JPH02232697A - Voice recognition device - Google Patents

Abstract

PURPOSE:To improve the recognition rate by providing an acoustic analyzing means which extracts feature data from a voice signal corresponding to a voice section of the voice signal from a microphone and outputs it to a comparing means. CONSTITUTION:A noise cut filter 2 outputs the voice signal inputted from the micriphone 1 after attenuating its band where more noise components are present and a voice section detection means 6 detects a voice section with the signal from the noise cut filter 2. Then an acoustic analyzing means 8 extracts the feature data from the voice signal corresponding to the voice section of the voice signal from the microphone 1 and outputs the data to the comparing means 14, which compares standard data stored in a voice storage means 12 with the feature data. Consequently, the feature data can be extracted from the inputted voice signal not through the noise cut filter 2 and the voice can be recognized based upon the feature data, so the recognition rate is improved.

Description

[Detailed description of the invention] [Industrial field of application 1 The present invention relates to a voice recognition system for determining voice information,
Special 1. Concerning speech recognition equipment in environments with a lot of unsteady noise. [Prior Art] Conventionally, voice recognition devices have been used to control the operation of various devices by recognizing voice signals based on the similarity between voice signals uttered by an operator, which serve as keywords, and registered voice signals. It is being done. For example, if a voice recognition device is installed in an air conditioner, noise from the surrounding environment in which the voice recognition device is used, such as noise from the air conditioner, vibration sounds, and other external noises, may This may cause misrecognition of speech, and countermeasures are being taken to prevent this. For example, while an air conditioner is operating, the noise signal is attenuated by passing the audio signal through a noise cut filter that has the characteristic of cutting the low frequency band as shown in Figure 5. As a result, the low frequency noise generated by the air conditioner as shown in Figure 6 or the external noise can be attenuated as shown in Figure 7. Then, feature data was extracted from the audio signal passed through the noise cut filter, compared with pre-stored standard data, and the audio information was determined based on the matching to control air conditioners, etc. [Problems to be Solved] However, with these conventional speech recognition devices, the input speech signal is processed by a noise cut filter, so noise can be reduced. For example, when the vowel ``a'' is input as an audio signal, an audio signal with frequency components as shown in Figure 8 f-2 is output from the microphone. When this audio signal is processed by a noise cut filter with the characteristics described above, the formant frequency that characterizes the audio is reduced.
This causes the peaks of the first and second formants, etc. of low frequencies in Figure 8 to be lost. Table 1: When extracting features after filter processing 1: of audio signals There was a problem that the most important feature quantity during recognition was reduced, leading to a decrease in the recognition rate.Therefore, the present invention was developed as described above. The aim is to solve the problems of
The purpose of this invention is to provide a speech recognition device that extracts features of speech signals without being affected by filters that attenuate noise and improves recognition rate. [Means for Solving the Problems] In order to achieve the above object, the present invention has the following configuration as a means for solving the problems. That is, the present invention has a voice storage means that stores standard data corresponding to the voice in advance. In the speech recognition device, the speech recognition device includes comparison means for comparing the standard data and characteristic data corresponding to the speech signal inputted from the microphone, and outputs the speech signal after attenuating a band containing many noise components. a noise cut filter; a voice section detecting means for detecting a voice section using a signal from the noise cut filter; a sound signal from the microphone; extracting feature data from the voice signal corresponding to the voice section and comparing the characteristic data; This is the configuration of a speech recognition device characterized by comprising: an acoustic analysis means for outputting an output to a speech recognition means. [Function] A speech recognition device having the above configuration (a) The noise cut filter attenuates and outputs a band containing many noise components in the speech signal input from the microphone, and the speech section comparison means Then, the acoustic analysis means extracts feature data from the voice signal corresponding to the first voice interval of the voice signal from the microphone and outputs it to the comparison means. compares the standard data stored in the audio storage means with the feature data.Therefore, feature data can be extracted from the input audio signal without passing through a noise cut filter, and speech recognition can be performed based on this feature data. [Examples] Examples of the present invention will be explained in detail below based on the drawings. Fig. 1 is a schematic configuration diagram of a speech recognition device that is an embodiment of the present invention. 1 is a well-known microphone that converts the voice emitted by the operator into an electrical signal and outputs it as an audio signal.The microphone lft, 1i is connected to the sound cut filter 2 and the pandobus filter 4, respectively. Noise cut filter 2 (9) This filter attenuates the frequency band containing noise components from the noise-containing audio signal input from the microphone 1.
In this embodiment, as shown in Figure 2, it has a correction characteristic that greatly attenuates sounds in the low frequency and high frequency ranges and is most sensitive in the vicinity of 2kHz to 4kHz, matching the human hearing characteristics. ．． Or even if it does not have such auditory correction characteristics, depending on the environment, if there are many high frequency components in the spectral components of the layered sound, LL may occur.
A high cut filter can be used as a noise cut filter. In addition, if you are placed in an environment with noise characteristics that include many low frequency components as shown in Figure 6, 1. Depending on the environment, you can set the frequency characteristics that attenuate the low frequency components as shown in Figure 5. It may be something that you have. If the noise is concentrated in a specific Nakagusuku area, a Nakagusuku cut filter can be used as the noise cut filter.Furthermore, if the noise is concentrated in a specific Nakagusuku area, a Nakagusuku cut filter can be used. A configuration may also be used in which a filter is used and its characteristics are switched depending on the surrounding noise condition, etc. If the filter's attenuation rate and other characteristics are set too strongly, the R sound will be further attenuated, but at the same time the pure audio signal will also be attenuated. Also, if the characteristics are weakened, the attenuation of the noise will be weak, which will affect the detection of voice sections, which will be described later. For example, approximately 2,00}-!z
It cuts components other than the frequency band of ~4KHz. A digital filter may be used as this filter. The noise cut filter 21 is connected to a voice section detection section 6, and this voice section detection section 6 detects the power information of the voice signal output from the noise cut filter 2 and a predetermined threshold value set in advance. By comparison, the start and end of the voice section are detected, and when the level is higher than the threshold level, it is output as a voice section, and when it is lower than the threshold level, it is output as a silent section. This threshold value can be set to one fixed value, but it is also possible to have multiple threshold values or to vary the threshold value according to the surrounding noise. Furthermore, it may be possible to detect a speech gap in conjunction with the slope of the speech spectrum, pitch information, etc. This voice section detection section 61 is connected to the acoustic analysis section 8 together with the bandpass filter 4, and the voice signal in the voice band that has passed through the bandpass filter 4 and the voice section signal are input to the acoustic analysis section 8. In addition to detecting the voice section using the signal that has passed through the accompanying noise cut filter 2, the gain of an amplifier (not shown) may be controlled based on this signal.If the input signal is large, the gain of the amplifier may be controlled. If the gain is small, the dynamic range of the audio signal can be increased by increasing the gain of the amplifier.At this time, if the reference signal contains many noise components, accurate gain control is impossible. By using the signal from which noise components have been cut by the noise cut filter 212 as a reference, more accurate gain control becomes possible.
The spectrum of the audio signal in the audio band that has passed through the bandpass filter 4 within the audio section is analyzed, and characteristic parameters representing the characteristics of the audio are extracted. In this extraction (for example, tf, the well-known fast Fourier transform (FFT)
), 1:. It extracts feature parameters from the audio signal within the input audio section and outputs them as a time series of feature vectors. In the acoustic analysis section 8, the audio signal may be further analyzed by performing processing to reduce the II sound component. This acoustic analysis section 81 can be selectively connected to an audio storage section 12 or an audio comparison section 14 via a changeover switch 10. The voice storage unit 12 (table) Extracted feature data, e.g., time series of melon vectors, is stored as standard data.The voice comparison unit 1
4 (Friend) The standard data stored in this voice storage section 12 and the characteristic data inputted via the changeover switch 10 are compared, and the degree of similarity is calculated and outputted. This voice comparison section 141 The table is connected to the determination unit 16,
Judgment unit 16 (good) If the voice is most similar according to the similarity from the voice comparison unit 14 and satisfies a predetermined condition 1:, it outputs a signal corresponding to it.This signal is output. It is designed to be output from the terminal 18, and other equipment 2 connected to this output terminal 18
0, for example, it is designed to control an air conditioner. Next 1: The operation of this embodiment will be explained. First, when the operator utters a voice, for example, ``Unden,'' the voice is picked up by the microphone 1, converted into an electrical signal, and output as an audio signal. At this time,
The microphone 1 inputs external noises other than the voices before and after the uttered voice, and outputs an audio signal with superimposed noise as shown in Figure 3. This output audio signal (above) is input to the noise cut filter 2 and the bandpass filter 4, respectively.The audio signal 1 input to the noise cut filter 2 is the noise component superimposed on the audio signal as shown in Figure 4. is attenuated and output to the voice section detection section 6.The voice section detecting section 6 compares the power information of the input voice signal with a predetermined threshold value and detects a section lower than this threshold value. is determined to be a silent section, and a section higher than the threshold value is determined to be a voice section, and the voice section signal is output to the acoustic analysis section 8.Low noise cut filter 2
It is possible to give ideal speech recognition analysis by giving characteristics that match the human hearing correction characteristics and performing analysis similar to human senses.

On the other hand, the bandpass filter 4 attenuates the input audio signal except for the audio band and outputs it to the acoustic analysis section 8.
The acoustic analysis section 8 (above) analyzes the speech signal within the speech section and extracts feature data from the speech signal output from the bandpass filter 4 and the speech section signal from the speech section detection section 6. That is, feature data is extracted based only on the audio signal of the section in which the operator uttered the audio,
Other sections {friend) are considered to be noise and no process is performed to extract feature data.

Then, when the changeover switch 10 is switched to the voice storage section 12 side, the feature data output from the acoustic analysis section 8 is stored in the voice storage section 12 as standard data. In this way, the voice storage unit 12 stores voice characteristic data such as words and short sentences uttered by a specific speaker, for example, lf.
When the sound is "Unden J", its characteristic data is stored as standard data. When the selector switch 10 is switched to the sound comparison section 14 side (. The feature data stored in the voice storage section 12 and the standard data stored in the voice storage section 12 are compared in the voice comparison section 14. Then, the voice comparison section 14 calculates the degree of similarity between the characteristic data of one company and the standard data. The determination unit 16 determines what kind of voice the operator has uttered based on the similarity output from the voice comparison unit 14, and outputs a signal corresponding to the voice uttered. .Then, the signal is output to other equipment 20 via the output terminal 18.For example, if it is used in an air conditioner, it will be 1.If the utterance is "Unden", it will be 1.Air conditioning Outputs a signal to start operating the device. Or outputs a signal in response to the utterance to execute control such as increasing the set temperature. However, it is also possible to perform speech recognition for unspecified speakers.As mentioned above, the speech recognition device 1 of this embodiment attenuates the noise by the noise cut filter 2, and the speech section detection unit 6 Detect a voice section based on the voice signal.Then,
Based on this audio section and the audio signal that does not pass through the noise cut filter 2, the acoustic splitting unit 8 extracts feature data.

The changeover switch 10 is switched to store this characteristic data as standard data in the voice storage section 12, and the voice comparison section 14 compares the characteristic data with the standard data. Therefore, the acoustic analysis unit 8 extracts feature data from the speech signal that does not pass through the noise cut filter 2. Therefore, the feature data of the speech is extracted from the speech signal whose speech components have been largely attenuated by the noise cut filter 2. Instead, it is possible to properly extract feature data from the unattenuated audio signal.This allows the features of the audio to be accurately grasped, and the audio comparison unit 14 can calculate the similarity between the feature data and the standard data. Calculation accuracy is improved, and the speech recognition rate is improved.Also, even if the noise overlaps with the spoken voice, by performing speech section detection from the signal whose noise components have been attenuated by the noise cut filter, it is possible to accurately recognize the speech. It is possible to perform section detection.Conventionally, noise overlaps with the speech section of the first uttered voice, making the speech section wide.As a result,
Since special image data was extracted, recognition errors could occur, but by properly detecting speech sections, the recognition rate can be improved. The present invention is not limited to these embodiments in any way, and may be implemented in various forms without departing from the gist of the present invention. [Effects of the Invention] As described in detail above, the speech recognition device of the present invention (attenuating noise components from input speech containing noise components and detecting speech sections from the signal) accurately detects speech sections. It is possible to perform detection and input and analyze audio information without losing any audio information to the acoustic analysis unit, improving recognition performance in noisy environments, especially in non-stationary noisy environments. It has excellent effects.

[Brief explanation of the drawing]

FIG. 1 is a schematic configuration of a speech recognition device as an embodiment of the present invention; FIG. 2 is a graph showing the characteristics of a noise cut filter having audibility correction characteristics; FIG. 3 is a graph showing audio signals from a microphone; Figure 4 is a graph of the signal processed by the noise cut filter, Figure 5 is a graph showing the characteristics of the low cut filter, Figure 6 is a graph showing the level of noise concentrated in the low range, and Figure 7 is the graph of the signal processed by the low cut filter. Figure 6 is a graph of the noise processed level, and Figure 8 is a graph explaining the formant frequency. 1... Microphone 2... Noise cut filter 6... Voice section detection section 8... Acoustic analysis section 1
2...Speech storage section 14...Speech comparison section agent Patent attorney Tsutomu Adachi (2 people) Figure 1 Figure 3 Figure 4 Figure 2 Figure 5 Frequency (Hz) Frequency

Claims (2)

[Claims] (1) A speech recognition device having a voice storage means for storing standard data corresponding to a voice in advance, and a comparison means for comparing the standard data with characteristic data according to a voice signal input from a microphone, a noise cut filter that attenuates and outputs a band containing many noise components in the audio signal; a voice section detection means that detects a voice section using a signal from the noise cut filter; , acoustic analysis means for extracting characteristic data from the voice signal corresponding to the voice section and outputting it to the comparison means. (2) The speech recognition device according to claim 1, wherein the filter is a filter having correction characteristics that match or approximate human hearing characteristics.