiment of the present invention will be described using a specific example. As shown in FIG. 7, for example, the prefix unnecessary word candidates 1 to m (m = 7 in this case) of the general-purpose word spotting (WS) grammar dictionary 132 include “a”, “a”, and “e”. “To”, “”, “”, “”, or the null word φ is registered, and “from” and “to” are included in postfix unnecessary word candidates 1 to n (n = 8 in this case). , "Is", "san", "manager", "chief", "kun", or the null word φ is registered.

However, the null word φ here means a word that makes a state transition without the output of a symbol, and is a word whose content is empty. Although it is treated the same as a word in terms of algorithm, it actually works as a path that connects the preceding and following words by tunneling them. For example, the grammar that accepts the following sentence is "I read this book." And "I read the book."
→→→→→→→→→→ It will be. With this grammar, when recognition processing is performed, it is recognized as "I read (N) books." However, in the actual processing, "ha ~ book" is tunneled and directly connected. So, "I read a book." Therefore, in the case of FIG. 8 described later, the utterance content “Yamada-kun” (32 is “Yamada”)
However, since it is a null word,
It is judged that there was no unnecessary word before, and the answer is "Yamada-kun".

In the service scenario S01, it is assumed that the place name word is first input (step S41 in FIG. 4) first, and then the person name word is input (step S42 in FIG. 4) (sequence in FIG. 4). S01). Grammar control unit 1
40 transmits grammar control information to the general-purpose WS grammar dictionary 132 based on the information from the service application. Based on the grammar control information, the selection means 141 changes the path permitted for transition from the prefix unnecessary word set 31 in the general-purpose WS grammar dictionary 132 through the WS target word 32 to the suffix unnecessary word set 33.

The specializing means 142 is used for the changing method. For example, when it is confirmed that the service pattern is input of a place name, the specializing means 142 selects the prefix unnecessary word "aa" and the suffix unnecessary word "san", "manager", "chief", and "kun". Substitute infinity for the initial value of the cumulative distance calculation table and remove it from the calculation practically (step B1 and step B in FIG.
2) Or, look at the flags of the table that manages the execution of the cumulative distance calculation set by the selection means 141, and those that are not flagged are removed from the calculation loop as being ungrammatical (step B2 in FIG. 5, FIG. Step C of 6
2, step C3, step C4).

As a result, the input of a place name word takes the form of a continuous grammar as shown in FIG. 8, and the input of a person's name word takes the form of a continuous grammar as shown in FIG. For each voice input, the recognition process is performed using the continuous grammar shown in FIGS.

According to the first embodiment of the present invention, if the dictionary corresponding to the assumed service scenario is registered in the voice recognition device only once, the service can be advanced without re-registering. There is an effect that the recognition rate can be expected to be improved in the assumed service scenario, and the recognition using the general-purpose word spotting grammar dictionary can be efficiently switched.

[Second Embodiment] Next, a second embodiment of the present invention will be described in detail with reference to the drawings.

(1) Description of Configuration As shown in FIG. 1, the speech recognition apparatus according to the second embodiment of the present invention has a speech input section 100 and a speech analysis section 110.
A distance calculation unit 120, a data unit 130 having a standard pattern dictionary 131 and a general-purpose word spotting grammar dictionary 132, and a grammar control unit 140 having a selection unit 141 and a specialization unit 142. Since each part has been described in the first embodiment, it will be omitted.

In the second embodiment of the present invention, it is possible to similarly register all keyword words at the beginning and remove them later. Referring to FIG.
The second embodiment differs from the general-purpose WS grammar dictionary shown in FIG. 7 of the first embodiment in that the keyword target word group is further divided into a plurality of groups. The divided keyword target word group is the grammar control unit 1 of FIG.
It is controllable by 40 and is classified into the optimum keyword group for the service scenario.

(2) Description of Operation Next, the operation of the second embodiment of the present invention will be described with reference to FIG.
This will be described in detail with reference to FIGS. 7 and 9 and FIGS.

As shown in FIG. 10, for example, the prefix unnecessary word candidates 1 to m in the general-purpose word spotting (WS) grammar dictionary 132 are "a", "a", "etto", and "e". , “Uu”, “that” or the null word φ is registered, and “kara” is added to the postfix unnecessary word candidates 1 to n.
"Up", "is", "san", "manager", "chief",
It is assumed that “kun” or the null word φ is registered. However, the null word here means a word that undergoes state transition without output of symbols.

Further, the keyword target word groups 62 and 63 store keyword word groups A and B (for example, two cases) classified into optimum groups for each service. In the service scenario S01, it is assumed that the place name word is mainly input first, and then the person name word is input.

The grammar control section 140 transmits the grammar control information to the general-purpose WS grammar dictionary 132 based on the information from the service application. General-purpose WS based on the grammar control information
The path permitted for transition from the prefix unnecessary word set 31 in the grammar dictionary 132 through the WS target word 32 to the suffix unnecessary word set 33 is changed. At this time, the WS target word 32 is shown in FIG.
0, 62 and 63 are pre-classified into a plurality of word groups, and therefore, the paths allowed for the transition from the prefix unnecessary word set 31 to the suffix unnecessary word set 33 are in the keyword spotting 62, 63. Therefore, the classification will be further restricted.

As a result, FIG. 11 shows the case of inputting a place name word.
In the case of inputting a personal name word, the continuous grammar as shown in FIG. For each voice input, the optimum recognition processing is performed using the continuous grammar shown in FIGS.

According to the second embodiment of the present invention, as shown in FIG. 10, a keyword word candidate is extracted from a smaller number of candidates by forming a plurality of keyword spotting target word groups. Can
The new effect of raising the recognition rate further occurs.

[0043]

As described above, according to the present invention, the following effects can be obtained.

The first effect is that the possible unnecessary word candidates are registered at a time and divided into a small set in the scene of the service for use, so that the dictionary corresponding to the expected service scenario is used in the voice recognition device. If you register only once, you can proceed with the service without re-registering.

The second effect is that if all possible unnecessary words are registered, it is expected that source errors due to other words will occur frequently. However, as in the present invention, a set of unnecessary words is generated. By switching the timely,
Since unnecessary recognition candidates can be eliminated, the recognition rate can be expected to improve in the assumed service scenario.

The third effect is that by registering frequently used grammar control information as a pattern in advance, recognition using word spotting can be efficiently switched.

[Brief description of drawings]

FIG. 1 is a block diagram showing a configuration example of a voice recognition device according to first and second embodiments of the present invention.

FIG. 2 is a flowchart showing a voice recognition process according to the first embodiment of the present invention.

FIG. 3 is a flowchart showing a voice recognition process according to the first embodiment of the present invention.

FIG. 4 is a flowchart showing a place name / person name input process according to the first embodiment of the present invention.

FIG. 5 is a flowchart showing a cumulative distance calculation process according to the first embodiment of the present invention.

FIG. 6 is a flowchart showing a cumulative distance calculation process according to the first embodiment of the present invention.

FIG. 7 is an explanatory diagram showing a specific example of voice recognition according to the first embodiment of the present invention.

FIG. 8 is an explanatory diagram showing a specific example of voice recognition according to the first embodiment of the present invention.

FIG. 9 is an explanatory diagram showing a specific example of voice recognition according to the first embodiment of the present invention.

FIG. 10 is an explanatory diagram showing a specific example of voice recognition according to the second embodiment of the present invention.

FIG. 11 is an explanatory diagram showing a specific example of voice recognition according to the second embodiment of the present invention.

FIG. 12 is an explanatory diagram showing a specific example of voice recognition according to the second embodiment of the present invention.

[Explanation of symbols]

100 voice input section 110 Voice analysis unit 120 distance calculator 130 data section 131 Standard Pattern Dictionary 132 General-purpose word spotting grammar dictionary 140 Grammar Control Unit 141 sorting means 142 Specialized means

Claims (4)

(57) [Claims] 1. A voice recognition device for performing voice recognition by removing unnecessary words added before and after a voice recognition target word, and is expected to be used with a standard pattern registration means in which a standard pattern of voice information is registered. Non-word registration means in which a continuous grammar holding all prefix and postfix non-words is registered, voice recognition means for performing voice recognition by referring to the standard pattern registration means and the non-word registration means, The search space of the non-word registration means by the voice recognition means is matched to the service scenario.
And control means for reducing Align Te Optionally, said control means, based on pre-registered grammatical control information
Change the path transition is permitted in the previous said rear置不term set as the speech recognition target word group from置不term set, when speech recognition by said speech recognition means, wherein the pre that is registered A voice recognition device characterized by reducing candidates for non-words and post-words. 2. The speech recognition target word group is a speech recognition request.
The speech recognition apparatus according to claim 1, comprising a plurality of speech recognition target word groups classified into optimum groups for each . 3. The control means appropriately reduces the search space of the non-word registration means by the voice recognition means according to a service scenario based on noun word input such as place name word input and person name word input. The voice recognition device according to claim 1. 4. A voice recognition method for performing voice recognition by removing unnecessary words added before and after a word to be voice-recognized, and all standard patterns of voice information registered in advance and use registered in advance are assumed. of pre when performing speech recognition on the basis of non-term and continuous grammar holding the rear置不terms, the front for置不terms and rear置不 said registered use is contemplated
It is appropriately scaled down to fit the search space of the word to the service scenario
In order, by changing the path transition is permitted in the previous said rear置不term set as the speech recognition target word group from置不term set, when the voice recognition, before the registered置不A speech recognition method characterized by reducing candidates for terms and postfix non-words.