JPH0314359B2 - - Google Patents

Description

[Detailed Description of the Invention] [Technical Field of the Invention] The present invention has a learning function and automatically redisplays the same prompt when there are rejects, similar words, or variations in pronunciation. The present invention relates to a user voice recognition device.

[Conventional technology and problems]

In speech recognition devices, one of the important factors for increasing the recognition rate is optimizing dictionary patterns to be compared and matched with input speech patterns.

The specific speaker speech recognition device has a learning function for optimizing this dictionary pattern, and the operator makes the device learn by uttering the same word several times during voice registration. With the advanced learning function, multiple dictionary templates are prepared for one word, and you can modify (average), add, or delete the dictionary.
Make additions.

There are two modes for voice registration: initial registration mode and learning mode. Initial registration involves uttering all the registered words once to create an initial pattern for the dictionary. Learning is performed after initial registration, that is, after initial registration. - Performed after a pattern is created, recognition processing is executed for the utterance of a certain word, and based on the ranking and distance information of the recognition result, rejects, corrections, or additions to the dictionary pattern are performed. As learning progresses further, dictionary patterns are deleted/added.

In voice registration, it is desirable for the operator to efficiently complete the dictionary with as few utterances as possible, and it is also desirable for the operator to avoid unnecessary key operations and judgments and concentrate on utterance.

[Purpose of the invention]

The present invention is based on the above considerations, and includes:
A specific speaker speech recognition device with a learning function automatically displays the next word to be uttered to the operator in sequence, minimizing operator judgment and key operations, thereby reducing the burden on the operator. The purpose of this invention is to provide a voice recognition device for users.

[Structure of the invention]

To this end, the specific speaker speech recognition device of the present invention performs recognition by comparing input patterns and dictionary patterns with a dictionary memory in which character strings of registered words are stored and dictionary patterns of speech for registered words are registered. A matching unit that outputs the results and input pattern correction information, a display, a dictionary control correction unit that performs additional registration and correction of dictionary patterns in the dictionary memory, and a recognition result and input pattern correction information output from the matching unit. a learning mechanism unit into which is input, and a prompt sending control unit which reads out the character string of the registered word from the dictionary memory and displays on the display a prompt requesting voice input of the registered word according to instructions from the learning mechanism unit. The learning mechanism section controls the dictionary control and correction section to perform processes such as rejecting, adding dictionary patterns, and correcting dictionary patterns, and is necessary for re-voicing. If there is no need to re-speak, control is performed to display a prompt on the display requesting voice input of the next registered word, and if re-speak is necessary, the same voice input is performed. It is characterized by being configured to perform control to display on the display again a prompt requesting voice input of the registered word, and to perform control to display on the display auxiliary information indicating the reason why re-voicing is necessary. That is.

[Embodiments of the invention]

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

The figure is a diagram showing the configuration of one embodiment of the present invention. In the figure, 1 is an input pattern buffer, 2 is a dictionary control correction section, 3 is a dictionary memory, 4 is a collation section, and 5 is a dictionary control correction section.
6 is a prompt sending control section, 6 is a learning mechanism section, and 7 is a display. Input pattern buffer 1
stores the input patterns voice input by the operator, and the collation unit 4 collates the input patterns stored in the input pattern buffer 1 with the dictionary patterns stored in the dictionary memory 3, and performs learning. In the mode, the recognition results (a plurality of candidates and their distances) and input pattern modification information are sent to the learning mechanism section 6 as the matching results. The dictionary memory 3 consists of a table and a dictionary pattern. The table stores character strings of registered words, and the dictionary pattern stores a plurality of dictionary templates for registered words and registers speech patterns. In initial registration, all the registered words in the table are uttered once, and the patterns are registered in the dictionary pattern. During learning, the input pattern uttered for the registered words is compared with the dictionary pattern, and the matching is performed. Depending on the results, rejects, dictionary patterns are added, and dictionary patterns are modified. These processes for the dictionary memory 6 are carried out by the learning mechanism section 6.
This is executed by the dictionary control correction unit 2 based on instructions from the dictionary control correction unit 2. Learning mechanism part 6 is correct (uttered word)
As described above, the dictionary control correction unit 2 is controlled to reject, add a dictionary pattern, or modify a dictionary pattern based on the distance information of , the distance difference between the correct answer and other words, and the dictionary template free information. Perform one of the following corrections. Rejecting is a process of not performing learning, adding a dictionary pattern is a process of using an input pattern as an additional dictionary pattern, and modifying a dictionary pattern is a process of averaging and replacing dictionary patterns. Furthermore, in relation to this learning result, the learning mechanism unit 6 performs processing for displaying on the display 7 what the operator should say next. The learning mechanism unit 6 causes the prompt sending control unit 5 to sequentially display the character strings in the table in the dictionary memory 3 on the display 7, but if, for example, the following cases occur, the prompt is required to be re-voiced. Display the same prompt again without proceeding. () When it is rejected (when it is not considered normal audio input).

When the correct answer is not recognized in the first place (similar words exist).

Even if the correct answer is recognized as the first candidate, the distance to the second candidate is close, that is, when similar words exist (similar words exist).

Even if the correct answer is recognized in first place, when the distance is greater than a certain threshold, that is, a word whose pronunciation tends to vary (there is variation in pronunciation).

Furthermore, the learning mechanism section 5 causes the display 7 to display the following information as auxiliary information in order to notify the operator why re-voicing is necessary. ().

Candidates ranked up to about 4th place in recognition results.

Information on whether there is a redirect, a similar word exists, or a variation in pronunciation.

〔Effect of the invention〕

As is clear from the above description, according to the present invention, in a specific speaker speech recognition device having a learning function, in the process of learning, how far (how many times) should the recognition device itself utter? Since the system automatically determines and displays prompts, the operator can simply follow the prompts and speak without making unnecessary judgments or key operations, reducing the operator's burden and improving learning efficiency. can.

[Brief explanation of the drawing]

The figure shows the configuration of one embodiment of the present invention. 1...Input parameter buffer, 2...Dictionary control correction section, 3...Dictionary memory, 4...Verification section,
5... Prompt sending control section, 6... Learning mechanism section, 7... Display.

Claims (1)

[Scope of Claims] 1. Correcting the recognition result and input pattern by comparing the input pattern and the dictionary pattern with a dictionary memory in which character strings of registered words are stored and dictionary patterns of speech for the registered words are registered. A collation unit that outputs information, a display, a dictionary control correction unit that additionally registers and corrects dictionary patterns in the dictionary memory, and a learning mechanism that receives recognition results output from the collation unit and input pattern correction information. and a prompt sending control unit that reads the character string of the registered word from the dictionary memory and displays on the display a prompt requesting voice input of the registered word according to instructions from the learning mechanism unit. In the recognition device, the learning mechanism section controls the dictionary control correction section to perform processing such as rejecting, adding a dictionary pattern, and correcting the dictionary pattern, and determines whether re-voicing is necessary. , If there is no need to re-speak, a prompt is displayed on the display to request voice input of the next registered word, and if re-voice is required, the voice input of the same registered word is performed. A specific speaker's voice characterized by being configured to perform control to display the requested prompt on a display again, and to perform control to display on the display auxiliary information indicating the reason why the re-speech is necessary. recognition device. 2. The specific speaker speech recognition device according to claim 1, wherein the auxiliary information includes recognition results from the first rank to the Nth rank (N is an integer of 2 or more).