JPH06242795A - Speech recognition device - Google Patents

Abstract

PURPOSE:To suppress a decrease in recognition rate only by slightly increasing both a throughput and a memory capacity by outputting a recognition result according to the 1st distance between an analyzed input speech and a standard pattern and the 2nd distance between an environmental noise at the time of the speech input and the standard pattern. CONSTITUTION:A feature parameter extraction part 11 analyzes the input sound signal, at every frame, to calculate feature parameters required for distance calculation. Then, an inter-frame distance calculation part (1) 15 finds the inter- frame distance between the standard pattern of the speech to be recognized and a noise pattern. When a speech section decision part 12 decides a speech section, the process is advanced to an inter-frame distance calculation part (2) not through a noise pattern update part 13, and the distances to all frames of respective standard patterns stored in a standard pattern storage part 17 are calculated and stored in a buffer (2) 19. A recognition decision part 24 finds a recognition result by using the obtained integral distance value and outputs the result.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device capable of improving the recognition rate under environmental noise.

[0002]

2. Description of the Related Art Speech recognition in a situation where environmental noise is large is
The recognition accuracy is significantly lower than when the environmental noise is small. Therefore, conventionally, various methods have been studied as measures against environmental noise.

As a conventional method, (1) standard patterns are created at various noise levels, the ambient noise level is measured at the time of recognition, and only the standard pattern registered at the noise level closest to the level is recognized. Method used for
(2) A method of removing noise from the input voice by separately analyzing the input voice and the ambient noise under the environmental noise and applying an inverse filter composed of the latter spectrum to the former, (3) using a standard pattern A method of superposing environmental noise at the time of recognition and the like are known.

[0004]

However, the above method (1) has a problem in that many standard patterns must be prepared in advance, which requires a large amount of memory. Further, the method (2) has a problem that the amount of calculation for removing noise becomes large. Further, the method (3) has a problem that the standard pattern has to be always recalculated at the time of recognition, resulting in a large amount of calculation.

Therefore, the conventional device for preventing the reduction of the recognition rate in the presence of noise has a problem that the calculation amount and the memory amount are likely to increase.

In view of the above-mentioned problems in the conventional method, an object of the present invention is to provide a voice recognition device which can effectively improve the recognition rate under noise without increasing the calculation amount and the memory amount so much. To do.

[0007]

SUMMARY OF THE INVENTION An object of the present invention is to acoustically analyze input voice and environmental noise at the time of voice input by a predetermined method, and a first distance between the analyzed input voice and a standard pattern. And a speech recognition device having means for calculating a second distance between environmental noise at the time of voice input and a standard pattern, and means for outputting a recognition result based on the calculated first distance and second distance. To be done.

[0008]

The voice recognition device of the present invention acoustically analyzes the input voice and the environmental noise at the time of voice input by a predetermined method,
The first distance between the analyzed input voice and the standard pattern and the second distance between the environmental noise at the time of voice input and the standard pattern are calculated, and the recognition result is output based on the calculated first distance and the second distance. .

[0009]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS An embodiment of a voice recognition device of the present invention will be described in detail below with reference to the drawings.

FIG. 1 is a block diagram showing the main part of the configuration of an embodiment of the speech recognition apparatus of the present invention.

The speech recognition apparatus of FIG. 1 applies a distance correction method to a typical dynamic programming (DP) matching method, and a characteristic parameter extraction unit 1 which is means for acoustic analysis.
1, a voice section determination unit 12, a noise pattern update unit 13, a noise pattern recording unit 14, an inter-frame distance calculation unit (1) 15 that is a part of a distance calculation unit, and a part of a distance calculation unit. Inter-frame distance calculation unit (2) 16, standard pattern storage unit 17, buffer (1) 18, buffer (2) 19, subtraction calculation unit 20, cumulative distance calculation unit 21, which is a part of distance calculation means, addition The calculation unit 22, the buffer (3) 23, and the recognition determination unit 24 that outputs the recognition result are included.

Next, the operation of the above components will be described.

First, the feature parameter extraction unit 11 analyzes the input acoustic signal for each frame, and calculates the feature parameters required for distance calculation. Based on this characteristic parameter, the next voice section determination unit 12 determines whether the current frame is a section containing voice or a noise section (non-voice section) containing no voice. This can be determined, for example, by the power or spectrum shape of the input sound.

If it is determined that it is in the non-voice section, the process proceeds to the noise pattern updating section 13, and from the noise pattern stored in the noise pattern recording section 14 and the newly analyzed characteristic parameter of the non-voice section, For example, the noise pattern is recalculated by the moving average method, or the noise pattern is updated by substituting new feature parameters as they are.

Before operating the voice recognition device for the first time, it is necessary to store an appropriate initial value in the noise pattern recording section 14. Next, the inter-frame distance calculation unit (1) 1
In 5, the inter-frame distance between the standard pattern of the voice to be recognized and the noise pattern is obtained. The standard pattern is stored in the standard pattern storage unit 17 in advance.
As a method of calculating the distance, for example, there is a method of obtaining the Euclidean distance between the characteristic parameters.

This distance calculation is performed for all frames of the noise pattern and each standard pattern, and the calculation result is recorded in the buffer (1) 18. This is the end of the processing in the non-voice section.

On the other hand, when the voice section determining unit 12 determines that the voice section is present, the noise pattern updating unit 13 does not pass and the process proceeds to the interframe distance calculating unit (2) 16 and is stored in the standard pattern storing unit 17. The distance of each standard pattern from all frames is calculated. The calculation result is recorded in the buffer (2) 19.

In FIG. 1, it seems that the inter-frame distance calculation section goes to different voice sections or non-voice sections, but the inter-frame distance calculation section may be a common processing module or a different processing module.

Here, a recurrence formula example of DP matching will be described.

The restricted path shape for calculating the cumulative distance is shown in FIG.
In such a case, an initial condition like the formula (1) and a recurrence formula like the formula (2) can be considered.

[0021]

[Equation 1]

[0022]

[Equation 2]

This recurrence formula is executed in the cumulative distance calculation unit 21 of FIG.

In the normal DP matching method, the equation (2)
The value of the buffer (2) 19 is used as it is as the first term Dk (i, j) of the right side of Dk (i, j) in the present invention.
The feature is that j) is obtained using the values recorded in the two buffers, buffer (1) 18 and buffer (2) 19.

The distance between the j-th frame of the standard pattern k in the buffer (1) 18 and the noise pattern is Nk (j), and the latest input speech frame i in the buffer (2) 19 and the j-frame of the standard pattern k. Rk is the distance between the eyes
Assuming that (j), Dk (i, j) is obtained by the equation (3).

[0026]

[Equation 3]

Here, α is a subtraction coefficient, and since it is more flexible to set an appropriate numerical value in accordance with the environmental noise level and the distance calculation scale, for example, α = 0.3 is set. The calculated cumulative distance is stored for each standard pattern as shown in FIG.

Returning to the explanation of FIG. 1, after calculating the distance between the input speech frame and each frame of each standard pattern and substituting it in the buffer (2) 19, the buffer (2 ) 19 each recorded value, each recorded value of the corresponding buffer (1) 18 is multiplied by a coefficient α, and a subtraction calculation unit 20
To subtract. The left side D of equation (3) thus obtained
After being added in the addition calculation unit 22 using k (i, j) and the cumulative distance calculation value up to the present, the obtained cumulative distance calculation value is the buffer (3) 23 constituting the cumulative distance calculation unit 21.
Recorded in.

Using the cumulative distance value thus obtained, the recognition determination section 24 obtains a recognition result and outputs the result.

In the above-mentioned configuration, compared with the conventional DP matching voice recognition device, the voice section determination unit 12, the noise pattern update process, and the increase in the processing amount of the subtraction of the equation (3),
Although the noise pattern recording unit 14 needs to increase the memory amount of the buffer (1) 18 for each standard pattern, the noise pattern is only a feature parameter for one frame,
The required memory amount of the buffer (1) 18 is the buffer (2) 1
The number of storage areas is exactly the same as 9, and the number of storage locations is the same as the total number of frames of all standard patterns.

This is obviously less than the amount of memory required to increase the standard pattern. Because
This is because in order to increase the number of standard patterns by one, a memory amount obtained by multiplying the total number of frames of all standard patterns by the number of dimensions of the characteristic parameter is required.

Regarding the processing amount, it is not necessary to calculate the cumulative distance in the non-voice section, and it is not necessary to update the noise pattern in the voice section. The increase is not so large.

Compared with the conventional method, both the processing amount and the memory amount can be slightly increased, and effective noise countermeasures can be taken.

FIG. 3 shows an example of the hardware configuration required to implement the voice recognition apparatus of FIG.

The voice recognition device shown in FIG.
1. Analog / digital (A / D) conversion circuit 32 connected to microphone 31, microprocessor 33 connected to A / D conversion circuit 32, data memory 34 connected to microprocessor 33, connection to microprocessor 33 Configured program memory 35.

Next, the operation of the voice recognition device of FIG. 3 will be described.

The acoustic signal collected by the microphone 31 is converted into a digital value by the A / D conversion circuit 32.
Then, in the microprocessor 33, processing such as feature parameter extraction and noise determination is sequentially performed according to the block diagram of FIG. Standard pattern required for processing,
Information that needs to be saved, such as noise patterns and distance data, is stored in the data memory 34, and is read or written as needed. A program for performing these processes is built in the program memory 35 and read by the microprocessor 33 to operate.

The reason why the voice recognition rate is lowered under high noise is that the voice pattern is deformed by the addition of noise. Therefore, even if the same voice as the standard pattern is input to the voice recognition device, it will be different from the standard pattern due to the influence of environmental noise, and the accurate distance between the standard pattern and the input pattern cannot be calculated. Will end up. This becomes remarkable as the level of environmental noise increases.

What should be noted here is the fact that some standard patterns are originally close or far from the environmental noise. Therefore, when noise is superimposed on the voice to be recognized, the distance to the standard pattern may be very large, and on the contrary, the distance may not be increased so much. The distance between the standard pattern close to the environmental noise and the recognized voice does not become so large, and the distance between the standard pattern far from the environmental noise and the recognized voice rapidly increases. By correcting this phenomenon, it is possible to suppress a decrease in recognition rate due to superposition of environmental noise.

In normal speech recognition, the inter-pattern distance is obtained from the characteristic parameter of the standard speech pattern and the characteristic parameter of the input speech pattern. From the inter-pattern distance, between the characteristic parameter of the environmental noise at the time of recognition and the characteristic parameter of the standard pattern. The above correction can be performed by subtracting the distance. That is, from the distance between the standard pattern and the input pattern obtained by the normal speech recognition method, the larger the standard pattern is to the environmental noise at the time of recognition, the larger the distance is subtracted. Since a too large distance is not subtracted, it is possible to reduce the difference in advantage and disadvantage due to the similarity between the standard pattern and the environmental noise at the time of recognition.

[0041]

The speech recognition apparatus of the present invention comprises means for acoustically analyzing the input voice and the environmental noise at the time of voice input by a predetermined method, the first distance between the analyzed input voice and the standard pattern, and the voice. Since the means for calculating the second distance between the environmental noise at the time of input and the standard pattern and the means for outputting the recognition result based on the calculated first distance and second distance are provided, the speech under noise In recognition, it is possible to suppress a decrease in the recognition rate with a slight increase in the processing amount and the memory amount by correcting the advantage or disadvantage due to the difference in the similarity between the environmental noise at the time of recognition and each standard pattern.

[Brief description of drawings]

FIG. 1 is a block diagram showing the configuration of an embodiment of a voice recognition device of the present invention.

FIG. 2 is an explanatory diagram of a DP matching limit path shape and a cumulative distance buffer in the voice recognition device of FIG.

FIG. 3 is a block diagram showing a configuration example of hardware for realizing the voice recognition device in FIG.

[Explanation of symbols]

 11 Characteristic Parameter Extraction Section 12 Voice Section Determining Section 13 Noise Pattern Updating Section 14 Noise Pattern Recording Section 15, 16 Interframe Distance Calculation Section 17 Standard Pattern Storage Section 18, 19, 23 Buffer Memory 20 Subtraction Calculation Section 21 Cumulative Distance Calculation Section 22 Addition calculation unit 24 Recognition determination unit

Claims (1)

[Claims] 1. A means for acoustically analyzing an input voice and environmental noise at the time of voice input by a predetermined method, a first distance between the analyzed input voice and a standard pattern, and environmental noise at the time of voice input. A voice recognition device comprising: a means for calculating a second distance between standard patterns; and a means for outputting a recognition result based on the calculated first distance and second distance.