JPH0588840B2 - - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a voice detection method used in a voice recognition device or the like to determine the time when voice exists.

(Prior Art) Conventionally, in this type of voice detection method, the duration of the voice power level exceeding the threshold value is determined to be the beginning of the voice when the duration time is longer than a certain period of time, and the duration of the voice power level exceeding the threshold value is determined as the beginning of the voice, and the continuation of the voice power level exceeding the threshold value A method in which audio ends when the time exceeds a certain time has often been used.

(Problems to be Solved by the Invention) In the conventional speech detection method described above, speech sections are detected based on the duration of the level, so in the case of speech with a deep power dip, the beginning of the word may be missing. Even if the noise is instantaneous, if the noise is close to the end of a word, the end will be extended and the noise will be included in the speech section. As described above, the conventional speech detection method has the drawback of erroneously detecting speech sections.

(Means for Solving the Problems) Means provided by the present invention to solve the above-mentioned problems includes a power calculation unit that calculates the power of an audio signal, and a power calculation unit that smooths the power calculated by the power calculation unit. a power smoothing unit that obtains a smoothed power by converting the rate of change into a smoothed power; a peak detection unit that detects an inflection point where the rate of change of the smoothed power changes from positive to negative as a peak candidate of the smoothed power; Selecting the peak candidate with the highest level as the maximum peak, calculating a peak selection threshold from the level of this maximum peak and a predetermined peak selection coefficient,
a threshold calculation unit that calculates a peak width calculation threshold from the maximum peak level and a predetermined peak width calculation coefficient, and a comparison between the level of the peak candidate detected by the detection unit and the peak selection threshold; a selection unit that selects only the peak candidates whose level is equal to or higher than the peak selection threshold as peaks; a peak width calculation unit that calculates the time including the peak as a peak width; a peak width comparison unit that outputs a peak width wider than a predetermined peak width threshold value as a voice section candidate among the peak widths; If there are a plurality of speech segment candidates, the time from the end of the previous speech segment candidate to the start of the next speech segment candidate among the adjacent speech segment candidates is calculated as a speech segment candidate time difference. The speech segment candidate time difference calculation unit includes a speech segment candidate time difference calculation unit, and a speech segment determination unit that determines a speech segment based on the output results of the peak width comparison unit and the speech segment candidate time difference calculation unit. If there is one speech segment candidate, the speech segment candidate is directly determined as the speech segment, and if there are multiple speech segment candidates and the speech segment candidate time difference between the adjacent speech segment candidates is shorter than a predetermined speech segment candidate time difference threshold. In such a case, the plurality of speech segment candidates are combined into one speech segment candidate, and the speech segment candidates are summarized so that a new speech segment candidate extends from the start of the previous speech segment candidate to the end of the subsequent speech segment candidate. Processing is performed, and this process of summarizing the speech segment candidates is repeated until one of the speech segment candidates that remains.
The voice section is characterized in that one or more of the voice sections are defined as the voice section.

(Example) Next, the present invention will be described with reference to the drawings.
FIG. 1 is a block diagram of one embodiment of the present invention.

This embodiment includes a power calculation section 1, a power smoothing section 2, a peak detection section 3, a threshold calculation section 4, a peak selection section 5, a peak width calculation section 6, a peak width comparison section 7,
It is composed of a speech section determination section 8 and a speech section candidate time difference calculation section 22. The input voice 10 is input to the power calculation section 1, the calculated power 11 is input to the power smoothing section 2, and the smoothed power 12 is inputted to the power calculation section 1.
is input to the peak detection section 3 and the peak width calculation section 6. The peak detection unit 3 detects an inflection point where the rate of change of the smoothed power 12 changes from positive to negative as a peak candidate 13 of the smoothed power, and applies the detected peak candidate 13 to the threshold calculation unit 4 and the peak selection unit 5. Output to. The threshold calculation unit 4 calculates the maximum peak lever from among the peak candidates 13, performs calculations with the peak selection coefficient 19, calculates the peak selection threshold 14, outputs it to the peak selection unit 5, and also calculates the maximum peak level and peak A calculation is performed with the width calculation coefficient 20 to calculate a peak width calculation threshold 15 and output it to the peak width calculation section 6. The peak selection unit 5 compares the peak level of the peak candidate 13 with the peak selection threshold 14 and outputs only peak candidates having a peak level equal to or higher than the threshold to the peak width calculation unit 6 as peaks 25 .

The peak width calculation unit 6 calculates a peak width calculation threshold value 15.
The time interval in which the smoothed power is 12 and includes the peak 25 is defined as the peak width 1
Output as 6. Peak width 16 is equal to peak 25
is the peak of smoothed power 12 specified by
It represents a time section of smoothed power 12 that is equal to or greater than the peak width calculation threshold value 15. This peak width 1
6 is output to the peak width comparison section 7. The peak width comparison unit 7 compares the peak width threshold 21 with the peak width 16 of each peak, and selects the starting and ending ends of the peaks having a peak width equal to or greater than the threshold 21 as voice interval candidates 17, and selects the voice interval determination unit 8 and the voice interval. It is output to the candidate time difference calculating section 22.

When there is only one speech segment candidate 17, the speech segment determination unit 8 converts the speech segment candidate 17 into a speech segment 18.
Output as . When there are multiple speech segment candidates 17, the speech segment candidate time difference calculation unit 22 calculates the speech segment candidate time difference 23, which is the time from the end of the previous speech segment candidate to the start of the next speech segment candidate among the adjacent speech segment candidates. Calculated as At this time, if the speech segment candidate time difference 23 of adjacent speech segment candidates is smaller than the speech segment candidate time difference threshold 24, the speech segment determination unit 8 repeatedly performs the process of combining the speech segment candidates into one speech segment candidate, and finally If it becomes one, the voice section candidate that has undergone the grouping process is output as voice section 18, and if it has not become one, the voice section candidate that has undergone the grouping process and has the maximum peak level is output as voice section 18. Output as .

In this way, when a plurality of speech segment candidates end up, the method of discarding speech segment candidates other than those having the maximum peak level is effective in removing noise segments. However, the voice section determination unit 8
In this method, when the speech section candidates become complex, it is effective to separate the speech sections when uttering continuous speech by using a method that determines each speech section candidate as a separate speech section. it is obvious.

FIG. 2 is a diagram showing the relationship between the waveform of the smoothed voice power 12, the voice detection threshold, and the detected voice section in the embodiment of FIG. 1. With reference to FIG. 2, it will be seen that according to the embodiment of FIG. 1, noise with a low peak level and noise close to speech can be removed, and even speech with a deep power dip can be prevented from missing the beginning of a word. Further, a detailed explanation will be given below in connection with FIG. The horizontal axis 30 in FIG. 2 represents time, and the vertical axis 31 represents smoothed power, and the waveform in this figure shows the waveform of the smoothed power 12 that is the output of the power smoothing section 2 in FIG.

The peak detection section 3 in FIG. 1 detects four peak candidates 32, 33, 34, and 35 in FIG. 2, and the threshold calculation section 4 in FIG. A peak selection threshold value 14 and a peak width calculation value value 15 are calculated from the above.
In the peak selection section 5, peak candidates 32 with low peak levels are removed using a peak selection threshold 14, and peak candidates 33, 34, and 35 are output as peaks. The peak width calculation section 6 calculates the peak widths 38, 39, and 40 of the peaks 33, 34, and 3 using the peak width calculation threshold 15, and the peak width comparison section 7 calculates the peak widths 38, 39, and 40 using the peak width threshold 21 and the peak widths 38, 3, respectively.
9 and 40, all peak widths are wider than the threshold value 21 in the example of FIG.
The sections from the start to the end of each of 4 and 35 are output as voice section candidates 17.

The voice section candidate time difference calculation unit 22 calculates peak 3
3 and 34 voice section candidate time difference 43 and peak 3
4 and the voice section candidate time difference 44 between the peak 35 and the peak 35 are calculated. The speech section determination unit 8 uses a speech section candidate time difference threshold 24 and each speech section candidate time difference 43, 4.
4, the voice section candidate time difference 43 is the threshold value 2.
Since it is shorter than 4, the voice section candidates of peaks 33 and 34 are combined into one, and the voice section candidates from peak 33 to peak 3 are
4 is set as a new speech section candidate, and since the speech section candidate time difference 44 is wider than the threshold value 24, the peaks 35 cannot be combined, and two speech section candidates remain. The speech section determination unit 8 then compares the peak levels of the two speech section candidates,
From the starting end 41 to the ending end 42 having the maximum peak 34
The speech section candidates up to this point are determined to be speech section 18 and output.

Therefore, according to the embodiment shown in FIG. 1, as in the example shown in FIG.
Even if the voice is separated into 34 parts, the voice section can be detected correctly.

(Effects of the Invention) As described above, the present invention detects the peak of a smoothed power waveform, calculates the peak selection threshold and the peak width calculation threshold from the level of the peak with the maximum level, The peak width of a peak having a peak level equal to or greater than a peak selection threshold is calculated using a peak width calculation threshold, and a peak whose peak width is equal to or greater than a predetermined width is determined to be a speech section candidate. If there are multiple peaks, calculate the voice interval time difference, and if it is shorter than a predetermined time, repeat the process of combining them into one voice interval, and if the peak does not become one in the end, one of them (e.g. By determining some of the speech section candidates (with the highest peak level) or a plurality of speech section candidates as speech sections, the speech section can be determined based on the peak height, width, and time difference with adjacent peaks. Even if noise with an instantaneous peak is close to the voice, it can be removed, and even in voice with a deep power dip, the peak at the beginning of a word can be prevented from being lost.
The drawbacks of the conventional method mentioned above can be removed,
When used in a speech recognition device, the recognition rate can be improved.

[Brief explanation of the drawing]

FIG. 1 is a block diagram of one embodiment of the present invention, and FIG.
The figure is a diagram showing the waveform of smoothed audio power in this embodiment. 1... Power calculation section, 2... Power smoothing section,
3...Peak detection section, 4...Threshold value calculation section, 5...
Peak selection unit, 6...Peak width calculation unit, 7...Peak width comparison unit, 8...Speech section determination unit, 10...
Input audio, 11... Power, 12... Smoothed power, 13... Peak candidate, 14... Threshold for peak selection, 15... Threshold for peak width calculation, 16
...Peak width, 17...Voice section candidate, 18...
Voice section, 19...Coefficient for peak selection, 20...
Peak width calculation coefficient, 21...Peak width threshold, 2
2...Voice segment candidate time difference calculation unit, 23...Voice segment candidate time difference, 24...Voice segment candidate time difference threshold, 30...Horizontal axis (time), 31...Vertical axis (smoothed power)), 32 ~35...Peak candidate, 3
8 to 40...Peak width, 41...Terminal, 42...
End, 43, 44 voice section candidate time difference.

Claims (1)

[Claims] 1. A power calculation unit that calculates the power of an audio signal; a power smoothing unit that smoothes the power calculated by the power calculation unit to obtain smoothed power;
a peak detection unit that detects an inflection point where the rate of change of the smoothed power changes from positive to negative as a peak candidate of the smoothed power; and a peak detector that selects a peak candidate with a maximum level among the peak candidates as the maximum peak. , a threshold calculation unit that calculates a peak selection threshold from the maximum peak level and a predetermined peak selection coefficient, and calculates a peak width calculation threshold from the maximum peak level and a predetermined peak width calculation coefficient; , a peak selection unit that compares the level of the peak candidate detected by the peak detection unit with the peak selection threshold and selects only the peak candidates whose level is equal to or higher than the peak selection threshold as peaks; a peak width calculation unit that calculates, as a peak width, a time during which the peak width calculation power is greater than or equal to the peak width calculation threshold and that includes the peak selected by the peak selection unit; and a predetermined width threshold among the peak widths. a peak width comparison section that outputs a wider peak width as a speech section candidate; and when there is a plurality of speech section candidates obtained by this peak width comparison section, the previous speech section candidate among the adjacent speech section candidates; a speech segment candidate time difference calculation unit that calculates the time from the end of the speech segment candidate to the start of the subsequent speech segment candidate as a speech segment candidate time difference; and a speech segment determination unit that performs the determination, and the speech segment determination unit directly determines the speech segment candidate as the speech segment when there is one speech segment candidate, and when there is a plurality of speech segment candidates, the speech segment determination unit directly determines the speech segment candidate as the speech segment. If the speech segment candidate time difference between adjacent speech segment candidates is shorter than a predetermined speech segment candidate time difference threshold, the plural speech segment candidates are combined into one speech segment candidate, and the speech segment candidates are combined into one speech segment candidate. A process of summarizing speech segment candidates is performed in which a new speech segment candidate is created from the beginning to the end of the voice segment candidate after the speech segment candidate, and this process of summarizing speech segment candidates is repeatedly performed to finally select one of the remaining speech segment candidates. A voice detection method characterized in that one or more of the voice sections are defined as the voice section. 2. In the speech detection method according to claim 1, when there are finally a plurality of speech section candidates, the speech section determining section selects the speech having the highest peak level among the speech section candidates. A speech section detection method characterized in that only section candidates are used as the speech section. 3. In the speech detection method according to claim 1, when there are finally a plurality of speech segment candidates, the speech segment determination unit determines each speech segment candidate to be a different speech segment. A voice section detection method characterized by: