JPH0376471B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION [Technical Field of the Invention] The present invention relates to a speech segment detection device capable of stably and reliably detecting a speech segment of isolated word speech.

[Technical background of the invention]

When recognizing a word sound using the overall sound pattern of an isolated word sound, the detection accuracy of the sound section greatly influences the recognition rate. In particular, when detecting this speech section, it is important to prevent so-called addition errors caused by incorporating noise etc. as part of the speech section, and to prevent so-called omission errors caused by defining the speech section with part of the speech missing. Preventing mistakes is an important issue.

However, in the case of a conversational recognition system, which is generally provided in a quiet environment where speech is provided in a quiet environment where the energy S/N ratio is guaranteed to be 30 dB or more, the above-mentioned problem can be solved relatively easily. That is, in this type of system, since the speaker is actively encouraged to speak, a silent period before speaking is guaranteed to some extent. Therefore, if the average value, variance, etc. of the energy and number of zero crossings in this section are determined, the threshold for speech section detection can be set in accordance with the background noise level, and accurate speech section detection becomes possible.

[Problems with background technology]

However, there is no guarantee that the environment in which voice input is performed is necessarily quiet, and it is generally not possible to reliably set the above-mentioned silent period. Furthermore, depending on the speaking environment, the surrounding noise level may be high or the noise level itself may vary. For this reason, conventional systems cannot deal with such background noise in any way, and it is extremely difficult to accurately and stably detect the speech section. For this reason, the recognition rate could not be made sufficiently high due to addition/omission errors.

Furthermore, in order to perform highly accurate speech recognition, it is necessary to stabilize the starting and ending positions of a speech section. For example, if the start point is defined as the start point of the vowel of the first phoneme of a word, it is necessary to detect the start point regardless of the speech level or background noise level. However, it is very difficult to satisfy these requirements using only a single audio parameter or only a single algorithm, and there are also problems in terms of practicality.

[Purpose of the invention]

The present invention has been made in consideration of these circumstances, and its purpose is to provide a voice that is uttered in an environment where the S/N ratio between the voice signal and background noise is poor and the background noise level fluctuates greatly. Accurately detects the beginning and end of the audio, stabilizes the audio section,
Another object of the present invention is to provide a highly practical voice section detection device that can detect with high accuracy.

[Summary of the invention]

The present invention extracts a certain audio parameter of the input audio, detects a general audio section of the input audio using the extracted audio parameter, and then detects a few frames before the tentative start of this audio section, and the above-mentioned audio parameter. The section from the tentative end of the speech section to several frames later is defined as the detection section, and the speech in this detection section is detected using another characteristic parameter of the input speech to accurately identify the start and end of the speech section. It was designed to be established as follows.

〔Effect of the invention〕

Therefore, according to the present invention, a speech section is temporarily determined using a speech parameter that is relatively resistant to noise, and then the start and end points of the speech are uniquely determined with high precision using another characteristic parameter. Since both are detected, it is possible to accurately detect the voice section. In addition, the voice section is tentatively detected using parameters that are resistant to noise, and the section with a predetermined number of frames added before and after the tentatively detected section is used as the detection target to perform highly accurate speech detection. It becomes possible to stably detect a voice section without being affected greatly. Therefore, it is possible to accurately detect voice segments without being affected by the speaking environment, and this provides great practical effects such as improving the voice recognition rate.

[Embodiments of the invention]

Hereinafter, one embodiment of the present invention will be described with reference to the drawings.

FIG. 1 is a schematic configuration diagram of the embodiment device, and FIG. 2 is a diagram for explaining the operation of the embodiment device.

An audio signal input from a microphone via an input amplifier is guided to an audio parameter extraction section 1. This audio parameter extraction unit 1 extracts, for each short time width T, the effective value E of the total band energy of the input audio signal and the output of each channel of the input audio signal extracted via a 4-channel wideband bandpass filter, for example. The effective values B ₁ , B ₂ , B ₃ , and B ₄ are determined and outputted as a time series of audio parameters. E, B ₁ , B ₂ obtained in this way,
The audio parameter time series of B ₃ and B ₄ are temporarily stored in the buffer memory 2 over a predetermined time width. Signal E shown in FIG. 2 shows an example of a voice pattern using energy as a voice parameter.

Therefore, the first stage speech section detection unit 3 inputs the speech parameter time series of energy E, compares the energy with a preset threshold value _E1 ,
The time point _a1 exceeding the threshold E is detected as the temporary start of the voice. This tentative starting point a ₁ detection algorithm is performed by determining the point in time when the input audio energy E exceeds the threshold _{E 1} and continues for a predetermined time as the tentative starting point a ₁ . If the period in which the threshold value _E1 is exceeded does not continue for a predetermined time (predetermined number of frames: 50 to 70 msec), this is regarded as noise and the tentative start point detection is performed again.
Afterwards, the tentative starting point a ₁ detected in this way
The information is given to the threshold calculation section 4. The threshold calculation unit 4 calculates, for example, the tentative starting point from the start of voice input.
Find the average value of the input audio energy E up to a ₁ ,
A predetermined value is added to this to set a threshold value E ₂ for detecting a temporary end, and this is provided to the first stage speech section detection section 3 . First stage speech section detection unit 3
Now, according to this newly given threshold value _E2 , the input audio energy E is below the threshold value _E2 , and for a predetermined period of time (predetermined number of frames: 250 to 300).
msec), the time point _b1 immediately before the energy E falls below the threshold value _E2 is detected as a temporary termination. Therefore, the detection of this temporary termination b ₁ is
Compared to the detection of the tentative starting point _a1 , the background noise level is taken into consideration to some extent. In this way, the first
The step speech section detection unit 3 detects a temporary speech section for the input speech by comparing it with thresholds E ₁ and E ₂ using energy E as a speech parameter to find a temporary start point a ₁ and a temporary end b ₁ . There is. The speech section indicated by the provisional start point a ₁ and provisional end point b ₁ is not affected by a section in which the energy is high due to pulse noise, a silent section in a word, etc.

Now, the labeling section 5 is given information on the tentative start point a ₁ and the tentative end point b ₁ detected by the first stage speech section detection section 3. In addition, the threshold calculation unit 4 reads out the audio parameters E, B ₁ , B ₂ , B ₃ , and B ₄ in the section from the temporary start point a ₁ to the temporary end point b ₁ from the buffer memory 2, and the energy E is the maximum value. The maximum energy value EM at a time point M and each channel output B _1M at this time point M,
B _2M , B _3M , B _4M are determined, and a predetermined value is subtracted from each of the above values for each parameter, and information on labeling thresholds E _T , B _1T , B _2T , B _3T , and B _4T is obtained by the labeling unit. 5 is given.

In the labeling unit 5, based on the information on the tentative starting point _a1 , a point in time several frames before the tentative starting point _a1 ( _a1-
N _F ) is set as the starting point a ₂ for section detection, and from the information on the tentative ending point b ₁ , a time point (b ₁ - N _E ) several frames after this tentative starting point b ₁ is set as the ending point b ₂ for section detection.
It is defined as And from this starting end a ₂ to the ending
The audio parameters are sequentially read from the buffer memory 2 for the section indicated by _b2 , and the previously set labeling thresholds E _T , B _1T , B _2T ,
It is compared with B _3T and B _4T respectively. Then, as shown in the table in Figure 2, at each time point,
The comparison results between each of the above-mentioned audio parameters and their threshold values are registered in sequence. This comparison result can be registered, for example, when the audio parameter is greater than the threshold, the audio element is strong and the data is "1", and when the audio parameter is smaller than the threshold, the audio element is weak and the data is "0". This is done by registering. Then, the table obtained in this way is subjected to, for example, logical sum processing at each time point, and as a result, a time series of "Q" and "V" is obtained. This Q-V table shows "V" as an element of a voice section and "Q" as an element of a silent section.

In the second stage speech section detection unit 6, based on the labeling results "Q" and "V" obtained as above,
Perform section detection for input audio. That is, the detection of the start and end points in this section detection is almost the same as the detection algorithm of the first stage speech section detection section 3 described above, but the detection algorithm is similar to the detection algorithm of the first stage speech section detection section 3 described above, but the speech elements are detected in the time direction by referring to the Q-V table. This is performed by searching for frames labeled "V". First, a frame a labeled "V" is detected, and then it is checked whether the labeling "V" continues for a predetermined number of frames, for example, a number of frames corresponding to 40 to 50 msec. and,
When this condition is met, the frame a is recognized as the beginning of the input audio. Then, first detect frame b labeled "Q",
After that, a predetermined number of frames, e.g. 250-300m
It is detected whether the labeling "Q" continues for sec. If this detection is negative, it is checked whether the frames labeled "V" continue for a predetermined number of frames, for example, a number of frames corresponding to 40 to 50 msec. If this condition is satisfied, it is assumed that another voice section in the word has appeared, and the end detection operation described above is performed again.
If the above conditions are not met, this is assumed to be due to noise and the number of frames labeled "V" is added to the count of "Q". As a result, the start point a and the end point b of the input voice are detected, and the voice section thereof is detected and determined.

Thus, according to the above-described voice section detection performed by the present apparatus, the voice section of the input voice can be detected stably and reliably with high accuracy without being adversely affected by background noise. This is because the beginning and end of a speech section are independent of the level of background noise, and therefore should be determined without being influenced by the level of background noise. Therefore, in the present invention, a threshold value is determined based on the maximum level of input audio, and each audio parameter is labeled in accordance with this threshold value. Moreover, since labeling is determined based on multiple audio parameters, even if energy is concentrated in a certain band,
Even phonemes whose energy level is low overall can be detected reliably. Therefore, it is possible to reliably detect each phoneme in each frame of input speech, and it is possible to accurately detect speech sections. Therefore, it is possible to stably and accurately detect voice segments without being influenced by the speaking environment and even when the background noise level is fluctuating, and its practical advantage is enormous.

Note that the present invention is not limited to the above embodiments. For example, as the audio parameters extracted from the input audio, feature parameters such as LPC prediction errors of various orders and correlation coefficients of audio signals may be employed. It is also useful to use the relationship between the outputs of each channel of the bandpass filter as a characteristic parameter, and the number of channels of this bandpass filter may be determined according to the specifications. In short, the present invention can be implemented with various modifications without departing from the gist thereof.

[Brief explanation of drawings]

FIG. 1 is a schematic diagram of an apparatus according to an embodiment of the present invention;
FIG. 2 is a diagram showing audio parameters and a labeling table for explaining the operation of the embodiment device. DESCRIPTION OF SYMBOLS 1... Speech parameter extraction section, 2... Buffer memory, 3... First stage speech section detection section, 4... Threshold calculation section, 5... Labeling section, 6... Second stage speech section detection section.

Claims (1)

[Scope of Claims] 1. Means for extracting a plurality of types of audio parameter time series of input audio, a memory for storing these extracted audio parameter time series, and a means for extracting a plurality of types of audio parameter time series from the input audio, and a memory for storing the extracted audio parameter time series, and Means for detecting a tentative starting point and a tentative ending point of a voice section from one speech parameter time series, and means for setting an interval from a predetermined number of frames before the tentative starting point to a predetermined number of frames after the tentative end as a labeling target section. and a means for detecting the maximum level of the audio parameter time series in the set labeling target section, and a means for detecting the maximum level of the audio parameter time series in the labeling target section based on this maximum level, at least other than the audio parameter time series used for detecting the tentative start and end. means for labeling whether each frame is silent or not using the stored audio parameter time series; and means for determining the start and end of the audio section of the input audio according to the labeling results. A voice section detection device characterized by the following features. 2. The audio parameter time series used to detect the tentative start and end points is based on the first claim in which the input audio energy is used as a characteristic parameter.
The speech interval detection device described in Section 1. 3. The tentative start point is detected by detecting the point in time when the level of the audio parameter time series with audio energy as a feature parameter exceeds a predetermined threshold and continues for a predetermined number of frames. Claim 1: This is performed by detecting a point in time when the level of the audio parameter time series continues for a predetermined number of frames below a threshold value determined according to the distribution of the audio parameters in the section up to the start end. The voice section detection device described above. 4. The voice section detection device according to claim 1, wherein the detection of the start end and the end end is performed by detecting the time immediately before and after the labeling indicating silence continues for a predetermined frame.