JPS6147437B2 - - Google Patents

Description

[Detailed Description of the Invention] In the present invention, the voice is expressed as shown in equation (1).
It is assumed that the input is expressed as a series of feature vectors obtained by sampling at certain time intervals.

X ₁ , X ₂ ,...,X _N (1) In the above formula, each Xi (i=1, 2,...,
N) are each m-dimensional vectors, and are expressed as Xi=(x _i1 , x _i2 , . . . , x _in ) (2). Here, for example, the outputs of the m-channel bandpass filter x ₁ (t), x ₂ (t), ...,
The feature vector can be captured as a time sample of x _n (t).

Furthermore, a section of audio represented by a feature vector is called a frame. The subscripts 1, 2, . . . , n in equation (1) are parameters representing time (ie, frame).

The series of feature parameters is, for example, a series of energies, and if the series of speech feature vectors shown in equation (1) is the output of a filter bank, the energy of the feature vector X _i is Pi= Σ ^m _j=1 x ² _ij (3) is defined. The energy sequence corresponding to the feature vector sequence in equation (1) is expressed as P ₁ , P ₂ , . . . , P _N (4) from equation (3).

With respect to fluctuations in the energy sequence shown by equation (4), sections that exceed a preset threshold are detected as speech sections, which is one of the important issues in automatic speech recognition. . The problem of speech interval detection can be divided into two sub-problems: detection of the start end, and detection of the end.

The present invention is concerned with one of these two sub-problems, and is aimed at detecting the termination.

By the way, when detecting speech intervals, the average value of the input level fluctuates for each utterance, not only for different speakers but also for the same speaker, and the energy sequence may fluctuate irregularly due to noise added to the uttered voice. accompanied by. Due to these fluctuations, the speech interval cannot be detected correctly, which often results in erroneous recognition of speech.

The present invention absorbs fluctuations in the average value of the input level and irregular fluctuations due to noise by adapting to fluctuations in the energy series and setting thresholds, and observes for a preset period of time to improve accuracy. The purpose is to detect speech intervals.

However, in the present invention, the series of equation (4) is used not only for the energy as given by equation (3).

Pi=(Σ ^m _j=1 x ² _ij ) ^1/2 (5) Pi=Σ ^m _j=1 |x _ij | (6) It is also valid to define. That is, P _i
is equivalent to the energy in Equation (3), and is sufficient as long as it can express the characteristics of the audio input level.

This paper describes a method that has been proposed in the past for detecting the end of a speech interval from an energy sequence, and describes its problems.

Conventionally proposed methods for detecting the end of a speech section do not correct the speech section if the predetermined energy level threshold is too high due to irregular fluctuations due to noise or fluctuations in the average level during speech. It is easy to falsely detect the end at a point earlier than the end. Also,
If the predetermined energy level threshold is too low, false positives due to fluctuations in the average level during speech will be eliminated to a large extent, but irregular fluctuations due to noise may cause false positives to occur at the end of the speech interval. The disadvantage is that it is easy to falsely detect the end at a late point.

Therefore, in order to correctly detect the end of a speech interval, it is necessary to observe the fluctuations in the input level of the speech sound, as well as the amount of noise added to the speech, and to determine the energy level that is appropriate for each speech sound. It is necessary to determine the threshold value.

The termination detection method described below was developed against this background. An example of a configuration for realizing the present invention will be shown and described.

FIG. 1 shows an example of a configuration for detecting the end of a speech section using an adaptively determined threshold.

The local maximum value determination unit indicated by 1 observes fluctuations in the input level, and each time a local maximum value of the energy level is detected, the maximum value is given to the threshold value determination unit indicated by 2. In the threshold value determination unit 2, when the output of the local maximum value determination unit 1, that is, the local maximum value exceeds the current threshold value, a new threshold value is determined based on the local maximum value. Even if a maximum value exists in an interval where the energy sequence does not exceed the current threshold, the threshold remains the same. Now, when the threshold value is θ and the maximum value is α, θ is given by θ=f(α) (7) for a certain function f. Specifically, it is given by, for example, θ=max(α/16,25) (8). Here, max(·) is a function that gives the maximum value in parentheses. The first threshold is given by equation (8) using the first local maximum value.

The state determining unit 3 determines whether the energy sequence is larger or smaller than the threshold based on the output of the threshold determining unit 2, that is, the threshold.
The state sequence holding unit indicated by 4 holds the output of the state determining unit 3, and stores the energy only after a preset time, that is, when the energy sequence becomes below the threshold value continuously for a preset number of frames or more. The time point at which the sequence becomes less than or equal to the threshold value is output to the end detection section indicated by 5. The end detection section 5 detects the end frame from the energy sequence based on the output of the state holding section 4 and outputs it.

FIG. 2 is an example of an energy series for specifically explaining the operation of the configuration example shown in FIG. 1. In the figure, the horizontal axis is time, ie, frames, and the vertical axis is energy level. The energy series corresponds to discrete points on the curve shown in the figure, but for convenience of notation, it is shown as a continuous curve.

The local maximum value determination unit 1 detects the local maximum value θ ₁ and provides α ₁ to the threshold value determination unit 2 . In the threshold determination unit 2, the formula
The threshold value θ ₁ is determined based on (8). The state determination unit 3 determines whether the energy sequence is larger or smaller than θ ₁ based on the threshold value θ ₁ and outputs it. The state series holding unit 4 holds the output from the state determining unit 3, and holds the output from the state determining unit 3 for a preset time _T.
We are observing whether the following energies are continuous, but in the example in Figure 2, with respect to θ ₁ , the energy series exceeds θ ₁ within T, so the energy series becomes smaller than θ ₁ for the first time. At the point of getting old
_t1 is not output to the termination detector. Similarly,
Although the threshold value θ ₂ is determined based on the maximum value α ₂ , there is no output to the end detection unit 5 . However, the maximum value α ₄
does not exceed the threshold θ, so the threshold θ remains unchanged until the local maximum value _α3 . However, the maximum value α
Regarding the threshold value θ ₃ determined based on _{θ 3}
Since the time below continues for T or more, the time t ₃ is when the energy series becomes smaller than θ ₃ for the first time.
is detected as the termination, and is output from the termination detection section 5.

As described above, the present invention is an end detection method that performs more accurate end detection by adaptively determining a threshold for end detection in accordance with fluctuations in the average value of the audio input level. Compared to the conventional method, the threshold for end detection is determined adaptively according to fluctuations in the average value of the audio input level.
It is possible to reduce false termination detections caused by fluctuations in the average value of the input level or irregular fluctuations in the energy sequence due to noise, and more accurate termination detection can be performed.

It has excellent characteristics such as. Accurately detecting speech intervals will lead to improved accuracy in automatic speech recognition, and will therefore absorb irregular fluctuations in the energy sequence due to input level fluctuations and noise, resulting in more accurate speech recognition. The present invention, which performs termination detection, is extremely effective in automatic speech recognition.

[Brief explanation of the drawing]

FIG. 1 is a block diagram showing an embodiment of a method for detecting the end of a voice section, and FIG. 2 is an energy series diagram for explaining the operation of the block diagram of FIG. 1...Local maximum value determination unit, 2...Threshold value determination unit, 3...
...state determining section, 4... state series holding section, 5... termination detecting section.

Claims (1)

[Claims] 1. Means for sampling audio input at certain time intervals to generate a sequence of audio feature vectors X ₁ , X ₂ , ..., X _N , and representing the audio input level using X ₁ , X ₂ , ..., _X Sequence of parameters P ₁ , P ₂ ,..., P
means for generating _N , means for setting thresholds for P ₁ , P ₂ ,...P _N , and detecting that P ₁ , P ₂ ,..., P _N do not exceed the thresholds. Means for determining the termination and the series of characteristic parameters P
_i (i=1, 2,...N), threshold determining means for comparing the maximum value and the threshold and setting the larger value as the threshold, and feature parameter P _i (i= 1, 2, ..., N) does not exceed the threshold set by the threshold determining means for a certain period of time, the end of the speech section is determined at the point in time when P _i becomes smaller than the threshold for the first time. A device for detecting the end of a voice section, comprising means for determining the end of a voice section.