JPS62237498A - Voice section detecting method - 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 Field of Application) The present invention relates to a speech segment detection method in a speech recognition device.

(Prior Art) Conventionally, speech sections have been detected in speech recognition processing. First, let us introduce the fourth aspect of this conventional voice section detection method.
This will be explained with reference to the figures.

According to the conventional voice section detection method, when the input level value of a voice signal is a function S (shi) of time t, the noise level value N when this voice signal is input, the input level value 5 (L), etc. Set a higher level threshold.

When the state in which the input level value 5(L) is large compared to the threshold value (S(j)>L) continues for a certain period of time, that is, the minimum input duration time TS for starting end determination, this duration time TS The start time of is the start of the audio section. After that, when the state in which the input level value S (t) is smaller than the threshold value (S (t) ≦L) continues for a certain period of time, that is, after the termination determination low level input duration time TE, this continuation The voice section was determined by a determination method in which the start time of time TE was taken as the end of the voice section.

In this case, the noise level value is the input level value 5 (L) in a temporally continuous noise measurement time interval TN of a predetermined length from time to when it is assumed that no audio signal is being input. Average values are used. The level threshold value has been generally determined by adding a predetermined constant CO to the noise level value N, such that the value L=N+CO.

According to this method, the time tl in FIG. 4 at which the level threshold value coincides with the input level value S(shi).

t2. t3. Assuming t/I, the start point of the section in which the input level value 5 (L) exceeds this level threshold by more than the aforementioned duration TE, for example time t3, is determined as the start point of the voice section.

Next, the starting point, for example, time t4, of the section where the level threshold is lower than the above-mentioned duration time TE is determined as the end of the voice section.

(Problem to be Solved by the Invention) However, when actually using a speech recognition device in an environment with severe noise fluctuations, the input level value at the noise level value measurement time and the time period from the noise level value measurement time are It is generally believed that the probability that the input level value will take a different value after elapsed time increases as time elapses. Therefore, the probability that the input level value (magnitude) will be different from the noise level value at a time within a relatively short period of time since the noise level value measurement time, for example, just before the start of speech production, is relatively small. There is a high probability that the starting point can be detected stably and accurately using the level threshold set as described above in the conventional manner.

However, since there is a high probability that the input level value (magnitude) will be different from the noise level value at a time when the elapsed time from the noise level value measurement time is relatively long, for example, immediately after the end of voice utterance, when detecting the end , the initially set level threshold is no longer suitable for terminal detection,
This increases the probability that the end cannot be detected accurately, which causes a problem in that speech recognition performance deteriorates.

This invention has been made to solve the above-mentioned problems.

Accordingly, an object of the present invention is to provide a voice section detection method that makes it possible to stably and accurately detect the end of a voice section.

(Means for solving the problem) In order to achieve this objective, according to the speech interval detection method according to the present invention, the following processing is performed (see Figure 1).
.

According to this invention, the temporary termination is determined in the first step, and the true termination is determined in the second step.

In this first stage, the value is smaller than the level threshold LS for start detection obtained by adding a predetermined positive constant aC1 to the noise level value N of the input audio signal, and the value is larger than the noise level value N. A level value is obtained, and this level value is set as E as a level threshold for temporary termination detection. Next, the input level value S(shi) is set to the level threshold for detecting the slope end, and E is predetermined for the temporary termination determination low level input minimum duration TK.
The start time t3 of the section below E or more is determined as the temporary end of that voice section.

Next, as a second step, the magnitude of the input level value immediately after the temporary termination is set as the noise level value NE for termination detection, and the value obtained by adding a predetermined positive constant 03 to this noise level value NE is used for termination detection. Let the level threshold value LE be. Next, the intersection time t2 of the stored input level value S(shi) from the start end to the slope end and this level value LE is detected as the true end. In this case, the level threshold LE for end detection can be set not only to the same value as the level threshold LS for start detection, but also to a value larger or smaller than this level threshold LS.

(Operation) As described above, according to the present invention, the input level value S (t) immediately after detecting the slope end is set as the noise level threshold LE for detecting the true end, so the end detection is performed based on the temporal It is less susceptible to fluctuations, so detection of voice sections is stable and accurate.

(Embodiments) Hereinafter, embodiments of the voice section detection method of the present invention will be described with reference to the drawings.

FIG. 1 is an explanatory diagram for explaining the voice section detection method of the present invention, and shows an example of changes in input level values, with time t plotted on the horizontal axis and level values plotted on the vertical axis. FIG. 2 is a block diagram showing an example of the configuration of a speech section detection section for implementing the speech section detection method of the present invention. Moreover, FIG. 3 is a flowchart of processing provided for detailed explanation of the present invention.

The speech section detection section shown in FIG. 2 includes a level extraction section 11, a provisional speech section detection threshold setting section 12, and a provisional speech section detection section I3.
, a level storage section I4, a termination detection threshold setting section 15, a termination detection section 16, and a control section 17. still,
In the following description, processing steps in the flowchart are denoted by S.

First, processing is started (Sl). input audio signal a
1 is input to the level detection section 11, and its level is extracted and converted into an input level signal a2 (S2). The input level value of this input level signal a2 is defined as sB) and is shown by a solid line in FIG. This input level signal a2 is outputted to the provisional speech section detection threshold setting section (2), the provisional speech section detection section 13, and the level storage section 14.

The control section 22 outputs a temporary voice section detection threshold setting command signal rl to the temporary voice section detection threshold setting section 12 at a time when it is assumed that the voice is not being uttered.

The provisional speech section detection threshold setting unit I2 receives the provisional speech section detection threshold setting command signal r1 at time t.

The input level signal a2 is received for a predetermined time period TSN, and the input level signal S during this time period is
The average value of (shi) is set as the noise level value N, that is. Next, a value obtained by adding a ninety-plus constant C1 learned and determined in advance to this noise level value N is determined as a level threshold LS for start edge detection (S3).

LS=N+C1 The signal b1 of this level threshold LS is sent to the temporary speech section detection unit 13.
send to

Next, in the provisional speech interval detection threshold setting unit 12, a positive value determined by r learning, which has a smaller value than the constant C1 used for setting the level threshold for start detection, is determined for this noise level value N. The value obtained by adding the constant 02 is used for slope edge detection and is determined as the health threshold LKE (S4).

LKE=N+C2 The method of determining the level threshold LKE for detecting the slope end is to calculate the level threshold LKE for starting edge detection, which is obtained by adding a predetermined positive constant CI to the noise level value N of the input audio signal, to a predetermined positive value LKE. By subtracting the constant C'2, a level value smaller than this level threshold value LS and larger than the noise level value N is obtained, and this level value is set as E as the level threshold value for slope end detection. Also good.

The provisional voice section detection threshold setting unit 12 uses this level threshold LK.
E signal b2 is sent to the temporary voice section detection section 13, and a temporary voice section detection threshold setting end signal b3 is sent to the control section 17.

The control unit 17 receives this provisional voice section detection threshold upper determination end signal b.
3 is supplied, a temporary voice section detection command signal r2 is output to the temporary voice section detection section 13.

The temporary voice section detection unit 13 receives a temporary voice section detection command signal r2.
After receiving the human level signal a2. Detection of a temporary speech section is started by inputting the level threshold value b2 for detecting the start end and the level threshold value b3 for detecting the temporary end point, and the start end and the temporary end point are detected.

In this process of detecting the starting point time tl, the input level value 5(t) changes over time to the starting point detection level threshold LS.
From time t1, which coincides with , this input level value 5(t) becomes
The starting point determination high level input minimum duration time TS predetermined by learning must be greater than this level threshold value LS5(t)>LS However, when a time period of 73 or more is detected, this duration time TS The aforementioned start time t1 of
is determined to be the starting point of the voice section (S5).

In addition, in the temporary termination detection process, the input level value 5 (
From time t3 when L) decreases in level after detecting the start end and matches the level threshold LKE for provisional end detection, this input level value S(shi) changes to the slope end determination low level input minimum duration time predetermined by learning. The value must be greater than TKE and less than this level threshold LS5(L)≦LKE However, when the duration time TKE or more is detected, the above-mentioned start time t of this duration time TKH
3 is determined as the temporary end of the temporary voice section (S6).

The signal d1 at the start time t1 detected in this way is output to the level storage section 14, the end detection section 16, and the control section 17 (S5), and the signal d2 at the detected temporary end time t3 is output to the level storage section 14. , end detection threshold setting unit 15,
The signal is output to the termination detection section 16 and the control section 17 (S6).

Signals d1 and d2 at the starting point time t1 and the tentative ending point time t3 are input to the level storage section 14, respectively.

When the start end time signal di is input, storage of the input level value 5(t) of the input level signal a2 is started from the start end time tl, and this input level value is stored for a predetermined time predetermined by learning from the temporary end time t3. Continue until the time elapses.

After receiving the provisional end time d2 from the provisional speech section detection section 13, the control section 17 outputs the end detection threshold setting command signal r3 to the end detection threshold setting section 15.

After receiving the end detection threshold setting command signal r3 from the control section 17, the end detection threshold setting section 15 sets the time axis from time t3 given by the temporary end time signal d2 from the temporary voice section detection section 13. The input level value S(shi) for a predetermined termination detection noise measurement time TEN in the positive direction is taken from the level storage unit 14 as the termination detection noise level signal e1. Then, the average value NE of the input level values 5(t) during this noise measurement time TEN is set as the termination detection noise level value NE. Subsequently, a positive constant C3 determined in advance by learning is added to this noise level value NE to set it as a level threshold value LE for end detection (S7). That is, LE=NE+C3 In this case, by selecting this constant 03,
The level threshold LE for end detection is set to the level threshold LE for start end detection.
Of course, this level threshold LE should be the same value as LE.
It is possible to set it to a value larger or smaller than ``'4-''.

The signal f1 of this level threshold LE is output to the termination detection section 16, and the termination detection threshold setting completion signal f2 is output to the control section 17.

When the control unit 17 receives the termination detection threshold setting completion 16, it outputs a termination detection command signal r4 to the termination detection unit 16.

After receiving the end detection command signal r4, the end detection section 16 receives the start end time signal dl input from the temporary voice section detection section 13.
A signal e with an input level value 5(t) of a temporary voice section from time ti determined by d2 to time t3 determined by the temporary end time signal d2.
2 from the level storage section 14, and further receives the termination detection level threshold signal fl from the termination detection threshold setting section 15.

Then, in this termination detection section 16, these signals e2
The input level value 5(t) of the temporary voice section determined by and fl is compared in magnitude with the level threshold value LE for end detection in the negative direction of the time axis from the temporary end time t3, and the input level of the temporary voice section is The time when the value 5(t) becomes larger than the end detection level value LE for the first time, for example t2, is detected as the true end time (S8).

The signal g of the end time determined in this way is sent to the control unit 1.
7, and this voice section detection process ends (
S9).

(Effects of the Invention) As is clear from the above description, in this invention, in order to cope with noise fluctuations when using a speech recognition device, the level threshold for termination detection is set to the average value of the input level values immediately after temporary termination. This is done based on. Therefore, since end detection can be performed without being affected by temporal fluctuations in the ambient noise level, stable and accurate 1:f inter-intentional detection can be performed, and improvement in the recognition performance of the speech recognition device can therefore be expected.

[Brief explanation of drawings]

FIG. 1 is an explanatory diagram showing how the input level value of an input audio signal changes over time to explain an embodiment of the voice section detection method according to the present invention, and FIG. 2 is an explanation of the voice section detection method according to the present invention. FIG. 3 is a flowchart of the processing of the speech section detection method of the present invention, and FIG. 4 is an explanatory diagram of the conventional speech section detection method. 11...Level extractor I2...Temporary voice section detection threshold setting section 13...Temporary voice section detection section 14...Level storage section 15...Temporary voice section detection threshold setting section 16...End Detection unit, 17... Control unit. Patent applicant: Oki Electric Industry Co., Ltd.

Claims (2)

[Claims] (1) The average value of the input level values of the input audio signal during a predetermined time period before the audio section of the input audio signal is taken as the noise level, and the level threshold set based on the noise level and the input level When detecting a voice section by comparing the noise level with a value, the value is smaller than the level threshold for start detection, which is obtained by adding a predetermined positive constant to the noise level value of the input voice signal, and the noise level is A level value larger than the above value is set as a temporary termination detection level threshold, and the input level value is lower than the temporary termination detection level threshold by a predetermined temporary termination determination low level input minimum duration time or more. The start time is determined as the temporary end of the speech section, the magnitude of the input level value immediately after the temporary end is set as the noise level value for end detection, and the value obtained by adding a predetermined positive constant to the noise level value is determined. Set as a level threshold for termination detection, and the input level value stored from the starting edge to the temporary termination,
A voice section detection method characterized by detecting a time closest to a provisional end among times of intersection with the end detection level threshold as a true end. (2) The level threshold for detecting the end point is set to the same value as the level threshold value for detecting the starting point, a value larger than the level threshold value, or a value smaller than the level threshold value.
The speech interval detection method described in Section 1.