JPH0634189B2 - Voice recognizer - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice recognition device, and more specifically, determines the suitability of a speaker, the suitability of a vocalization method, the suitability of a recognition word set, and the like in voice recognition. A voice recognition device capable of

(Prior Art) Conventionally, the recognition rate has been used to make the above determination. A conventional voice recognition device will be described with reference to FIG. FIG. 5 is a block diagram of a conventional voice recognition device, in which 1 is a voice input terminal, 2 is a recognition unit, 201 is a frequency analysis unit, 202 is a pattern matching unit, 203 is a standard pattern unit, and 204 is a judgment unit. 3 is a correct number counter, 4 is a pattern number counter, 5 is a divider, 6 is a multiplier, and 7 is an output terminal. To explain the operation, first, the input terminal 1
The frequency analysis unit 201 analyzes the voice input from the above to obtain a two-dimensional voice input pattern having a frequency axis and a time axis. Next, the pattern matching unit 202 performs pattern matching with the standard pattern for each recognition category stored in the standard pattern unit 203 to obtain the distance. Next, the determination unit 204 obtains the category with the smallest distance and sets it as the recognition result. If the recognition result is equal to the category of the initially input voice, it is a correct answer, and if not, it is a false recognition.
If the answer is correct, the correct answer number counter 3 is incremented. Also, the pattern number counter 4 is incremented each time recognition is performed. In this way, hundreds to thousands of voices are recognized, and the number of patterns and the number of correct answers are obtained. Next, the corrector number is divided by the number of patterns by the divider 5, and the multiplier 6 multiplies by 100 to obtain the recognition rate, which is output from the output terminal 7.

(Problems to be solved by the invention) However, the recognition rate is, so to speak, only the number of erroneously recognized patterns, and if 3-digit accuracy is required (for example, 98.8%), at least 1000 times. Has to be recognized, and it has a drawback that it is a very time-consuming and troublesome work.

The present invention has been made in order to solve the above-mentioned drawbacks of the prior art, and includes the aptitude of a speaker, the adequacy of a vocalization method,
An object of the present invention is to provide a voice recognition device that can obtain an index for determining the suitability of a recognition word set with a small number of recognition times.

(Means for Solving Problems) In order to solve the problems of the related art, a standard pattern storage unit that stores a standard pattern for each voice category and a pattern that compares an input voice pattern and a standard pattern In a voice recognition device including a pattern matching unit that obtains a distance representing the similarity between both patterns by performing matching, and a recognition determination unit that performs recognition determination based on the size of the distance obtained by the pattern matching unit, A first distance holding circuit that holds a distance D1 based on the standard pattern of the same category for each pattern, and a second distance holding circuit that holds a minimum distance D2 among the distances based on the standard pattern of a category different from the input voice for each pattern. Circuit, pattern number counting circuit for counting the number of recognized patterns, and pattern number counting It is characterized in that it has a geometric mean calculation circuit for obtaining a geometric mean of the values of D2 / D1 of each pattern based on the number of patterns obtained by the circuit.

(Operation) According to the present invention, since the voice recognition device is configured as described above, each means operates as follows.

The first distance holding circuit holds the distance D1 according to the standard pattern of the same category as the input voice for each pattern, while the second distance holding circuit holds the minimum distance D2 among the standard patterns of the category different from the input voice. Is held for each pattern. The pattern number counting circuit counts the number of recognized patterns. The geometric mean calculation circuit inputs the information from the first and second distance holding circuits and outputs D for each pattern.
2 / D1 is obtained, the geometric mean of the number of D2 / D1 values corresponding to the number counted by the pattern number counting circuit is obtained, and this is defined as recognition stability. Then, the voice recognition device of the present invention determines the suitability of the speaker in the voice recognition, the suitability of the utterance method, the suitability of the recognition word set, etc. based on the recognition stability.

(Embodiment) FIG. 1 is a circuit diagram showing a first embodiment of the present invention, in which a voice input terminal 1 is connected to a recognition unit 2, and the recognition unit 2 is connected to a data register D1 11, a data register D2 12, It is connected to the pattern number counter 17. Data register D1
11 and the data register D2 12 are connected to the divider 13,
The output of the divider 13 is connected to the multiplier 14, and the output of the multiplier 14 is connected to the register 15. The output of the register 15 is connected to the other input of the Nth root calculator 16 and the multiplier 14. N
The other input of the root calculator 16 is the pattern number counter 17
Connected with the output of. The output of the Nth root calculator 16 is the recognition stability and is output from the output terminal 18. The operation of the first embodiment will be described. First, 1 is set in the register 15 and the pattern number counter 17 is cleared to zero. Next, a voice with a clear category is input from the input terminal 1 and recognized by the recognition unit 2. From the recognition unit 2, the distance value and the stegory name for each category are obtained as a result in the ascending order of the distance value. Then, the distance (D1) of the category of the input voice is set in the data register D1 11, and the data register D2 is set.
12 is the minimum distance (D
2) Set. Also, the pattern number counter 17 is incremented each time recognition is performed. Next, the contents of the data register D2 12 are transferred to the data register D1 by the divider 13.
The value in register 15 is multiplied by the value in divider 13, and the result is set in register 15.
This operation is repeated N times, the Nth root is calculated from the value of the register 15 by the Nth root calculator 16, and is output from the output terminal 18.
Here, N is in the range of several tens to several hundreds.

The above operation is expressed by the equation (1).

Here, A is the recognition stability.

Next, a second embodiment of the present invention will be described. Number (1)
The formula is transformed as follows.

The second embodiment is based on the equation (4), and a block diagram is shown in FIG. In FIG. 2, the voice input terminal 1 is connected to the recognition unit 2, and the recognition unit 2 is connected to the data register D1 11, the data register D2 12, and the pattern number counter 17. Data register D1 11 and data register D2
12 is connected to the divider 13 and the output of the divider 13 is a logarithmic calculator
Connected to 21. The output of the logarithmic calculator 21 is connected to the adder 22, the output of the adder 22 is connected to the register 23, and the output of the register 23 is connected to the divider 24 and the other input of the adder 22. The other input of the divider 24 is connected to the output of the pattern number counter 17, the output of the divider 24 is connected to the exponential function calculator 25, and the recognition stability A is output from the output of the exponential function calculator 25. It is obtained and output from the output terminal 18.

To operate the circuit of the second embodiment, first, register 23
And the pattern number counter 17 is cleared to zero. Next, voice is input, the pattern number counter 17 is counted up, and the operation until the divider 13 obtains the value of D2 / D1 is the same as in the first embodiment.

Next, the logarithm calculator 21 calculates the logarithm of D2 / D1,
The value of the register 23 is added to the value by the adder 22, and the result is set in the register 23.

The above operation is repeated N times, and the value of the register 23 is divided by the divider.
It is divided by N by 24, and the exponential function of the result is obtained by the exponential function calculator 25 and output from the output terminal 18.

Next, a third embodiment of the present invention will be described.

Equation (4) is transformed as follows.

The third embodiment is based on the equation (5), and a block diagram is shown in FIG. In FIG. 3, the voice input terminal 1 is connected to the recognition unit 2, and the recognition unit 2 is connected to the data register D1 11, the data register D2 12, and the pattern number counter 17. The output of the data register D1 11 is connected to the logarithmic calculator 31, and the output of the data register 12 is connected to the logarithmic calculator 32. The output of the logarithmic calculator 32 is connected to the + input of the subtractor 33, and the output of the logarithmic calculator 31 is connected to the − input of the subtractor 33. The output of the subtractor 33 is connected to the adder 22,
The output of 22 is connected to the register 23, and the output of the register 23 is connected to the divider 24 and the other input of the adder 22. The other input of the divider 24 is connected to the output of the pattern number counter 17, the output of the divider 24 is connected to the exponential function calculator 25, and the recognition stability A is obtained from the output of the exponential function calculator 25. It is output from the output terminal 18.

To operate the third embodiment, first, the register 23 and the pattern number counter 17 are cleared to zero. Next, input the voice,
When the distance (D1) of the category of the input voice is set in the data register D1 11, the minimum distance (D2) of the categories other than the input voice is set in the data register D2 12, and the pattern number counter 17 is counted up. The steps up to this are the same as in the first embodiment.

Next, the logarithmic calculator 31 calculates the logarithm of D1, and the logarithmic calculator 32 calculates the logarithm of D2.
Subtract the logarithmic value of D1. Note that the output of the subtractor 33 is accumulated, divided by N, and the exponential function is calculated in the second place.
It is similar to the embodiment of.

FIG. 4 shows a histogram of D2 / D1 values when recognition is performed 600 times. In the figure, D2 /
A place where D1 <1 (indicated by) is D2 <D1 and therefore is erroneously recognized. Note that * indicates the correct answer.

The recognition rate used in the prior art is, so to speak, a problem of the number of numbers, and as is clear from FIG. It requires at least 1000 recognitions if needed, which is very time consuming. Therefore, according to each of the embodiments of the present invention described above, the geometric mean of D2 / D1 is defined as the recognition stability, and the easiness of recognition is determined from the entire histogram. It is sufficient to perform recognition about 100 times to obtain, and there is an advantage that determination does not take time.

(Effect of the Invention) As described in detail above, the present invention recognizes and stabilizes the geometric mean of the ratio (D2 / D1) of the distance (D1) of the correct category and the minimum distance (D2) of the other categories. It is defined as a degree, and based on this recognition stability, it is determined whether the speaker is suitable for speech recognition, the utterance method, the recognition word set, the necessity of re-registering a standard pattern, etc. The number of times can be determined in a short time.
Therefore, the present invention can be preferably used in a voice recognition system.

[Brief description of drawings]

1 is a block diagram of a first embodiment of the present invention, FIG. 2 is a block diagram of a second embodiment of the present invention, and FIG. 3 is a block diagram of a third embodiment of the present invention. The figure shows D2 / D1
FIG. 5 is a block diagram of a conventional voice recognition device. 1 ... Voice input terminal, 2 ... Recognition section, 11 ... Data register, 12 ... Data register, 13 ... Divider, 14 ... Multiplier, 15 ... Register, 16 ... N root calculator , 17 …… pattern number counter, 18 …… output terminal, 21 …… logarithmic calculator, 22 …… adder, 23 …… register, 24 …… divider, 25 …… exponential function calculator, 31, 32… … Logarithmic calculator, 33 …… Subtractor.

Claims (1)

[Claims] 1. A standard pattern storage unit for storing a standard pattern for each voice category, and a pattern matching for comparing an input voice pattern and the standard pattern to perform pattern matching to obtain a distance representing the similarity between both patterns. And a recognition determination unit that performs recognition determination based on the magnitude of the distance obtained by the pattern matching unit, a distance D1 based on a standard pattern in the same category as the input voice.
For each pattern, a second distance holding circuit for holding D2, which is the minimum distance among the distances according to the standard pattern of a category different from the input voice, for each pattern, and the number of recognized patterns. A voice recognition device, comprising: a pattern number counting circuit for counting; and a geometric mean arithmetic circuit for obtaining a geometric mean of D2 / D1 values of each pattern based on the number of patterns obtained by the pattern number counting circuit.