JPH0950290A - Voice recognition device and communication device using it - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a voice recognition device having an expansibility without being accompanied by increase in a processing amount and improving a recognition rate, and to provide a communication device using it. SOLUTION: In the voice recognition device, a vocal state detection means 7, an LR table means 4 being plural voice recognition grammar corresponding to respective vocal states and voice recognition means 5, 6 using the LR table means 4 and voice recognizing next vocalization are included. Since the voice recognition is performed using the voice recognition grammar corresponding to the next estimated vocal state, the processing amount is reduced when compared with the case where general grammar is used, and since useless grammar is not included, the recognition rate is improved. Further, an application range is expanded to conversation for an optional business depending on the method of the definition of the vocalizing state.

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 communication device using the device, and more particularly to a voice recognition device capable of improving voice recognition accuracy and a communication device using the device.

[0002]

2. Description of the Related Art Conventionally, in the field of voice recognition, for example, JP-A-2-113297 or JP-A-4-18.
HMM-LR as disclosed in Japanese Patent Publication No. 2000
A law has been proposed. The HMM-LR method uses LR (Lef
The phoneme in the input speech data is predicted by a syntactic analysis method called “t to Right” method, and the likelihood of the predicted phoneme is calculated by HMM (Hidden Markov).
This is a method of simultaneously advancing speech recognition and language processing by investigating by a phoneme recognition method called Model method.

[0003]

When the conventional HMM-LR method is used as described above, it is necessary to generate an LR table from a large amount of general grammar in order to recognize general dialogue. Therefore, there is a problem that the processing time for recognition increases and the recognition rate also decreases. An object of the present invention is to solve the above-mentioned problems of the prior art, and to provide a speech recognition apparatus that has expandability and an improved recognition rate without increasing the processing amount, and a communication apparatus using the apparatus. To do.

[0004]

According to the present invention, in a voice recognition device, there is provided utterance state detecting means for detecting which one of predefined utterance states the recognized dialogue belongs to, and corresponding to each utterance state. , Speech recognition for performing speech recognition of the next utterance using a plurality of speech recognition grammars generated from recognition rules that can be used in the next utterance and a speech recognition grammar corresponding to the utterance state output by the utterance state detecting means. And means. Further, a communication device using the device is also characterized.

According to the present invention, since the phoneme recognition is performed by using the voice recognition grammar corresponding to the next predicted utterance state, the processing amount is reduced as compared with the case of using a general grammar. In addition, the recognition rate is improved because unnecessary grammatical rules are not included. Further, depending on how to define the utterance state, it is possible to extend the scope of application to dialogue for any business. Further, since the communication device using the device has a small amount of processing, it becomes possible to perform a dialogue in a more real time manner, and the communication cost is reduced.

[0006]

DETAILED DESCRIPTION OF THE INVENTION The present invention will be described in detail below with reference to the drawings. FIG. 1 is a block diagram showing the functions of a voice recognition device to which the present invention is applied. Reference numeral 1 is grammatical rule data for voice recognition, which will be described later, and 2 is utterance state transition matrix data, which will also be described later. The LR table compiler 3 extracts a voice recognition rule corresponding to each utterance state based on the recognition grammar 1 and the utterance state transition matrix 2 and generates an LR table 4 for each. Note that these processes are performed in advance prior to voice recognition.

The voice recognition unit 6 predicts a phoneme in the input voice data by a syntax analysis method called LR method,
A method of simultaneously advancing speech recognition and language processing is adopted by checking the likelihood of the predicted phoneme by a phoneme recognition method called HMM method, and speech recognition is performed using the LR table selected by the table selection unit 5. I do. Specifically, a frame synchronization HM different from the conventional phoneme synchronization method
It uses the M method. In this frame synchronization method, pronunciation data is divided into frames at predetermined time intervals, and phonemes of each frame are predicted and collated. Based on the output result of the voice recognition unit 6, the utterance state detection unit 7 refers to the utterance state information added to the recognition grammar rule applied to the recognition, the utterance state information of the conversation partner input from the outside, and the like. Determine the utterance status. The table selection unit 5 selects the corresponding LR table based on the utterance state information output from the utterance state detection unit 7.

FIG. 3 is an explanatory diagram showing an example of the contents of the utterance state transition matrix of FIG. In addition, in the following description, as a specific job, for example, a job related to a hotel room reservation will be described. As the utterance state in such work, the information of the utterer is effective first, and for example, the hotel side is represented by "H" and the user side is represented by "C". Next, as the utterance type, for example, the following utterance types can be considered. "GO" (greeting at the beginning of the dialogue), "G
"C" (greeting at the end of the dialogue), "QP" (yes / no question), "QW" (question using question words such as What, When, etc.), "RP" (affirmative response), ""RN" (negative response), "CF" (confirmation), "EX" (thanks). The utterance type has little dependence on work and is versatile. In the present embodiment, the utterance state is expressed by combining the speaker and the utterance type.

For example, the current utterance state is "H / GO",
That is, in the case of a greeting at the beginning of the dialogue on the hotel side, referring to the first line of the matrix in FIG. 3, the utterance state that may have the next transition is 1 in the matrix. State, ie "C / GO" (greeting at the beginning of the user's dialogue) and "C / QP"(user's yes /
It turns out that it is a (no question sentence). Therefore, if the current utterance state is "H / GO", only the grammatical rules that may be used in the "C / GO" and "C / QP" states in the next speech recognition. It turns out that it is sufficient to use the LR table generated from

FIG. 4 is an explanatory diagram showing an example of the contents of the recognition grammar data 1 of FIG. The recognition grammar is described in the form of a branch-free grammar. Each line represents one grammar rule, and each rule is provided with a header. The header is composed of four-digit characters, and the first digit represents speaker information and the second and third digits represent speech type information. It should be noted that "-" is unlimited, that is, it is applied to any speaker or utterance type. The recognition grammar 1 is
By adding a header to the grammar rule, it can be managed as one database. Therefore, for example, when adding a general-purpose grammar rule, by adding an unlimited header, it is reflected in the grammar in all utterance states, which is more efficient than maintaining the grammar for each utterance state. Can be managed and updated.

FIG. 5A shows the table selection unit 5 of FIG.
6 is a flowchart showing processing in the voice recognition unit 6 and the utterance state detection unit 7. In step S1, the initial value of the utterance state at the start of the dialogue is set in the utterance state detecting unit 7. In step S2, the table selection unit 5
Selects the corresponding LR table based on the utterance state information input from the utterance state detection unit. In step S3, the voice recognition unit 6 inputs the preprocessed voice data, and in step S4, performs voice recognition using the LR table selected in step S2.

In step S5, a predetermined number of candidate sentences obtained by the recognition process are displayed on, for example, a display device, and the speaker selects the sentence using a pointing device such as a mouse to confirm the confirmed recognition result. To get In step S6, for example, by checking whether or not the speaker has performed a dialogue ending operation, it is determined whether or not the dialogue has ended. If the result is negative, the process proceeds to step S7.

In step S7, the utterance state is detected based on the recognition result. In the detection of the utterance state, for example, when the recognition result is confirmed, the recognition grammar applied to obtain the recognition result is known. The recognition grammar also includes top-level rules regarding the structure of sentences, each of which has utterance state information added as a header. Therefore, the utterance state can be detected from the header added to the rule regarding the structure of the sentence used at the time of recognition. It is also possible to detect the utterance state from the header of a specific phrase included in the sentence, and both data may be referenced.

In step S8, for example, as in the automatic translation telephone system described later, this voice recognition system is in charge of voice recognition of only one speaker, and the other speaker is recognized by another voice recognition system. When the resulting text data is received, the utterance state of the other party is detected by referring to the recognition dictionary from the specific phrase in the received text information from the other party. Step S9
In step S7, the current utterance state is determined based on the information about the utterance state obtained in steps S7 and S8, and the process returns to step S2 to perform voice recognition of the next utterance.

FIG. 5B is a flowchart showing the LR table generation process. However, steps S20 to S22
The processing of is performed by a person. In step S20, recognition grammar data is created from general grammatical rules used in an arbitrary job and grammatical rules such as words and phrases used in a specific job. It should be noted that the grammar rules include a plurality of levels of rules ranging from a sentence structure rule to a phoneme arrangement rule (pronunciation information for each word). In step S21, FIG.
Header information is added as shown in. In step S22, the utterance state to be used is determined, and the transition matrix between the states as shown in FIG. 3 is determined.

In step S23, the grammar subset corresponding to the utterance state is extracted by referring to the header information given to the grammar rule based on the transition matrix. For example, if the current utterance state is "H / GO",
Since it is found from the transition matrix that the next utterance state is either "C / GO" or "C / QP", referring to the header information, the grammar that may be used in the utterance state Extract only rules.

In step S24, an LR table corresponding to each utterance state is generated from each of the extracted subsets. The LR table is, for example, data of a multistage tree structure in which all phonemes that may possibly follow a certain phoneme are connected. Since the LR table generated in this way is composed of only the grammar necessary for the corresponding utterance state, the number of data is smaller than that of the general-purpose LR table, and the recognition processing time is shortened and unnecessary. Since the data is not included, the risk of erroneous recognition is reduced,
The recognition rate is improved.

Next, an automatic translation telephone device using the voice recognition device of the first embodiment will be described. FIG. 2 is a block diagram of an automatic translation telephone device to which the present invention is applied. For example, the automatic translation telephone devices 10 and 12 configured by workstations or the like are connected via an arbitrary communication network 11 such as a telephone network, a data network, a private line, and a LAN. The voice signal input from the microphone 20 is subjected to acoustic analysis (VQ) of the voice signal in the pre-processing unit 21, and the voice signal is input to the voice recognition unit 22 having the same function as the voice recognition device of the first embodiment shown in FIG. Is entered. Speech recognition unit 2
In No. 2, voice recognition is performed by the same processing as that of the first embodiment, and if necessary, a speaker uses a display device and an input device such as a mouse (not shown) to select from among a plurality of candidates of the recognition result. Let the correct answer be selected.

The voice recognition unit 22 outputs to the translation unit 23 the text information of the recognition result and the information about the structure of the sentence and the word breaks found at the time of recognition. The translation unit 23 translates the text information into another language while referring to the translation dictionary based on the input information, and outputs the translated text information. The communication management unit 24 interfaces with the communication network 11 and the partner terminal 12 and transfers text data. As the communication network, any communication network can be used as long as it can communicate text data with each other, and the data transfer rate is very low compared to voice data.

The translated text data transferred from the partner terminal is received by the communication management unit 24 and output to the voice synthesis unit 26 and the utterance state detection unit 25. The voice synthesizer 26 converts the text information into a voice signal, and the speaker 27 produces a sound. Also, the utterance state detection unit 25
Cuts out a word from the text information, matches it with a grammatical rule (word dictionary) in the recognition grammar, and detects the utterance state of the other speaker from the header information of the corresponding rule. This information is
The speech state is input to the speech state detection unit 7 in the voice recognition unit 22, and the speech state is determined. Alternatively, the utterance state information detected by the voice recognition unit 22 may be transmitted to the other terminal together with the text information, and the utterance state detection unit 25 may extract the information. With the above configuration, the speaker can talk with the foreign conversation partner in almost real time.

Although the embodiment has been disclosed above, the following modifications are also conceivable. I disclosed an example of preparing the LR table in advance, but each time the utterance state changes,
The LR table compiler may be used to generate the LR table corresponding to the utterance state. Although the example in which the content of the utterance state transition matrix is 0 or 1 is disclosed,
The transition probability may be represented by a weighting coefficient of three values or more, and the coefficient may be used to calculate the likelihood at the time of phoneme determination.

Although the HMM method has been disclosed as an example of the phoneme matching method, it is not limited to the method of the embodiment, and any phoneme matching method can be used. Although the example of the LR method has been disclosed as the language analysis method, any language analysis method is not limited to the method of the embodiment as long as it is a language analysis method that uses a knowledge source of language constraints such as speech recognition grammar. It can be used. Regarding the utterance state, the example using the combination of the speaker and the utterance type is disclosed, but the present invention is not limited to this, and other information such as the type of topic (topic) and attributes can be adopted (combined). is there.

Although an example of an automatic translation telephone device has been disclosed as a communication device using a voice recognition device, the device configuration can be applied to a video telephone device. In addition, the device may be a device that conducts a dialogue using text data, except for the translating means, which enables a dialogue using a very narrow band transmission channel. Further, it may be a device that displays the voice of the telephone in characters for the hearing impaired. In the automatic translation telephone device, it is possible to share grammar (dictionary) data for voice recognition and translation.
For example, for a Japanese vocabulary heading, its reading, syntactic /
Semantic attributes, bilingual foreign languages, and the like may be described in a table format, and vocabulary-level grammatical rules for voice recognition may be generated using the reading and syntactic information in the table format. By doing so, the dictionary can be efficiently generated and updated.

[0024]

As described above, according to the present invention,
The speech recognition apparatus includes a speech state detection unit, a plurality of speech recognition grammars corresponding to each speech state, and a speech recognition unit that performs speech recognition of the next speech by using the grammar, and is predicted next. The phoneme recognition is performed using the speech recognition grammar corresponding to the utterance state, so the processing amount is reduced compared to the case of using a general grammar, and the unnecessary grammar is not included, so the recognition rate is improved. To do. Further, there is an effect that the application range can be expanded to a dialogue for any business depending on how the utterance state is defined. Further, since a communication device such as an automatic translation telephone device using the device has a small processing amount, it is possible to have a dialogue closer to a real time and to reduce communication cost.

[Brief description of drawings]

FIG. 1 is a block diagram showing functions of a voice recognition device to which the present invention is applied.

FIG. 2 is a block diagram of an automatic translation telephone device to which the present invention is applied.

FIG. 3 is an explanatory diagram showing an example of contents of an utterance state transition matrix 2.

FIG. 4 is an explanatory diagram showing an example of contents of recognition grammar data 1.

FIG. 5 is a flowchart showing voice recognition and LR table generation processing.

[Explanation of symbols]

1 ... Recognition grammar, 2 ... Utterance state transition matrix, 3 ... L
R table compiler, 4 ... LR table by state, 5 ...
Table selection unit, 6 ... Voice recognition unit, 7 ... Speech state detection unit

Claims (5)

[Claims] 1. A utterance state detecting means for detecting which one of predefined utterance states the recognized pronunciation belongs to, and a speech recognition rule which can be used in the next utterance corresponding to each utterance state. It is characterized by including a plurality of speech recognition grammars generated from the set and a speech recognition means for performing speech recognition of the next utterance using the speech recognition grammar corresponding to the utterance state output by the utterance state detection means. Speech recognizer. 2. The utterance state information is added to at least a part of the speech recognition rules, and the speech recognition grammar corresponding to each utterance state is based on the utterance state transition information, and the next utterance is recognized from the entire speech recognition grammar. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus is created by extracting a grammar that can be used in. 3. The utterance state detecting means, based on the utterance state information added to the voice recognition grammar applied to the voice recognition,
The voice recognition device according to claim 2, wherein the current utterance state is detected. 4. A communication device including the voice recognition device according to claim 1. 5. The speech recognition device further includes a translation unit that translates the text information recognized by the speech recognition device into another language, and a speech synthesis unit that converts the received text information into a speech signal. The communication device according to claim 4, wherein the translation is performed based on the syntax information found at the time of recognition.