JPS62191895A - Voice recognition equipment - Google Patents

Abstract

(57) [Summary] This bulletin contains application data before electronic filing, so abstract data is not recorded.

Description

[Detailed description of the invention] (Industrial application field) The present invention relates to a speech recognition device.

(Prior art) Speech recognition technology has recently reached the practical stage, but it is still far from perfect.

FIG. 4 is a functional block diagram showing the configuration of the conventional speech recognition device.

Reference numeral 11 denotes a microphone, and the sound or noise inputted from the microphone is A/D converted by a preprocessing circuit 12, and further, anomalous phenomena etc. are removed by passing it through an LPF (low pass filter) to obtain a sample value X. Thereafter, it is applied to the power calculation circuit 13 and the power is calculated. This power is applied to the noise learning circuit 15 by the switching circuit 14,
Switching is performed between selecting the noise learning mode or directly selecting the speech section detection circuit 16 and selecting the speech recognition mode. 17 is a speech recognition circuit, and 18 is a recognition result output circuit.

In the above noise learning mode, the noise learning circuit 15 sets a threshold for detecting a voice section, the voice section detecting circuit 16 detects a voice section based on the set threshold, and the microphone 11 in that section The voice input from the voice recognition circuit 17 is recognized by the voice recognition circuit 17.

The power calculation circuit 13 calculates the unit time (
The power is calculated for each frame (hereinafter referred to as a frame).

However, P(J); A direct X(i) of the J-th frame power;
i-th sample value N within one frame; is the number of samples within one frame.

The noise learning circuit 15 uses the power value P calculated above.
(J) is used to set the threshold value TP for the voice section according to the following equation (2). Here, L is the time required for noise learning, and is a parameter that is arbitrarily set depending on the specifications of the speech recognition device.

TP=TP'+A =ΣP(J)/L,+A...(2) However.

TP; Threshold value P(J) for voice section detection: J
Power value L for the frame Noise learning time A Constant Next, the selector switch 14 is switched to the a side to select the voice recognition mode and perform voice recognition.

FIG. 5 is a flowchart of the voice recognition operation. The input speech to be recognized is preprocessed under the same conditions as in the noise learning mode described above, and the power is calculated (step (hereinafter abbreviated as S) 1). Based on the power time series thus obtained, a voice section detection S2 is performed using the threshold value TP obtained in the noise learning mode.

FIG. 6 shows an example of voice section detection in S2 above.
Power P (J) when saying /akita/ (Akita)
) shows the time series. In Figure 6, the threshold value T
Using P, the chevron part with large power, sl, s,,
s, , s4 and the valley-shaped portion p1. P, , P, and their respective corresponding times SL1 B2983184+
Using the values of PxvPz+Pi, the consistency with the conditions shown in equation (3) below is verified, and the starting point S is determined.

Detect the voice section at end E.

However, T1. T,,T,is a constant.

FIG. 7 shows the steps from setting the threshold TP to detecting the voice section when noise learning is performed under a different environment from that shown in FIG. According to this, it can be seen that there is noise that shows periodic level changes from the time of noise learning until after the sound is uttered.

Therefore, since the threshold value TP is set larger than in the previous case of FIG. 6, the voice section is detected incorrectly as starting point S and ending point E, and the original starting point S is 111 ending point E.
In the uttered part of j, t a /, the beginning / a / is dropped and becomes /kita/. Furthermore, the audio within the audio section is also affected by periodic noise, and its waveform differs from the original audio waveform.

Therefore, when speaking in such a noisy environment, even if it is possible to detect the voice section, the power spectrum structure of the voice changes, resulting in erroneous voice recognition.

(Problems to be Solved by the Invention) As described above, in the conventional speech recognition device, when inputting a voice,
Regardless of whether the ambient noise level or power spectrum is stationary over time, the system performs noise learning, sets the threshold for detecting speech sections, inputs the speech to be recognized, and detects the speech sections. , because voice recognition was performed using this method, there was a high frequency of voice recognition errors.

SUMMARY OF THE INVENTION In view of the above-mentioned drawbacks of the prior art, it is an object of the present invention to provide a speech recognition device with a low error rate.

(Means for Solving the Problems) In order to achieve the above object, the present invention uses speech recognition based on the average value of the noise level and power spectrum and the magnitude of variation thereof during noise learning. To provide a speech recognition device equipped with a display function capable of determining whether or not to perform speech recognition and, based on the judgment result, prompting or inhibiting a speaker who wishes to perform speech recognition.

(Function) According to the present invention configured as described above, when inputting voice,
Based on the surrounding noise level and power spectrum, it is determined whether the noise is suitable for speech recognition or not, and a warning is given to the speaker, making it possible to perform more reliable speech recognition.

(Example) Hereinafter, the present invention will be explained in detail by way of an example using the drawings.

FIG. 1 is a schematic diagram of the configuration of one embodiment of the present invention, and 1.2 is the conventional example shown in FIG. 4. ．． A microphone and preprocessing circuit similar to 12, 3 is a band power calculation circuit,
It is composed of a group of band pass filters (hereinafter abbreviated as BPF). Reference numeral 4 designates a noise learning circuit and determination circuit, 5 a switching circuit for switching between noise mode and voice mode, 6 a lamp lighting circuit that does not allow voice input, and 7 a lamp lighting circuit that allows voice input. Reference numerals 8, 9, and 10 are a voice section detection circuit, a voice recognition circuit, and a recognition result output circuit similar to the conventional example shown in FIG.

In the present invention configured as described above, first, the switching circuit 5 is set to the state a in which voice input is disabled and the noise is inputted, and the noise from the microphone 1 is guided to the preprocessing circuit 2.
/D conversion, passing through an LPF to remove pseudonym phenomena, etc., and inputting it to the band power calculation circuit 3. This band power calculation circuit 3 is composed of a BPF group of multiple channels, and noise power spectrum analysis is performed here.
The output is used as data for setting a threshold value for voice section detection by a noise learning and determining circuit 4, which determines whether the input noise can be used for voice recognition.

Note that the band power calculation circuit 3 calculates the band power for each frame of each channel (three channels in this embodiment) using the following equation (4).

However, Pk(J); Output power value of J-th frame of each channel of 1.2.3 of BPF (however, the same below k=1.2.3) xx(i); of each channel of BPF 1
i-th sample value N within a frame; is the number of samples within one frame.

The noise learning and judgment circuit 4 calculates the power average value PA(k) of each channel within the noise learning time and the standard deviation of its fluctuation from the value of the power Pk(J) according to the following equation (5). ), the range R(k) taken by the value of power Pk(J) is determined from equation (6) below.

R(k)=PA(k)±5(k)...
(6) Next, it is determined whether the power average value PA(k) and the range R(k) obtained above are within a preset threshold range.

Figure 2 shows the analysis results of the noise power spectrum, where the horizontal axis is the channel number and the vertical axis is the power average value PA (k
) and the range R(k) are shown. In addition, in the same figure, the shaded area is the range based on the preset threshold value, and the thick actual M box is the range obtained using equation (6). However, the C1 range R(k) is the preset threshold value. If the input noise is within the range of , the input noise can be used as data for setting a threshold value for detecting a voice section, thereby driving the voice input enable lamp lighting circuit, assuming that the noise is in a voice input enabled state; As in the conventional example, find the threshold value for the input voice section.

Detects speech sections and performs speech recognition.

If the range R(k) is outside the preset threshold, it is assumed that voice input is not possible in the input noisy environment, and the voice input disabled lamp lighting circuit 6 (FIG. 1) is activated. Perform noise learning again.

The present invention has been described above, and FIG. 3 is a flowchart of its operation.

The present invention first inputs noise from the microphone 1, performs noise learning such as pre-processing and power detection (Sl), and when the noise is unsuitable for voice recognition, the voice input disabled lamp lights up. Drives circuit 6 and returns to the initial operation (S2
), if the noise can be recognized by voice, drive the voice input possible lamp lighting circuit 7 (S3) and set the threshold value.
4) As in the case of noise input, preprocessing and power detection are performed (S5), and after voice section detection, voice recognition is performed (S6).

(Effects of the Invention) As is clear from the above description, the present invention is a voice recognition device that can determine whether an external noise environment is suitable for voice recognition and display the determination result. Therefore, if the display is followed, incorrect speech recognition will be extremely rare, and if put into practical use, a great advantage will be obtained.

[Brief explanation of drawings]

Figure 1 is a diagram showing an outline of a voice recognition device as an embodiment of the present invention, Figure 2 is a diagram showing the results of noise power spectrum analysis, Figure 3 is a flowchart of the operation of Figure 1, Figure 4 5 is a flowchart of the operation of FIG. 4, and FIG. 6 is a diagram showing the utterance of a conventional speech recognition device.
Figure 7 is a diagram showing noise learning and voice segment detection using t-B/ as an example. Figure 7 is a diagram showing noise learning and voice segment detection when the same utterance is performed in a different noise environment than that in Figure 6. It is. 2.12...Preprocessing circuit, 3...Band power calculation circuit, 4...Noise learning and judgment circuit, 6.
...Voice input disabled lamp lighting circuit, 7...Voice input possible lamp lighting circuit, 8.16...Voice section detection circuit, 9,17...Voice recognition circuit, 10.18...
- Recognition result output circuit, 13...power calculation circuit. Patent applicant Matsushita Electric Industrial Co., Ltd. Figure 2 (Shuno Campus) Figure 3 Figure 5

Claims (2)

[Claims] (1) In a device that performs speech recognition based on speech sections detected by noise learning, the noise environment is determined based on the average values of the noise level and power spectrum input in the noise learning and the magnitude of their fluctuations. 1. A speech recognition device characterized by having a configuration for determining whether or not the speech is suitable for recognition and, depending on the result of the determination, displaying a message to prompt or inhibit the speaker of the speech to be recognized. (2) As a method for determining the average value of the noise level and power spectrum and the magnitude of fluctuation thereof, a plurality of bandpass filters are used.
1) The speech recognition device described in section 1).