JPH0362279B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (a) Technical Field of the Invention The present invention relates to a speech recognition device, and more particularly to a speech recognition device that is an improved multi-template method.

(b) Background of the technology Conventional speech recognition devices register the user's voice, compare the registered voice with the input voice, and output the most similar voice as the recognition result. There is a recognition device.

This type of speech recognition device includes a multi-template method that prepares multiple dictionaries for one word, and a further development of the above method by introducing learning processing, but this latter method However, there is still insufficient analysis of the manner of pronunciation and the nature of the words that are created in the dictionary, and as a result, the amount of dictionary capacity that must be prepared is still large. Not only is this not possible, but there is still room for further improvement, such as the recognition time becoming longer.

(c) Prior art and problems In the early stages of the above-mentioned speaker-specific speech recognition devices, dictionaries were developed to deal with variations in pronunciation in order to avoid misrecognition that would occur when one dictionary was prepared for one word. A so-called multi-template speech recognition device has been developed that prepares multiple dictionaries for one word. This device can be expected to be effective if the various sounds that may occur when each word is sufficiently registered. However, in reality, voice registration is often performed one after another, and the utterances during the registration tend to be similar, so that the effect of creating multiple templates is diluted.

Therefore, once the registration is completed, the learning speech obtained by synthesizing the input speech and the registered speech at a specific ratio is re-registered as the registered speech for that word for those whose matching distance between the input speech and the corresponding word in the dictionary is close. This learning process is used to gradually create a dictionary with different pronunciations.

In this way, even if multiple dictionaries are provided for every word, if similar utterances are registered, the effect of multi-template creation as described above will still not be achieved. .

Furthermore, some words are easy to pronounce and are difficult to produce variations in, and some words have no similar words and are difficult to misrecognize. There is little need to perform multi-template conversion as described above for such words.

(iv) Purpose of the Invention The present invention was devised in view of the drawbacks of the conventional devices as described above, and its purpose is to significantly reduce the number of required dictionaries and to shorten the recognition time. There are some that provide recognition devices.

(d) Structure of the Invention The object of the present invention is to provide a speech recognition device that has a dictionary using a multi-template method and updates the contents of the dictionary through a learning process, from comparing each dictionary contents of the dictionary with input speech parameters. The shortest matching distance output means outputs the shortest matching distance, and the dictionary contents corresponding to the word corresponding to the next candidate are matched with the input speech parameters, and the recognition result is obtained.
and an output means for outputting matching distance information between the dictionary content and the dictionary content, an input means for inputting a next candidate request input, and a matching distance whose matching distance is close to the dictionary content of the shortest matching distance in response to the input means. control means for causing the output means to process different words, and a difference between a matching distance output from the output means and the shortest matching distance when the output means processes the different words; detection means for outputting a signal when the value is smaller than a predetermined reference value; and registration means for continuing the operation of each of the means until the generation of the signal and additionally registering the input audio parameter in the dictionary at the generation of the signal. This can be achieved by providing

(f) Embodiments of the invention Examples of the invention will be described below with reference to the accompanying drawings.

The accompanying drawings illustrate one embodiment of the invention. Reference numeral 1 denotes a microphone that receives the voice generated by the user U, and 2 denotes a voice feature extractor connected to the microphone 1. Reference numeral 3 denotes an input voice parameter buffer that stores the input voice parameters output from the voice feature extraction section 2. The output of the input voice parameter buffer 3 is connected to the registration learning control section 4 and also to the parameter matching section 5, and the input voice parameters thereto are connected to the registration parameter dictionary section under the control of the registration learning control section 4. The parameter collation section 5 compares the registered voice parameters output from the registered voice parameter 6 with the parameter collation section 5, and provides the recognition control section 7 with a collation distance for each registered voice parameter. Further, the registered voice parameter output of the registered voice parameter dictionary section 6 is connected to the registration learning control section 4, and the registration learning control section 4 inputs the learning voice parameters or input voice parameters as a result of the learning process as described later to the registered voice parameter dictionary section. 6. The device is configured to be re-registered or additionally registered.

The recognition control unit 7 receives the matching distance from the parameter matching unit 5, sets the dictionary with the closest matching distance as the dictionary of the recognition result, sends out the dictionary information as recognition dictionary information on the line 8, and also sends the dictionary information on the line 8 as the recognition dictionary information. Information indicating what language is in the dictionary is obtained from the dictionary directory section 9 under the control of the registration learning control section 4, and the recognition result is sent out on the line 10, and the matching distance at that time is sent out on the line as matching distance information. 11.

12 receives the matching distance information on the line 11, and when the relationship between the sequential matching distances is a predetermined relationship, for example, the difference between the matching distances is smaller than a reference value, a dictionary is displayed on the line 13. This is a dictionary addition control unit that generates an addition request signal.

Lines 8 and 13 are connected to the registration learning control section 4.

Reference numeral 14 denotes a buffer for storing the above-mentioned learning voice parameters or input voice parameters from the registration learning control section 4.

Further, 15 is a registration learning control section 4 and a recognition control section 7.
This is an operation panel that is connected to the device, sets the registration mode, learning mode, or recognition mode in the device, and also generates a request signal for the next candidate in the event of erroneous recognition as described below.

Next, the operation of the apparatus of the present invention having the above-described configuration will be explained.

A device implementing the present invention has a registration mode, a learning mode, or a recognition mode, and the device first has a registration mode, a learning mode, or a recognition mode.
The registration mode is set by the user U using the operation panel 15.

In this registration mode, the user U sequentially speaks words to be registered into the microphone 1.
Speech input from the microphone 1 is converted into speech parameters for recognition by a speech feature extraction section 2, and the input speech parameters are stored in an input speech parameter buffer 3. The stored input voice parameters are registered in the registered voice parameter dictionary section 6 under the control of the registration learning control section 4, and the identification information of the registered word is registered in the dictionary directory section 9 corresponding to the registered dictionary.

Such processing is performed for all words to be registered.

After that, the user U sets the learning mode from the operation panel 15 and inputs voice again in the same manner as in the registration mode. The input voice is given to the voice feature extraction section 2, where voice parameter extraction processing is performed. The extracted input voice parameters are stored in the input voice parameter buffer 3.

The input voice parameters of the input voice parameter buffer 3 are sequentially compared with all the registered voice parameters registered in the registered voice parameter dictionary section 6 under the control of the registration learning control section 4 in the parameter matching section 5. Each time the matching is completed, the matching distance is sent to the recognition control section 7.

When these comparisons are completed, the recognition control unit 7 sets the dictionary with the closest matching distance as the dictionary of recognition results, obtains from the dictionary directory unit 9 what words the dictionary contains, and displays the recognition results on the line 10. Send upward. At the same time, the recognition control unit 7 generates recognition dictionary information on line 8 and collation distance information on line 11.

The dictionary addition control unit 12 that receives the matching distance information on the line 11 compares the matching distances indicated by the matching distance information, that is, compares the matching distance in the first recognition result and the matching distance in the second recognition result. conduct,
If the difference is smaller than a predetermined reference value, it is assumed that there is a high possibility of erroneous recognition, and a dictionary addition request signal is generated.

As described above, the user U checks the recognition result output via the line 10, and if the recognition result is a misrecognition, the user U selects the next step for recognition from the operation panel 15. Request candidates. When the recognition control unit 7 receives the next candidate request, the recognition control unit 7 outputs another word with the next closest matching distance onto the line 10 as a new recognition result. Also, the matching distance information at that time is shown on line 11, and the recognition dictionary information is shown on line 8.
Send upward.

The dictionary addition control unit 12 compares the matching distance of the first candidate and the matching distance of the new recognition result, and sends out a dictionary addition request signal onto the line 13 if the difference is smaller than a reference value. In this case, if there is a signal on the line 13, the input voice parameters of the input voice parameter buffer 3 are stored as they are in the learning voice parameter buffer 14 via the registration learning control section 4, but in the opposite case, The registration learning control unit 4, which also responds to the recognition dictionary information on the line 8, synthesizes the input voice parameters and the registered voice parameters of the recognition dictionary at a specific ratio to create learning voice parameters, as in the past.
This is stored in the learning speech parameter buffer 14.

If the recognition result is correct, either by repeating or not repeating this process, the user U starts learning the next word by speaking the next word into the microphone 1.
input from In response to the input of the next word, the registration learning control section 4 starts a process of registering learning speech parameters into the registration speech parameter dictionary section 6. If no dictionary addition request has been issued, the learning voice parameters in the learning voice parameter buffer 14 are re-registered in the recognition dictionary of the registered voice parameter dictionary section 6. Also, if there is a request to add a dictionary,
The input speech parameters stored in the buffer 14 are registered in the empty dictionary of the registered speech parameter dictionary section 6, and the identification information of the word is registered in the dictionary directory section 9 corresponding to the empty dictionary.

In this way, multi-template conversion is performed only for templates with close matching distances.

Note that the degree of closeness between matching distances is preferably set to be optimal for each word or word group or in accordance therewith.

(g) Effects of the Invention As described above, according to the present invention, the same learning process as the conventional one is performed until a predetermined relationship occurs between the matching distances of the recognition results until a correct answer is obtained. In response to this occurrence, the input speech parameters are additionally registered in the dictionary, so only words with similar words that are easy to recognize or words whose matching distance is close to other words due to variations in pronunciation are converted into multi-templates. will be done. Therefore, it is possible to significantly reduce the dictionary capacity, thereby achieving effects such as shortening the collation time and, by extension, shortening the recognition time.

[Brief explanation of drawings]

The accompanying drawings illustrate one embodiment of the invention. In the figure, 1 is a microphone, 2 is a speech feature extraction section, 3 is an input speech parameter buffer, 4 is a registration learning control section, 5 is a parameter matching section, 6 is a registered speech parameter dictionary section, 7 is a recognition control section, 9 12 is a dictionary directory section, 12 is a dictionary addition control section, 14 is a learning speech parameter buffer, and 15 is an operation panel.

Claims (1)

[Scope of Claims] 1. A speech recognition device that has a dictionary using a multi-template method and updates the contents of the dictionary through a learning process, wherein the shortest matching distance is determined by comparing each of the dictionary contents of the dictionary with an input speech parameter. A shortest matching distance output means for outputting, and an outputting means for matching the dictionary content corresponding to the next candidate word with the input speech parameter and outputting the recognition result and matching distance information between the dictionary content and the dictionary content. an input means for inputting a next candidate request input; and in response to the input means, the output means causes the output means to perform processing for a word whose matching distance is close to the dictionary content of the shortest matching distance, but which is different from the dictionary content. a matching distance output from the output means when the output means processes different words;
a detecting means for outputting a signal when the difference from the shortest matching distance is smaller than a predetermined reference value; and a detecting means for outputting a signal, the operation of each of the means being continued until the generation of the signal, and upon the generation of the signal, the input audio parameter is transferred to the dictionary. A voice recognition device comprising: a registration means for additionally registering the voice recognition device.