JP2807457B2 - Voice section detection method - Google Patents

Description

Description: TECHNICAL FIELD The present invention relates to a voice section detection method, and more particularly to a voice section detection method in a voice response device, a voice recognition device and the like, and is applied to a toy, a musical instrument, an automatic accompaniment device, and the like. It is applicable. 2. Description of the Related Art There is a level detection method as a voice section detection method with a small hardware scale and a clear detection theory. this is,
The start and end of the voice section are determined by comparing a time-series feature amount obtained from the voice signal with a predetermined threshold value. However, in this method, it is difficult to accurately determine the start and end of the voice section, and there are problems such as malfunction due to background noise and detection of a weak signal. SUMMARY OF THE INVENTION The present invention has been made in view of the above situation, and has as its object to provide a voice detector that detects only a section of a voice signal from an input signal. Configuration In order to achieve the above object, the present invention forms one frame with a predetermined number of samples with respect to an input signal, sets the window length as a detection unit, and defines a feature amount within the window length and a predetermined threshold value. In the voice section detection method for detecting voice sections by comparison, a plurality of window lengths having different time lengths and threshold values corresponding to the respective window lengths are set, and the beginning of the voice is a middle time window length and a medium level. Using the threshold of, the number of frames exceeding the threshold for the start edge detection of the subsequent predetermined frames after exceeding the medium level threshold for the start edge detection repeatedly obtained until the number exceeds the predetermined number, the end of the voice, Using a long window length and a high-level threshold, until the number of frames below the threshold for the end detection exceeds a predetermined number in predetermined subsequent frames after falling below the high-level threshold for the end detection. Asked repeatedly, The start / end correction is performed using a short window length and a low-level threshold value, and comparing the feature amount calculated within the short window length with the threshold value, a time series from the previously detected start position is obtained. The point below the low-level threshold at the previous point in time is defined as a new starting point, and the point below the low-level threshold at a point in time later than the previously detected end point is determined as the new end point. It is what it was. Hereinafter, a description will be given based on examples of the present invention. As described above, the level detection method for comparing the energy (power, amplitude, and the like) of a voice signal with a preset threshold value and detecting the start and end of a voice section has a small hardware scale and a clear detection theory. On the other hand, the start and end of the voice section are not accurately determined, but there are problems such as malfunction due to background noise and detection of a weak signal. On the other hand, in the present invention, a plurality of window lengths having different time lengths are provided for the audio signal, and as a feature amount within the window length, the sum of absolute values of the difference values of adjacent audio signals is used, and the beginning of the audio is determined. Set a medium time window length to detect, set a long window length to determine the end to determine the temporary start and end using a different time length window and threshold, respectively, to determine the end, Further, the present invention is characterized in that a short window length and a threshold value are used to determine the exact start and end positions, and the accurate start and end of the voice section are detected. FIG. 1 is a configuration diagram for explaining an embodiment of a voice section detection system according to the present invention, wherein 1 is an A / D converter, 2
Is a difference value calculation unit, 3 is a feature amount calculation unit for each window length, 4 and 5 are comparison units, 6 and 7 are threshold units, A is an input voice signal, B is a start and end frame number output signal of a voice section. And the input signal A
After the A / D conversion, the difference value from the previous sample point is sequentially calculated. The difference value is converted to an absolute value, added for each predetermined number of samples, and used as a feature value for one frame. The setting of window lengths having different time lengths is performed based on the number of samples in one frame. In the first rain, for example, a window length of 20 ms is used for detecting the beginning of a voice section, and the feature amount within the window length is set to fsi. Similarly, for the end detection, a window length of 30 ms is used, and the feature amount within the window length is set to fei. i is a frame number. In addition, a window length of 10 ms is set so as to be more sensitive than the feature values fsi, fei, which are used when obtaining accurate start and end points, and the feature value is fsei. Features fsi, fei, fsei for finding the beginning and end of a voice section and their exact positions
Are TH, TL, and TE, respectively. FIG. 2 is a waveform diagram for explaining the operation principle of the present invention, FIG. 2 (a) is an input signal waveform, and FIG. 2 (b) is a characteristic amount fsi when the window length is 20 ms. Fig. 2 (c)
Shows the feature value fei when the window length is 30 ms, and FIG. 2D shows the feature value fsei when the window length is 10 ms. First, from the input point in time using the feature quantity of FIG.
To search for a frame that exceeds. Here, TH is set to be slightly higher so that noise and the like are not picked up. Further, in order to prevent sudden noise as shown in FIG. 2 (b) from being picked up, when the number of frames exceeding the threshold value for the detection of the starting end does not reach the predetermined number of the following predetermined frames after the detection of the starting end. In addition, a process of performing the start detection again from the next point of the start detection point and repeating the start detection until the number of subsequent frames exceeds a predetermined number is added. As a result, the position of the start end is determined as s ″ at the point where the threshold value is first exceeded, but becomes s ′ by the above-described operation. For the end detection, the frame after the point of s ′ is similarly shown in FIG. The feature quantity of (c) is used to search for a frame that falls below the next threshold value TL of the start detection point s', where the TL is exceeded within a predetermined number of subsequent frames after the end detection in order to prevent missing in the word. If it appears, update the end position and search for a point below the TL in the subsequent frame from the position exceeding the TL. In FIG. 2 (c), the point e ″ which is terminated first by this operation is updated to e ′, which is obtained, and the speech loss in the word is eliminated. When the start and end points are determined by the feature amounts of FIGS. 2B and 2C, a point lower than the threshold value TE is obtained at a time before the time s ′ using the feature amounts of FIG. 2D. Is the starting end s. Similarly, at a point after the provisional end point e ', the threshold TE
A lower point is obtained and this is set as a terminal e. Here, the plurality of window lengths and their feature amounts may be separately calculated for each time length. For example, only the feature amount of the short window length is calculated, and based on this, the middle time period is calculated. The sum of two adjacent features may be used as the window length, and the sum of three consecutive features may be used as the long-term feature. Effects As is clear from the above description, according to the present invention, it is possible to accurately and easily detect the start and end of a voice section.

BRIEF DESCRIPTION OF THE DRAWINGS FIG. 1 is a block diagram for explaining an embodiment of the present invention, and FIG. 2 is a waveform diagram for explaining the operation principle of the present invention. 1 ... A / D conversion unit, 2 ... Difference value calculation unit, 3 ... Feature amount calculation unit for each window length, 4,5 ... Comparison unit, 6,7 ... Threshold unit.

Continuation of the front page (58) Field surveyed (Int. Cl. ⁶ , DB name) G10L 3/00-9/18 JICST (JOIS)

Claims (1)

(57) [Claims] A voice section detection method that forms one frame with a predetermined number of samples for an input signal and uses this as a detection unit to determine a window length, and compares a feature amount within the window length with a predetermined threshold to detect a voice section. In, a plurality of window lengths with different time lengths and a threshold value corresponding to each window length are set, and the beginning of the voice uses a middle time window length and a middle level threshold, and the middle level for the leading edge detection is used. Of the subsequent predetermined frames after exceeding the threshold, the number of frames exceeding the threshold for the start detection is repeatedly determined until the number exceeds the predetermined number, and the end of the sound is determined using a long window length and a high-level threshold. Of the predetermined frames following the lower than the high-level threshold for the end detection, it is repeatedly determined until the number of frames below the threshold for the end detection exceeds a predetermined number, and the start and end corrections are performed.
By using a short window length and a low-level threshold value and comparing the feature amount calculated within the short window length with the threshold value, the low-level value is obtained at a time point earlier in time series than the previously detected start position. A voice section detection method characterized in that a point lower than the threshold value is determined as a new start point, and a point lower than the low-level threshold value is determined as a new end point in time series later than the previously detected end position.