JPH0250197A - Dictionary pattern producing device - Google Patents

Abstract

PURPOSE:To easily get a dictionary for an undefined speaker by indicating a held classification to a part, where voice patterns are held, from a part which indicates the classification of voice and selecting indicated patterns, and transferring patterns, whose quantity is equal to or smaller than the number of all selected patterns, to another medium. CONSTITUTION:Several patterns for dictionary generation, namely, standard patterns are held in a pattern storage part 1, and a list of contents is seen on a CRT display device 3. Required object names are inputted from a keyboard 4 to take out them from the storage part 1, and they are outputted to, for example, a floppy disk 5 to generate a dictionary for an unspecific speaker. The dictionary outputted to the floppy disk 5 is loaded to a recognizing device and is used. Thus, the dictionary for the unspecific speaker is easily generated.

Description

DETAILED DESCRIPTION OF THE INVENTION The present invention relates to a speech recognition dictionary pattern creation device.

111jlRecently, research on voice recognition has been active, and it is beginning to spread into the market. This device has two types: a specific speaker system in which the user registers their voice in advance, and an unspecified speaker system in which anyone's voice can be recognized without registration. To explain specifically using word speech recognition as an example, a word speech recognition device for unspecified speakers collects speech data of the words to be used in many places, and stores the pre-analyzed results in the device. I'll let you.

Next, it is determined which of the analysis results the unknown input is most similar to, and this is used as a recognition result.

Although the speaker-independent recognition device has the advantage of being able to recognize anyone's voice as described above, it has the disadvantage that the user of the device cannot freely register words for the reasons mentioned above.

The present invention has been made in view of the above-mentioned circumstances.
In particular, the purpose of this invention is to provide a device that allows a user of the device or a store selling the device to easily create a speech recognition dictionary for a speaker-independent recognition device. be.

In order to achieve the above-mentioned object, the present invention has a part that holds the audio pattern, a part that specifies the type of audio, and a part that selects the designated audio and transfers it to another medium. death,
Instruct all or part of the types held in the voice pattern holding part from the voice type indicating part, select the specified pattern, and select a pattern that is less than or equal to the number of all the selected patterns. A part that holds the sound pattern, a part that specifies the type of sound, a part that extracts the specified type of sound and performs calculations, and the result of the calculation. It has a means to transfer the obtained pattern to another medium, holds one or more patterns that can be standard patterns for one type of voice, and transfers each pattern equivalent to the unknown input pattern at the time of recognition. One or more types of voices are retained, a voice of the specified type is selected from the voice holding unit, and a recognition operation is performed between the voice and the human-powered equivalent pattern, and the number of voices with a minimum recognition error or equivalent thereto is selected. This system is characterized by selecting a pattern that can be a standard pattern for such a combination and transferring it to another medium. Hereinafter, the present invention will be explained based on examples.

FIG. 1 is a configuration diagram for explaining one embodiment of the present invention, in which 1 is a pattern storage section, 2 is a calculation section, and 3 is a CRT.
, 4 is a keyboard, and 5 is a floppy disk. In this embodiment, several patterns (standard patterns) for dictionary creation are held in the pattern storage unit 1.
The contents can be viewed on the CRT display 3.

On the other hand, by inputting the necessary target names from the keyboard 4, they are retrieved from the storage unit 1 and output to another medium, in this case a floppy disk 5, to create a dictionary for unspecified speakers. . The dictionary output to the floppy disk 5 is loaded into the recognition device and used. In this way, the dictionary -Jl for unspecified speakers can be created using a simple method. However, when considering word recognition, it is desirable that the dictionary patterns differ depending on the m-word group used. for example"
If the word ``Ricoh'' is in the word group ``1 piece'', ``2 pieces'', ``3 pieces''...rn i koJ and rrikO
” are easy to mistake because they are similar, and in order to distinguish between the two, it is necessary to emphasize the difference between /n/ and /r/, and in “Japanese,” “Science,” “Social Studies,” etc. If you want to put "Ricoh" in , use rrikaJ and r r i ko
The distinction between J, ie /a/ and /o/ must be emphasized. Therefore, it is clear that the recognition rate will be higher if different patterns are used for the same word "Ricoh" in the first word group and the second word group.

Therefore, as a countermeasure to this, multiple patterns of different types are maintained for each category (for example, words, which will be explained below using examples of words), and the necessary word groups are provided, and each 1jL pattern is created using the determined word group. It is desirable to select the combination that causes the least number of errors from among a plurality of word patterns to create the best dictionary. For this purpose, the embodiment shown in FIG. 2 is used.

In the embodiment shown in FIG. 2, 6 is an evaluation pattern storage unit, and 1 to 5 have the same functions as 1 to 5 in the embodiment shown in FIG. A part that holds the pattern, a part that specifies the type of audio, and a part that extracts the specified type of audio and performs calculations.
It has a means to transfer the pattern obtained as a result of the calculation to another medium, holds one or more patterns that can be a standard pattern for one type of voice, and has a means to transfer the pattern obtained as a result of the calculation to another medium. One or more patterns are held for each type of voice, and a voice of the specified type is selected from the voice holding unit, and a recognition operation is performed between it and the human-equivalent pattern, and the recognition error is minimized or This feature is characterized in that a pattern that can be a standard pattern of combinations that result in a number of KQs is selected, and the pattern is transferred to another medium. In the embodiment shown in FIG. 2, there is a section 1 for storing patterns for dictionary creation and an evaluation pattern storage section 6 for storing patterns for evaluating the dictionary. Multiple patterns are stored for each Togo language. For example, fu
A method announced as a recognition method using ZZy theory (Information Theory and Its Application Symposium (1987, 112
19-210) When using JC: 3-3-1, in the method described in JC 3-3-1 above, first, recognition is performed using each registered standard pattern and an unknown input pattern. Find similarities using pattern matching and classify into categories that show the greatest similarity. Manipulate a set of registered words.

1=(i□, 12....in), the set of word patterns is X, X= (x i□, x i2.-x i
n). is a finite set, and X is an infinite set.

In addition, the set of membership functions representing wordsmith-likeness is M, M= (m i□, m izt −min)
It is defined as Unknown voice pattern y (y (F-X
) is input, the degree of similarity Sjy between each word j and the unknown pattern is calculated, and j=max (Sjy) (1) A word name j that is a JC code is generated as a recognition result. A membership function for each word is registered inside the device, and the degree of similarity is determined from the membership function for unknown input.

The membership function m(x) must represent the likeness of the word pattern X in the presence of frequency fluctuations and time fluctuations due to one speaker. In this example, a binarized result of frequency analysis of word sounds is used as the word pattern.

FIG. 3(a) shows the word pattern of the word "s t a r t", and FIG. 3(b) shows the membership function of the same word.

In an actual system, the similarity Sjy between the membership function mj, which is a standard pattern, and the unknown input pattern y is expressed as follows. However, . Here, . is a normal product, and book is a logical product of α level cuts of y and mj, which is expressed by the following formula.

In many cases, α takes values from O to 3.

The dictionary pattern is a combination of word patterns superimposed as shown in FIG. 3, and the evaluation pattern is binary data in which each element is 1 or 0. Of course, for evaluation purposes, it is possible to store the sampled data itself, rather than binary data, and convert it into binarization during the next calculation, but it is better to binarize it in advance because the amount of data is smaller and unnecessary calculations are required. It turns out that you don't have to do it. It is preferable that the dictionary pattern has a plurality of different patterns for one word. Also, evaluation patterns should have multiple patterns for one word, and if possible, have multiple patterns so that the length and formant frequency distribution of these patterns are close to the distribution of people who speak the target language. This is desirable. This can be explained by taking as an example a case where a dictionary is needed in the Kanto region.When many words are uttered by many people, the average length of each word and the distribution of the number of people are taken as shown in Figure 4 (a). It is thought that the distribution will be similar. Similarly, if a graph is plotted based on the formant frequency of a certain vowel and the number of people, the result will be as shown in FIG. 4(b). Therefore, the evaluation patterns in the evaluation pattern storage section 6 in FIG. 2 are assembled into a set having the same distribution as in FIGS. 4(a) and 4(b). At this time, there is no particular restriction on the number of people for one word. There is no particular method for selecting dictionary patterns in the dictionary pattern storage section 2, and it is better to register a variety of patterns.

FIG. 5 is a diagram showing an example of the contents of the dictionary pattern storage section 2. The dictionary pattern storage section 2 is divided into a directory section A and a pattern memory B, and the word names and patterns of each word registered in the directory A. The address where each number is stored is held. In the pattern data memory, pattern data is arranged according to the values indicated in the directory.

On the other hand, FIG. 6 shows the evaluation data, where C is the directory section and D is the evaluation data, and the evaluation data is also stored in the same way as the dictionary pattern. In the example scheme, the number of bits representing one element of the dictionary is greater than that for evaluation.

In other words, the evaluation pattern is binarized.

Dictionaries are created by overlapping and adding many patterns equivalent to those for evaluation, resulting in a large amount of data.

Next, a usage example will be explained. When creating a standard pattern of 10 numbers, first input numbers 0 to 9 from the keyboard 4. Accordingly, patterns O, 1, . . . 9 are extracted from the dictionary pattern storage unit 1. At this time, if n patterns are registered for each word, 10×n patterns are loaded into the RAM of the calculation unit 3.

This means that n types of 10-digit dictionaries have been created. Next, patterns 0 to 1 for evaluation are taken out from the evaluation pattern storage section 6. This means that if there are m patterns, each pattern is loaded into the RAM of the calculation unit 3 so that the distributions shown in FIGS. 4(a) and 4(b) are obtained, and the n
Recognition calculations are performed using a dictionary of species and m types of evaluation patterns. This calculation may be performed in accordance with a flowchart as shown in FIG. First, recognition is performed using the first dictionary, and the recognition rate is set as Rs. The recognition at this time may be the sum of the recognition rates of all the evaluators.The zero-order word Ij is set to 1. j indicates the number of patterns for each word in the dictionary, and this box is also set to two. The pattern of word j is recognized by replacing it with the dictionary pattern of i, but at this time, all the evaluators recognize it, and the sum of all the evaluators is taken as the recognition rate Rj. Rj is compared with Rs, and if Rj is larger, word j is replaced with the pattern of i to create a new dictionary. If Rj > Rs does not hold,
Without replacing the pattern, try replacing it with pattern j from the next dictionary and recognize it again. In this way, by repeating i from 1 to n and j from 1 to the number of ilt words, a set of the best dictionary patterns of 10 words is selected from a large amount of dictionary data. A set of patterns selected in this way (one word may have two or more patterns if necessary) is output to the floppy disk 5, and the file a (floppy disk 5) is completed. This floppy is inserted into the floppy driver of the speech recognition device mentioned above, and the created dictionary is loaded into the device.
o Number speaker-independent recognition, 7a device. The storage area for such dictionary patterns and evaluation patterns includes m-words that are frequently used, such as 10 numbers, start, stop, end, hajime, hai, no, white, black, red, etc. ,
You can create a dictionary for Ren and Wu by adding words such as tiger, rabbit, etc. and selecting the necessary words.

In addition, this method selects words based on dictionary patterns stored in memory, but if you need something that is not stored in the dictionary, you can ask the manufacturer to rewrite the ROM. I have to take the troublesome method I mentioned. Therefore, this data is read from an external storage medium, such as a floppy disk, and loaded into the RAM.

Such an example is shown in FIG.

In the embodiment shown in FIG. 8, a floppy disk 7 is provided in place of the dictionary pattern storage section 1 and the evaluation pattern storage section 6 shown in FIG. The dictionary is then designated and retrieved from the keyboard 4, and after the above-mentioned calculations are performed, the completed dictionary is output to the floppy disk 5. FIG. 9 is a diagram showing such a procedure, and the method of selecting patterns based on similarity may be the same as that described above, but is not limited to this method. For example, in FIG. 8, the dictionary pattern of the word input from the keyboard 11 is retrieved from the floppy disk 7. At this time, if all n patterns are extracted for each word, n sets of dictionaries will be created. This may be output to the floppy disk 5. At this time, the recognition device may sequentially use n sets of dictionaries and use the one that is most convenient for use. In the case of the apparatus shown in FIG. 8, by replacing the floppy disk 7, it is possible to handle an infinite number of words.

Although this method is aimed at creating a dictionary for non-specific speakers, it is convenient because if the non-specific speaker device can be used as a specific speaker device, the user can register any a-voice. Therefore, the system shown in FIGS. 10 and 11 is such that the user's voice is added to the dictionary for unspecified speakers. As the recognition method, a method using the binary TSP described above is shown, but the present invention is not limited to this.

First, the user speaks the required words into the microphone 8 the required number of times. After each voice is amplified by a microphone amplifier 9, it passes through a filter bank 10 and a binarization unit 11 to
The values are converted into values and superimposed in the calculation unit 2 to create a standard pattern. This is stored in the dictionary pattern storage section 12 together with other patterns. By doing this, it is possible to create dictionary patterns not only for unspecified speakers but also for specific speakers. 1. A normal speech recognition device for unspecified speakers can be used as a recognition device for specific speakers. This makes it possible to register any word with a higher correct answer rate than the speaker-independent method.

Note that in the above explanations, dictionary patterns are stored in memory in advance, but this is not necessarily necessary, and the data may be input into a calculation unit and subjected to statistical processing.

In addition, such a device is not kept for personal use, but is provided for business use, and the user inserts a recording medium such as a floppy into a designated position, enters words from a designated keyboard, and inserts money. It is also possible to create a so-called vending machine that transfers necessary dictionary patterns to one's own floppy disk.

As is clear from the above description, according to the invention, a dictionary with a high recognition rate for any speaker can be easily obtained.

[Brief explanation of the drawing]

Figures 1 and 2 are configuration diagrams for explaining the present invention in detail, Figure 3 is a diagram showing a configuration example of a membership function, and Figure 4 is a diagram showing the average length of each word and the number of people. A diagram showing the distribution, FIG. 5 is a diagram showing an example of registration of JR quasi-patterns, FIG. 6 is a diagram showing an example of registration of evaluation patterns, and FIG.
FIG. 8 is a diagram showing a flowchart of recognition calculation; FIG. 8 is a diagram for explaining another embodiment of the present invention; FIG. 9 is a diagram showing the operating procedure of the embodiment shown in FIG. 8; FIG. and 11th
The figure is a diagram showing an example of a recognition method for unspecified speakers. 1... Pattern storage section, 2... Calculation section, 3... C
RT, 4...keyboard, 5...floppy disk, 6...evaluation pattern storage section, 7...floppy disk. 8... Microphone, 9... Microphone amplifier, 10... Filter bank, 11... Binarization section, 12... Dictionary pattern storage section. Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures Figures

Claims (1)

[Scope of Claims] 1. It has a part that holds a sound pattern, a part that instructs the type of sound, and a part that selects the designated sound and transfers it to another medium, and Instruct all or part of the types held in the part that holds the audio pattern from the instruction part, select the specified pattern, and transfer patterns that are less than or equal to the number of all selected patterns to other media. A dictionary pattern creation device characterized in that it transfers data. 2. A part that holds the sound pattern, a part that specifies the type of sound, a part that extracts the designated type of sound and performs calculations, and a means for transmitting the pattern obtained as a result of the calculation to another medium. It has one or more patterns that can be standard patterns for one voice type, holds one or more patterns for each voice type that are equivalent to the unknown input pattern during recognition, and Selects the specified type of voice from the holding unit, performs a recognition operation with the input equivalent pattern, and selects a pattern that can be a standard pattern of combinations that will result in a minimum number of recognition errors or an equivalent number. A dictionary pattern creation device characterized in that the dictionary pattern creation device is configured to create a dictionary pattern and transfer it to another medium.