JPS6073697A - Preparation of phoneme dictionary - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

DETAILED DESCRIPTION OF THE INVENTION (1) Technical Field of the Invention The present invention relates to a method for creating a phonetic dictionary used in a speech recognition device.

(2) Background of the Technology Generally, when inputting voice data using a voice recognition device, comparison data with which the voice data is compared is required.

A device that includes a large amount of such comparison data is generally known as a phoneme dictionary.

(3) Prior art and its problems Conventionally, this type of phonetic dictionary prepares a group of voice data for dictionary creation of, for example, 100 words, extracts the characteristics of each voice data, and It is created by sequentially writing subdivided phonetic symbols and their characteristic data.

However, in the case of such a conventional phoneme dictionary, since the phoneme dictionary is created independently of actual speech recognition, it is difficult to use the phoneme dictionary when the number of voice data for dictionary creation is relatively small. When voice recognition is performed using , there is a problem that errors tend to occur in the recognition results and the voice recognition tends to be inaccurate.

In order to solve this problem, it is possible to improve the reliability of speech recognition by increasing the number of speech data for dictionary creation to some extent, but in this type, the capacity of the phonetic dictionary itself increases. In addition, there is a problem in that the processing for voice recognition takes time, which does not meet the demand for speeding up voice recognition processing.

(4) Purpose of the Invention The present invention has been made based on the above points of view, and its purpose is to accurately recognize speech without increasing the number of speech data for dictionary creation. The purpose of the present invention is to provide a method for creating a phonetic dictionary that can be used to create a phonetic dictionary.

(5) Structure of the Invention The basic structure of the present invention is that when creating a phonetic dictionary, a primary phonetic dictionary is created in advance based on a group of speech data for dictionary creation, and then this primary phonetic dictionary is created. Each created speech data is recognized using the primary phonological dictionary excluding the dictionary based on the speech data to be recognized, and recognition error rules are extracted from this recognition result and applied to the primary phonological dictionary. The purpose is to add the above error rules to create a second-order phoneme dictionary.

(6) Embodiments of the Invention Hereinafter, the present invention will be described in detail based on embodiments shown in the accompanying drawings.

FIG. 1 is a block diagram showing an embodiment of an apparatus for carrying out the method of creating a phonetic dictionary according to the present invention.

In the figure, 1 is a microphone for detecting audio data, and 2 is a microphone that converts the audio data detected by the microphone 1 into N widths and slices it into several msec widths, which is called frame data. The input unit 3 that converts into a series extracts the features of the audio data by calculating the audio power spectrum etc. based on the frame data of the input unit 2, and converts the time series of the frame data into a time series of feature data. 4 is a display; 5 is a predetermined number (for example, 100) as shown in FIG. 2;
A registration word phoneme string file 6 stores words to be registered in advance as a phoneme string (5APORO, etc.), for example, the third
As shown in the figure, this is a word speech feature file that stores the time series of feature data from the feature extraction section 3 together with the word phoneme string from the registered word phoneme string file 5, and these constitute the speech registration section A as a whole. are doing.

Further, reference numeral 7 is a phoneme division unit that divides the phonemes of the feature data time series according to the phoneme string of the word in the word voice feature file 6. For example, as shown in FIG. between the dots (e.g. 08AKA
> This process involves matching one character of a phoneme string. As shown in FIG. This is a phoneme labeling unit that outputs information indicating whether the feature data occupies the range of the feature data. and,
As shown in FIG. 5, 9 is a primary phoneme dictionary that stores feature data together with phoneme symbols and word phoneme strings, and the phoneme segmentation unit 1, phoneme labeling unit 8, and primary phoneme dictionary 9 store feature data together with phoneme symbols and word phoneme strings. It constitutes a secondary phoneme dictionary creation section B. The above-mentioned phoneme division information is stored in the word voice feature file 6 as shown in FIG.

Further, reference numeral C denotes a secondary phoneme dictionary creation unit, which includes a phoneme recognition unit 10, an error rule extraction unit 11, a phoneme label addition unit 12, and a secondary phoneme dictionary 13. The phoneme recognition unit 10 recognizes the feature data time series of word phoneme strings in the word phonetic feature file 6 using the primary phoneme dictionary 9, and the recognition method is a method used in a speech recognition device, for example. This is the same as the dissimilarity calculation method, similarity calculation method, etc. However, when performing phoneme recognition of a certain word, the primary phoneme dictionary created from the feature data of the word is not used. Further, as shown in FIG.
Compare the phoneme recognition results obtained from It and find out where the recognition errors are.

This method extracts all rules. Then, the phoneme label addition unit 12 adds the error rule obtained from the error rule extraction unit 11 to the primary phoneme dictionary 9, as shown in FIGS. is created.

Therefore, when creating a phoneme dictionary using the above device, the contents of the registered word phoneme string file 5 are displayed one after another on the display 4, and the displayed words are uttered into the microphone 1. Bye.

At this time, the uttered words are first stored in the word voice feature file 6 together with the registered words as voice data for dictionary creation, and the voice registration is completed at this stage. Thereafter, the contents of the word phonetic feature file 6 are stored in the primary phoneme dictionary 9 via the phoneme division section 1 and the phoneme labeling section 8, and the primary phoneme dictionary 9 is created. Furthermore, after this,
The phoneme recognition unit 10 performs phoneme recognition of speech data using the primary phoneme dictionary 9, but since the dictionary of the speech data to be recognized is not used, there are no information in the primary phoneme dictionary 9 during phoneme recognition. There is no perfect match to the speech data, and the phonological recognition of the speech data will be subject to some error tendency. For this reason, errors occur in the recognition of speech data that includes phonemes with a high tendency for errors, such as the one shown in FIG.
:), the rule that [:I:] is easily mistaken for [E] is extracted by the error rule extraction unit 11, and the phonological label addition unit 12
The information is stored in the secondary phonetic dictionary 13 together with the primary phonetic dictionary 9, and the secondary phonetic dictionary 13 is created at this stage.

Since the 1F second-order phoneme dictionary 13 created in this way has automatically learned phoneme error rules, the second-order phoneme dictionary is used as a phoneme dictionary for the speech recognition device. In this case, speech recognition can take into account the tendency for errors in phonemes, and speech recognition becomes more accurate accordingly.

Note that the specific device for carrying out the method for creating a phonetic dictionary according to the present invention is not limited to that shown in the above embodiments, and the design may be modified as appropriate.

(7) As described in detail, according to the method for creating a phonetic dictionary according to the present invention, it is possible to create a phonetic dictionary in which phonetic error rules have been learned. It becomes possible to know error trends, and speech recognition can be made more accurate accordingly. Further, according to the present invention, there is no need to increase the number of voice data for dictionary creation when creating a phoneme dictionary, so it is possible to effectively avoid the situation where the capacity of the phoneme dictionary increases unnecessarily, and There is no risk of impairing the demand for faster recognition processing.

[Brief explanation of drawings]

FIG. 1 is a block diagram showing an embodiment of a device for carrying out the method for creating a phoneme dictionary according to the present invention, FIG. 2 is an explanatory diagram showing the contents of a registered word phoneme string file, and FIG. 3 is a word phonetic diagram. An explanatory diagram showing the contents of the feature file, Fig. 4 is an explanatory diagram showing the operation of the phoneme segmentation section and the phoneme labeling section, Fig. 5 is an explanatory diagram showing the contents of the primary phoneme dictionary, and Fig. 6 is an error. FIG. 7 is an explanatory diagram showing the operation of the rule extracting section, and FIG. 7 is an explanatory diagram showing the contents of the secondary phoneme dictionary. 1...Microphone 2...Input unit 3...Feature extraction unit 4...Display 5...Registered word phoneme sequence file 6...Word audio feature file...Phonological segmentation unit 8...・Phonological labeling section 9...Primary phonological dictionary 10...Phonological recognition section 11...Error rule extraction section 12...Phonological label adding section 13...Second phonological dictionary

Claims (1)

[Claims] When creating a phonetic dictionary used in a speech recognition device,
After creating a primary phonetic dictionary based on a group of audio data for dictionary creation in advance, each audio data for which the primary phonetic dictionary was created is converted into a primary phonetic dictionary excluding the dictionary based on the audio data to be recognized. Creation of a phonological dictionary, characterized in that recognition is performed using a dictionary, and error rules in recognition are extracted from the recognition results, and the error rules are added to the primary phonological dictionary to form a secondary phonological dictionary. Method.