JPH06149285A - Speech recognizing device - Google Patents

Abstract

PURPOSE:To reduce speech misrecognition of the speech recognizing device which controls equipment by recognizing a speech. CONSTITUTION:The speech recognizing device which recognizes the speech by preprocessing the input signal of the speech and controls the equipment on the basis of the recognition result is provided with a parameter setting part 10 in which parameters for roughly classifying spectrum patterns of the speech by respective features and optimizing preprocessing by the roughly classified spectrum patterns of the speech are stored and a parameter switching part 11 which analyzes the frequency of the input signal of the speech, decides which of the roughly classified speech spectrum patterns the frequency-analyzed spectrum belongs to, and switches the set parameters of the parameter setting part 10.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition apparatus for controlling equipment by recognizing voice, and more particularly to reducing false recognition of voice recognition in the present invention.

[0002]

2. Description of the Related Art Conventionally, there have been techniques described below as techniques in such a field. FIG. 5 is a diagram showing a control system using a conventional voice recognition device. As shown in this figure,
A control system using a voice recognition device is used in a vehicle interior 30
A microphone 200 for capturing the voice of the speaker in 0;
A speaker direction / distance determining unit 201 for identifying a speaker based on the direction of the voice from the microphone 200 and the distance from the sound source.
And a speech recognition preprocessing unit 202 which is connected to the speaker direction / distance determining unit 201 and performs adaptive processing for eliminating noise from a voice signal that identifies a speaker and further performs automatic gain control (AGC).
And a voice recognition unit 203 connected to the pre-processing unit 202 for recognizing which of the registered words the voice matches, and various control units for forming control signals based on the words recognized by the voice recognition unit 203. 204, and a voice synthesis unit 205 for synthesizing words recognized through the various control units 204 into voice.
A speaker 206 connected to the voice synthesizing unit 205 for reproducing synthesized voice, an audio 207 controlled by the various control units 204, and an air conditioner 2
08, telephone 209, navigation 210, auto drive 211 and the like.

That is, the voice captured by the microphone 200 is processed by the speaker direction distance determination unit 201 and the preprocessing unit 20.
2 is recognized by the voice recognition unit 203, the result is transmitted to the speaker by the speaker 206 via the various control units 204 and the voice synthesis 205, and the various control units 204 control the audio 207 and the like. Here, the microphone 200 to the speaker 206 constitute a voice recognition device. In such a control system, a high voice recognition rate is required from the viewpoint of improving control reliability. Therefore, the performance of the voice recognition unit 203 is required to be improved, but the result of the signal processing that is the preceding stage is also greatly affected. Therefore, particularly for the microphone 200, the speaker direction distance determination unit 201 optimizes the difference in delay amount and the difference in gain amount between the microphones. Further, the preprocessing unit 202 optimizes the tap length, delay amount, and update coefficient of the noise reduction adaptive filter (ADF), optimizes the setting value of the automatic gain control device (AGC), and cuts off the frequency of the band limiting filter. And the cutoff characteristics (attenuation characteristics) are optimized.

[0004]

However, in the pre-processing unit 202 of the conventional speech recognition apparatus, although the various optimizations described above are performed empirically, it is difficult to optimize because of the strong influence of the speaker. There is a problem that the recognition rate varies widely among persons and a stable high recognition rate cannot be obtained.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a voice recognition device capable of performing signal preprocessing for increasing the recognition rate even for different speakers in view of the above problems.

[0006]

SUMMARY OF THE INVENTION In order to solve the above problems, the present invention sets a parameter in a voice recognition device which preprocesses an input signal of voice to recognize the voice and controls a device based on the recognition result. Section and parameter switching section.
The parameter setting unit roughly divides the spectrum pattern of the voice on the basis of each characteristic, and stores a parameter capable of optimizing the preprocessing for each broad spectrum pattern of the voice.

The parameter switching unit frequency-analyzes the input signal of the voice, determines which of the broad-spectrum voice spectrum patterns the frequency-analyzed spectrum analysis belongs to, and sets the setting parameter of the parameter setting unit. It is designed to be switched. Further, the spectrum pattern of the voice is roughly classified based on the first formant frequency which characterizes the spectrum pattern of the voice, and the parameter is switched according to the first formant frequency of the input signal.

[0008]

According to the speech recognition apparatus of the present invention, the speech spectrum patterns are roughly classified based on their respective characteristics, and the parameters for optimizing the preprocessing are stored for each of the roughly classified speech spectrum patterns, The input signal of the voice is frequency-analyzed, the frequency-analyzed spectrum analysis is determined to belong to the broadly divided spectrum spectrum of the voice, and the parameter is switched, so that the parameter is conventionally fixed. Since the speaker is variable depending on the speaker, there is no variation in the recognition rate among the speaking speakers, and a high recognition rate can be stably obtained.

Further, the first formant frequency that characterizes the spectrum pattern of the voice is used as a reference to roughly divide the spectrum pattern of the voice, and the parameter can be switched according to the first formant frequency of the input signal. it can. Further, the first formant frequency is used as a reference, the spectrum pattern of the voice is roughly classified into male and female, and the parameters of the male and female are switched by the first formant frequency of the input signal, so that the first formant frequency of the male and female is significantly different. Can be realized more easily by using.

[0010]

Embodiments of the present invention will be described below with reference to the drawings. FIG. 1 is a diagram showing a voice recognition device according to an embodiment of the present invention. In addition, the same reference number or symbol is used to represent the same component throughout the drawings. The voice recognition device shown in the figure includes a plurality of microphones 200 for capturing a voice of a speaker, a speaker direction distance determination unit 201 connected to the microphones 200 for identifying a speaker based on a direction of the voice and a distance from a sound source. , A speech recognition preprocessing unit 202 which is connected to the speaker direction distance determination unit 201 to eliminate noise from a voice signal which identifies a speaker and further performs automatic gain control (AGC), and the preprocessing unit 202 Connected to the voice recognition unit 20 for recognizing which word matches the registered voice and outputting the result to the various control units 204 (see FIG. 5).
3, a parameter setting unit 10 for switching and setting various processing parameters of the preprocessing unit 202, and a parameter switching determination unit 1 for determining switching of parameters of the parameter setting unit based on a voice signal from the microphone 200.
1 is provided.

Next, the parameter setting unit 10 and the parameter switching determination unit 11 will be described. FIG. 2 is a flowchart for explaining signal processing in the parameter setting unit 10 and the parameter switching determination unit of FIG. As shown in the figure, steps 1 and 2 are processes outside the recognition system, and determine the following parameter values in advance, and after step 3, the processes inside the recognition system are performed.

First, in step 1, the parameter switching determination unit 11 is based on the voice from the microphone 200.
According to the difference in the spectrum pattern (first formant frequency) of the voice waveform, the voice is roughly divided into n patterns and set, and the n spectrum patterns are stored. Since this storage technology itself is well known,
The description is omitted. Here, as a method of roughly classifying speech into n patterns, a method of patterning speech by the first formant frequency will be described below. First, the voice generation will be briefly described. It is said that the physical factors that determine the acoustic characteristics of voice are the characteristics of the sound source, the resonance characteristics of the vocal tract, and the emission characteristics of sound waves from the lips or nostrils. FIG. 3 is a diagram showing a spectrum of a sound wave. As shown in this figure, in the spectrum of a voice wave, the intensity of the voice decreases with a certain slope as the frequency increases, and there are several peaks corresponding to the resonance of the vocal tract, which is called a formant. The mountain with the lowest frequency is called the first formant. The first formant frequency generated by the first formant varies from person to person and varies from person to person. The present inventor has noticed that the recognition rate varies depending on the variation of the first formant frequency. Therefore, the preprocessing can be optimized by changing the parameter to be set in the preprocessing unit 202 according to the first formant frequency. Therefore, the parameter switching determination unit 11 includes n first formant frequency regions as spectral patterns,
For example, the first formant frequency is 100Hz, 125Hz,
A certain width is stored around 150 Hz and 175 Hz.

In step 2, first in step 1
Simulation and emulation are repeated for n kinds of voice patterns roughly classified by formant frequencies, and the optimum of each control parameter is determined. This optimum value has no theoretical support and is determined by experimental empirical rules. In the simulation, the voice recognition device of the present control system is not used, but for example, the preprocessing unit 202 and the voice recognition device are configured in a personal computer, and in the ideal state, the optimal parameters are roughly classified by the first formants. is there. In emulation, DSP (Digital Si
gnal processor) and the actual voice recognition device of this control system checks whether the parameters determined by simulation can be used.

Here, the contents of the parameters are, as described above, the tap length, the delay amount, the update coefficient of the noise reduction adaptive filter (ADF) in the pre-processing unit 202, the set value of the automatic gain controller (AGC), and the band. The cutoff frequency of the limiting filter, the cutoff characteristic (attenuation characteristic), and the like. In step 3, the parameter setting unit 1 including a memory
The optimum parameters obtained as described above are stored in 0 for each first formant frequency.

In step 4, the microphone 2
The voice input to 00 is spectrum-analyzed by the parameter switching determination unit 11, and the voice pattern of the speaker is compared with the n types of voice patterns stored in the parameter switching determination unit 11. That is, the first formant frequency obtained by the spectrum analysis determines the first formant frequency stored in the parameter switching determination unit 11, and the region of the n first formant frequencies in which the first formant frequency is stored is determined. Compare if they belong.

In step 5, from the comparison in step 4, the parameter switching determination unit 11 selects the most similar to the speaker's voice pattern from the n types of voice patterns. In step 6, each control parameter of the voice pattern selected in step 5 is set to the parameter setting unit 10
, The signal is pre-processed by the pre-processing unit 202 using this parameter, and the voice recognition unit 203 performs voice recognition based on the pre-processed signal.

Therefore, according to the present embodiment, the optimum parameters for preprocessing have been fixed in the past, but since the optimum parameters are changed by the speaker, it is possible to stably obtain a high recognition rate without depending on the speaker. Will be possible. FIG. 4 is a flowchart illustrating another signal processing in the parameter setting unit 10 and the parameter switching determination unit in FIG. As shown in the figure, step 11 describes signal processing outside the recognition system, and step 12 and subsequent steps describe signal processing inside the recognition system. In step 11, the optimum value of each control parameter is determined for each gender by simulation and emulation. As described above, the patterns of the voice spectrum are roughly classified into the first formant frequency of 100 Hz to 175 Hz in the case of men,
In the case of women, the first formant frequency is 200 Hz to 3
This is because it is at 00 Hz. That is, the difference between men and women is particularly remarkable in the first formant frequency.
As described above, the optimum value of the control parameter is not theoretically supported and is determined from an empirical rule.

In step 12, the control parameters determined in step 11, that is, the optimum parameters of the two series of parameters for male and female, are set to the parameter setting unit 1.
Store in 0. In step 13, the parameter switching determination unit 11 analyzes the spectrum pattern of the input voice from the microphone 200, and the first pattern is analyzed by this analysis.
From the formant frequency, it is determined whether the voice pattern is male or female.

In step 14, one of the parameters stored in the memory of the parameter setting section 10 is selected according to the gender of the speaker. In step 15,
Using the parameters selected in step 4, the preprocessing unit 20
Set to 2 for voice recognition. According to this signal processing example, there is an effect that the configuration is simplified as compared with the above example.

[0020]

As described above, according to the present invention, the speech spectrum patterns are roughly classified based on their respective characteristics, and the parameters for pre-processing optimization are stored for each of the roughly classified speech spectrum patterns. Then, the input signal of the voice is frequency-analyzed, the frequency-analyzed spectrum analysis is roughly classified to determine which voice's spectrum pattern belongs, and the parameters are switched. Since it is variable depending on the person, it is possible to stably obtain a high recognition rate without variations in the recognition rate depending on the speaker. This can be easily realized by roughly classifying the speech spectrum pattern based on the first formant frequency that characterizes the speech spectrum pattern, and by switching the parameter according to the first formant frequency of the input signal.

[Brief description of drawings]

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

2 is a flowchart illustrating signal processing in a parameter setting unit 10 and a parameter switching determination unit 11 of FIG.

FIG. 3 is a diagram showing a spectrum of a voice wave.

FIG. 4 is a flowchart illustrating another signal processing in the parameter setting unit 10 and the parameter switching determination unit 11 in FIG.

FIG. 5 is a diagram showing a control system using a conventional voice recognition device.

[Explanation of symbols]

 10 ... Parameter setting unit 11 ... Parameter switching determination unit 200 ... Microphone 201 ... Speaker direction / distance determination unit 202 ... Preprocessing unit 203 ... Speech recognition unit

Claims (2)

[Claims] 1. A voice recognition device for pre-processing an input signal of voice, recognizing voice, and controlling a device based on the recognition result, wherein the spectrum pattern of the voice is roughly classified according to each feature, and A parameter setting unit (10) that stores a parameter capable of optimizing the pre-processing for each of the separated speech spectrum patterns; and frequency analysis of the input signal of the speech, and the spectrum analysis subjected to the frequency analysis is roughly classified into the following. A voice recognition device, comprising: a parameter switching unit (11) for determining which of the voices the spectrum pattern belongs to and switching the setting parameter of the parameter setting unit (10). 2. The spectrum pattern of the voice is roughly classified on the basis of a first formant frequency which characterizes the spectrum pattern of the voice, and the parameter is switched to a parameter having the most similar spectrum pattern based on the result of the rough classification. 1. The voice recognition device according to 1.