JPS6223096A - Detection of voice section - 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 (Field of Industrial Application) The present invention relates to a speech segment detection method in a speech recognition device.

(Prior Art) FIG. 2 is a functional block diagram for executing an example of a noise learning method in a conventional speech recognition device.

A conventional example will be explained with reference to FIGS. 2 and 3.

In FIG. 2, reference numeral 10 denotes a microphone from which voice or noise is input, a preprocessing section 11 performs preprocessing, and a power calculating section 12 calculates the power. 13 is a changeover switch for switching between voice recognition mode a and noise learning mode b. 14 is a noise learning section, and 15 is a speech section detection section, which sets a threshold value for detecting a speech section in the noise learning mode, and detects the speech section in the speech section detection section. The voice recognition unit 16 recognizes the voice within the detected voice section. 17 is a recognition result output unit.

Next, the operation of the above conventional example will be explained. First, before entering the voice recognition mode, the changeover switch 13 is used to select the noise learning mode to perform noise learning. In FIG. 2, the noise manually generated from the microphone 10 is processed in the preprocessing section H.
The signal is A/D converted, passes through an LPF to remove pseudonym phenomena, and obtains a sample value X. Next, in the power calculation unit 12, according to equation (1), the unit time (hereinafter referred to as frame)
The power P(J) is calculated for each time.

P(J): Power value of the J-th frame x(i): First sample value within one frame N: Number of samples within one frame The noise learning unit 14 calculates (2 ) The threshold value TP of the voice section is set according to the equation, where L is the time required for noise learning, and is a parameter that is arbitrarily set according to the specifications of the recognition device.

Note that the value of TP' is the average power of environmental noise during the learning time.

TP= TP'+ A = Σ P (J)/L + A...(2
) J=1 TP: Threshold value for voice section detection P (J): Power value of J frame L: Noise learning time A: Constant Next, select the recognition mode with the changeover switch 13 and perform voice recognition. .

In FIG. 2, the input voice is preprocessed under the same conditions as in the noise learning mode, and the power is calculated. Based on the obtained power time series, a voice section is detected using the threshold value TP obtained in the noise learning mode. Figure 3 shows /a
It shows the time series of P(j) when uttering kita/ (Akita).

In FIG. 2, using the threshold value TP, the chevron-shaped portions Sl, S2, . Detect S3 and the valley-shaped part Pl, P21 P3 sandwiched between the chevron-shaped part, and calculate the corresponding times SL+32+ S3 and pl+ 'I'2+
The value of p3 is used to test whether it matches the condition shown in equation (3), and the voice section, the start point S, and the end point E of the voice are detected.

Figure 4 shows learning under different noise conditions than in Figure 3.
It shows the process from setting the threshold TP to detecting the voice section when uttering ta/ (Akita). Shocking noise occurs during the time required to set the threshold value, and the level is louder than when the voice is being uttered. For this reason, the threshold value TP is set larger than in the example shown in FIG.
The /a/ at the beginning of the word is dropped, making it sound like /k it a/. Therefore, in the conventional method, as in the case of FIG. 4, when the noise level during noise learning and the noise level during voice production are significantly different, the voice section detection is incorrect.

(Problems to be Solved by the Invention) In the conventional speech section detection method described above, during the time required to learn the threshold setting, due to shocking noise, etc., the noise level is lower than the noise level when the speech is being uttered. There was a problem that the noise level during learning was overestimated, leading to incorrect threshold settings and, in turn, incorrect speech recognition.

The present invention solves these conventional problems,
It is an object of the present invention to provide a voice section detection method that can detect voice sections with high accuracy.

In order to achieve the above object, the present invention, when performing noise learning, does not set a threshold value from the noise data of all frames taken into the learning time, but uses a threshold value set in advance. The threshold value is set using only the noise data within the range.

(Function) Therefore, according to the present invention, by selectively handling the noise data for learning, it is possible to reduce threshold setting errors due to impact noise during noise learning, and it is possible to accurately detect speech sections. As a result, speech recognition errors can be reduced.

(Example) Below, the configuration of an example of the present invention will be described with reference to FIG.

In FIG. 1, a microphone 1, a preprocessing section 2, a power calculation section 3, a noise learning section 6, a speech section detection section 7, a speech recognition section 8
．． The recognition result output unit 9 is similar to the conventional example. 5
is the noise data selection section.

Next, the operation of the embodiment of the present invention will be explained.

First, the mode changeover switch 4 is set to the noise learning mode. The noise input from microphone 1 is processed by preprocessing section 2 A/
The D-converted signal passes through the LPF and enters the power calculation unit 3, where the power of the frame is calculated according to equation (1) as in the conventional example. The noise data selection unit 5 selects noise data to be used for threshold setting for voice section detection. In this method, only noise data whose manually inputted noise level falls within a preset range during noise learning is used as threshold value setting data.

This range is determined as follows.

In FIG. 3, the portion of ρ3 after the rear end E of the voice section is
Since the analysis is performed before determining the voice section, that is, before determining E, the power for each frame in the section p3 has been calculated. In the conventional method, the data in the section p3 is
Once a voice section is determined, it is discarded, but in the present invention, the average value Nε and the variance σε of the power for each frame in this section are determined according to equation (4) and sent to the noise data selection section 5. The noise data selection unit 5 newly calculates N and σ2 from the average value Np and variance σP2 of the noise level used in setting the threshold for voice section detection, NE and σε2 according to equation (5).

Only the noise data whose input noise level is within the range of N±σ using N and σ2 are used as the noise data for setting the threshold value.

The initial values of Np and σP2 are determined according to equation (6).

Using noise data in the range of N±σ, a threshold value TPx for detecting a voice section is set according to equation (7) in the same way as in the conventional example, and the voice section is detected using this threshold value TPX.

P(I): First noise power value within the range of N±σ within the learning time interval M: Number of samples of noise data within the range of N±σ within the learning time interval B: Constant TPX FIG. 4 shows a case where a voice section is detected using the method. In this figure, the starting end is S×, and the trailing end is E, and it can be seen that unlike the conventional example, the /a/ at the beginning of the word /akita/ is not dropped, and the speech section can be detected properly.

As described above, according to this embodiment, even if shocking noise enters noise learning, it will not be used as learning data unless the noise level is within a preset range, so it is possible to incorrectly set the threshold value for speech section detection. There is no. Therefore, it has the advantage that voice sections can be detected with high accuracy.

(Effects of the Invention) As is clear from the above explanation, when performing noise learning, the present invention does not set a threshold value from the noise data of all frames captured during the learning time, but instead uses noise within a preset range. Since the threshold value is set using only data, there is an advantage that errors in setting the threshold value for voice section detection can be reduced and voice sections can be detected with high accuracy. Furthermore, since speech sections can be detected with high accuracy, it has the effect of improving the speech recognition rate.

[Brief explanation of the drawing]

FIG. 1 is a schematic block diagram of a speech recognition device in one embodiment of the present invention. FIG. 2 is a schematic block diagram of a conventional speech recognition device. FIG. 3 shows temporal changes in noise power and voice power when /akita/ is uttered at a certain noise level. FIG. 4 shows temporal changes in noise power and voice power when /akita/ is uttered in an environment different from that in FIG. 3. l... Microphone, 2... Preprocessing section, 3... Power detection section, 4... Changeover switch, 5... Noise data selection section, 6... Noise learning section, 7... - Speech section detection unit, 8... speech recognition unit, 9... recognition result output unit. Patent applicant: Matsushita Electric Industrial Co., Ltd.

Claims (2)

[Claims] (1) In the method of detecting speech sections by learning the noise level and adaptively setting the threshold for speech section detection, the noise level is determined in advance based on the average value and the magnitude of fluctuation of the noise level. A voice section detection method that limits a range and uses only noise level data within that range as data for setting a threshold for voice section detection during noise learning. (2) Claim (1) characterized in that the method for determining the preset noise level range uses the portion after the rear end of the voice section detection as input data for determining the range. The voice interval detection method described.