JP2001134285A - Speech recognition device - Google Patents

Abstract

PROBLEM TO BE SOLVED: To solve such problem that a conventional language model has increased an erroneous recognition and has been also unable to recognize such expression and phrasing which have not been learned by the model beforehand because a single language model is used disregarding the other party, situation, contents, or the like of a speaker. SOLUTION: According to the contents which a user intends to speak and the other party to whom the user intends to speak, a learning data selecting part 1 selects a language model or a corpus to be used in a speech recognition part 5 from a learning data storage part 2, and after it has been converted into a format to be used in the speech recognition part 5 by a learning data conversion part 3, the speech recognition part 5 performs speech recognition by using the language model suitable for the utterance contents and the other party.

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a dialogue device for outputting characters, images, voice data, and the like using voice as input.

[0002]

2. Description of the Related Art A conventional speech recognition technique will be described below.

Conventionally, in learning a language model for speech recognition, a large-scale database of written words, such as newspaper articles, which is generally easily available, has been used. However, such a language model learned from written words cannot sufficiently learn the free utterance expression of the target user by speech recognition, so that a large-scale written word database must be used.
They collected and added a database of small-scale spoken words obtained from a collection of dialogues, etc., and made them learn so that they could respond to free speech.

Further, in order to improve the recognition performance of the user in speech recognition, the acoustic model is adapted by learning the acoustic characteristics of the utterance of the user.

[0005]

There is no particular criterion for selecting a spoken language database to be added to a large-scale written language database, and since it has been collected and added as much as possible, the task contents of the spoken language database (for example, No particular consideration was given as to whether it was a conversation during a meeting, a conversation at a restaurant, whether the speaker was a customer, a clerk, etc.).

However, experiments have confirmed that adding a spoken language database does not always correspond to free speech, that is, does not directly lead to an improvement in recognition rate.

FIG. 12 shows the results of measuring the Perplexity for several types of language models trained on various databases. Here, the perplexity is a measure indicating the ease of predicting the next word, and it is generally said that a language model having a lower value of perplexity has better performance.

As a large-scale written language database, the Mainichi Shimbun Article 95 version (described as daily DB in FIG. 12), and a small spoken language database which is considered to have different task contents due to different speakers, and a clerk at a hotel.
And customer conversation (in Figure 12, clerk and custo respectively)
clerk, cust for Mainichi Shimbun DB
When a language model was learned by adding a database of each omer's conversation, the value of Perplexity for the evaluation data of each clerk's and customer's utterances was measured.

For the evaluation data of clerk, the spoken language database to be added is clerk or customer,
In both cases, as the value increases, the value of Perplexity decreases, suggesting the effect of adaptation by increasing the spoken language database.However, with regard to customer evaluation data, the value of Perplexity is the highest when the customer's spoken language database is added. Conversely, adding a clerk spoken language database (that is, a spoken language database with different task content) to it will increase the value of Perplexity. From this, it is not considered that adding a spoken language database and learning a language model is necessarily adaptation to a user's free speech expression.

[0010] Expressions that are easy for the user to speak vary, but the language model provided to the user does not reflect the ease of speaking of expressions that are unique to each user. The adaptation of the acoustic model by learning the features alone cannot sufficiently compensate for the user's recognition rate.

In view of the above problems, the present invention changes the type and number of spoken language databases to be added to a large-scale written language database during language model learning, or changes the language model itself in accordance with the task contents spoken by the user. To provide a language model that adapts to the user by adding a new spoken language database from the history of user-specific expressions in order to change the language and reflect the utterance expressions that are diverse for each user in the language model. Is what you do.

[0012]

In order to solve the above-mentioned problems, a first aspect of the present invention (corresponding to claim 1) is to provide a voice input unit for inputting voice and learning data for storing language data. A storage unit, a learning data selection unit for selecting language data stored in the learning data storage unit, and learning for statistically processing the language data selected by the learning data selection unit or converting the data format A data conversion unit, a voice recognition unit that recognizes a voice input to the voice input unit using data processed by the learning data conversion unit, and a character or voice based on a recognition result of the voice recognition unit. Or a data output unit for outputting an image or data obtained by combining the images.

Further, a second aspect of the present invention (corresponding to claim 2)
2. The learning data selection unit according to claim 1, wherein the user selects one or a plurality of language data from a plurality of language data stored in the learning data storage unit according to a user's utterance purpose. It is a speech recognition device of the description.

A third aspect of the present invention (corresponding to claim 3)
2. The voice according to claim 1, wherein the learning data selection unit automatically selects from a plurality of language data stored in the learning data storage unit in accordance with a user's utterance purpose. It is a recognition device.

A fourth aspect of the present invention (corresponding to claim 4)
The speech recognition apparatus according to claim 1, wherein the language data described in claim 2 or 3 is a connection frequency or a connection probability of a basic unit constituting the language.

A fifth aspect of the present invention (corresponding to claim 5)
The language data according to claim 2 or 3 is a word string, and the learning data conversion unit according to claim 1 converts the word string selected by the method according to claim 2 or 3 into a language string. 2. The speech recognition apparatus according to claim 1, wherein the speech recognition apparatus converts the connection frequency into a connection frequency or a connection probability.

The sixth invention (corresponding to claim 6)
Is a speech recognition apparatus according to claim 1, wherein the language data according to claim 2 or 3 is classified in units of utterance situations.

The seventh invention (corresponding to claim 7) is:
The language data stored in the learning data storage unit is
A speech recognition apparatus is characterized by being created from a language data input unit for inputting a character string and a language processing unit for processing the input character string.

An eighth aspect of the present invention (corresponding to claim 8) is:
The language data input unit is a voice recognition device that explicitly requests a user to input a character string.

A ninth aspect of the present invention (corresponding to claim 9) is:
The language data input unit is a voice recognition device that extracts a partial character string of the character string from a character string input by a user without making the user conscious.

[0021]

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

First, the entire speech recognition apparatus will be described with reference to FIG.

Before the user makes an utterance whose utterance changes according to the task or purpose, the learning data selecting unit 1 selects language data to be used by the speech recognition unit 5 from the learning data storage unit 2. The learning data conversion unit 3 receives the selected data from the learning data storage unit 2, determines the format of the data, and converts the data into a usable format if different from the format used by the speech recognition unit 5. The conversion of the data format will be described later.

The voice recognition unit 5 loads data output from the data output unit 6.

When the user speaks, the speech content is input to the voice input unit 4. The voice recognition unit 5 recognizes the input voice based on the loaded data, and determines an output for the voice. The data output unit 6 determines the voice, character, image, and other data according to the determination of the voice recognition unit 5. Is output.

In the following first to fourth embodiments, dictation of electronic mail is assumed.

(First Embodiment) In the first embodiment, the language model shown in FIG. 2 is selected as the language data stored in the learning data storage unit 2 in FIG. It is assumed that the user is the subject to perform.

This embodiment will be described with reference to FIG.

The user selects a language model suitable for the content of the utterance from the contents of the e-mail or the transmission destination, and inputs the information to the selection data input unit 21. Based on the input to the selection data input unit 21, the corresponding language model is sent from the learning data storage unit 2 to the learning data conversion unit 3. Voice recognition unit 5
Loads the language model output from the learning data converter 3.

(Second Embodiment) In the second embodiment, the language data stored in the learning data storage unit 2 in FIG. 1 is used as the corpus shown in FIG. It is assumed that the selecting unit 1 automatically performs the processing.

It is also assumed that the types of corpus are tagged with a large-scale corpus of newspaper articles, a corpus of conversation at a hospital, and a corpus of conversation between a user and a friend.

This embodiment will be described with reference to FIGS.

When the user tries to send an e-mail in the format shown in FIG. 5, the subject and the recipient are input to the determination condition input unit 51 in FIG. In the selection data determination unit 52, the Subject input to the discrimination condition input unit 51
From the discriminant condition database 53 which stores the concept structure and the relation between the concept and the corpus corresponding thereto based on the recipient and the recipient, the contents of the body of the e-mail to be uttered by the user from now on are predicted. Select the data to use. For example, in the example of FIG. 5, the subject 42 is “amusement park” and the recipient (To 41 in FIG. 5) is “yoshinobu”. From the conceptual diagram shown in FIG. 7, “amusement park” is a “playground” and “yoshinobu” is a “friend”. The selection data determination unit 52 in FIG. 6 determines that the corpus to be selected is the newspaper article corpus and the conversation corpus with friends from FIG. 8 showing the concept and the corpus corresponding thereto.

After receiving the corpus selected by the selection data decision unit 52, the learning data conversion unit 3 performs the morphological analysis shown in FIG. 9 and then counts the frequency of adjacent occurrences between words, phrases, and classes. Is converted into a language model format that can be used by the speech recognition unit 5 as shown in FIG. Here, the class is a group of words and phrases having similar connection relations before and after.

(Third Embodiment) In the third embodiment, it is assumed that language data stored in the learning data storage unit 2 in FIG. 1 is newly created.

This embodiment will be described with reference to FIG.

The user inputs an expression or a character string to be added to the learning data storage unit 2 to the language data input unit 91. In this case, the expression or character string to be added may be, for example, a phrase frequently used by the user or an expression unique to the user.

The language processing unit 92 receives the input character string and converts it into a format stored in the learning data storage unit 2.

For example, if the learning data storage unit 2 stores the data in the form of a language model, the language processing unit 92 performs morphological analysis of the input data, calculates the frequency of adjacent occurrences between words, phrases, and classes. 2 stores the information.

(Fourth Embodiment) In the fourth embodiment, even if the user is not conscious of the input to the language data input section 91, the user can connect to another device or system for another purpose. As a case where an expression or a character string to be input is stored in the learning data storage unit 2, an environment in which a user uses an electronic mail is assumed.

This embodiment will be described with reference to FIG.

The user inputs the body of the electronic mail to the electronic mail software 101 as a character string every day. Language processing unit 9
2 receives the character string input to the e-mail software 101 and performs the same processing as in the third embodiment.

[0043]

As is apparent from the above description, according to the present invention, the speech recognition unit can use an appropriate language model each time according to the contents and purpose of the utterance of the user. An improvement in performance can be expected.

Also, by storing expressions unique to the user and phrases often used by the user, it is possible to recognize expressions that are not present in the language model that has been learned in advance, and to recognize the frequently used expressions. Can be improved.

[Brief description of the drawings]

FIG. 1 is a configuration diagram of a speech recognition device.

FIG. 2 is a diagram of a format of a language model used in a speech recognition unit.

FIG. 3 is a configuration diagram of a speech recognition device according to the first embodiment.

FIG. 4 is a diagram of a dialog corpus.

FIG. 5 is a diagram of an e-mail transmission template.

FIG. 6 is a configuration diagram of a speech recognition device according to a second embodiment.

FIG. 7 is a diagram showing a tree structure of a concept.

FIG. 8 is a diagram showing a relationship between a concept and a corpus corresponding thereto.

FIG. 9 is a diagram showing a result of morphological analysis.

FIG. 10 is a diagram showing a structure of learning data addition in the third embodiment.

FIG. 11 is a diagram showing an additional structure of learning data by e-mail according to the fourth embodiment.

FIG. 12 is a diagram showing a relationship between a language model and a Perplexity value.

[Explanation of symbols]

 Reference Signs List 1 learning data selection unit 2 learning data storage unit 3 learning data conversion unit 4 voice input unit 5 voice recognition unit 6 data output unit 11 language model 21 selection data input unit 31 dialog corpus 41 recipient 42 Subject 51 determination condition input unit 52 selection Data determination unit 53 Discriminant condition database 61 Recipient's conceptual structure 62 Subject's conceptual structure 71 Concept and corresponding corpus 81 Morphological analysis target sentence 82 Morphological analysis result 91 Language data input unit 92 Language processing unit 101 E-mail software 111 Evaluation data Perplexity for

 ──────────────────────────────────────────────────続 き Continued on the front page (72) Inventor Yumi Wakita 1006 Kazuma Kadoma, Kadoma City, Osaka Matsushita Electric Industrial Co., Ltd. F-term (reference) 5D015 AA01 AA05 HH11

Claims (9)

[Claims] An input unit for inputting a voice; a learning data storage unit for storing language data; a learning data selection unit for selecting language data stored in the learning data storage unit; A learning data conversion unit that statistically processes the language data selected by the data selection unit or converts a data format; and a voice input to the voice input unit using the data processed by the learning data conversion unit. And a data output unit for outputting a character, a voice, an image, or a combination thereof based on the recognition result of the voice recognition unit. . 2. The learning data selection unit, comprising: a plurality of language data stored in the learning data storage unit;
2. The speech recognition device according to claim 1, wherein the user selects one or a plurality of items according to the utterance purpose of the user. 3. The learning data selecting unit, from among a plurality of language data stored in the learning data storage unit,
2. The speech recognition device according to claim 1, wherein the selection is automatically made in accordance with the utterance purpose of the user. 4. The speech recognition apparatus according to claim 1, wherein the language data according to claim 2 or 3 is a connection frequency or a connection probability of a basic unit constituting the language. 5. The language data according to claim 2 or 3, wherein the language data is a word string, and the learning data conversion unit according to claim 1 is a word string selected by the method according to claim 2 or 3. 2. The speech recognition apparatus according to claim 1, wherein is converted into a connection frequency or a connection probability of a language. 6. The speech recognition device according to claim 1, wherein the language data according to claim 2 or 3 is classified in units of utterance situations. 7. The language data stored in the learning data storage unit includes: a language data input unit for inputting a character string;
A speech recognition apparatus, which is created from a language processing unit that processes an input character string. 8. The speech recognition device according to claim 1, wherein the language data input unit explicitly requests a user to input a character string. 9. The speech recognition device according to claim 1, wherein the language data input unit extracts a partial character string of the character string from a character string input by a user without making the user aware.