TW201633291A - Speech segment detection device and method for detecting speech segment - Google Patents

Abstract

A speech segment detection device comprises: a speech segment detection unit (4) for detecting a provisional start time indicating the start point and a provisional end time indicating the end point of the speech segment included in an input signal using a pattern recognition model for distinguishing speech from noise included in an input signal on the basis of a first feature calculated by a first feature calculation unit (1); and a start-end time correction unit (5) for correcting the start time and end time on the basis of a comparison between a second feature calculated by a second feature calculation unit (2) and a threshold value.

Description

Sound interval detecting device and sound interval detecting method

The present invention relates to a technique for detecting a sound interval in an input signal using a complex feature quantity.

The sound section detection processing of the section in which the sound is extracted from the input signal is a very important process as the pre-processing of the voice recognition. In general, the voice recognition processing is performed on the section detected by the voice section detection processing. Since the pattern recognition is performed to obtain the recognition result and the sound section is detected incorrectly, the recognition accuracy of the voice recognition processing is greatly reduced. As a basic method of the sound section detection, there is a method of calculating the power of the input signal and detecting a section in which the calculated power is equal to or greater than the set threshold value as the sound section. The above detection method is preferable in the case where the background noise is small and the stable sound section is detected.

On the other hand, as a result of the inspection result input and the various FA (factory automation) equipment and the like in the conservative operation of the factory equipment, the user is free from the user interface that is very effective in sound recognition. However, the conservative working environment of the plant equipment and the operating environment of the FA machine are environments in which unsteady noise such as knocking sounds and hammering sounds often occur. Therefore, only the method of detecting the sound interval based on the power calculated based on the above input signal is used, because the undetected noise is erroneously detected as the sound, and the detection accuracy of the sound interval is lowered, and the subsequent sound recognition processing cannot obtain sufficient recognition performance. problem.

For the above problem, for example, Patent Document 1 discloses a sound interval detecting method in which a feature amount used in sound area detection is converted into a power of an input signal, and a cepstrum representing a spectral characteristic of an input signal is used, and the above is used. CMM (Hidden Markov Model) with cepstrum as a parameter. Specifically, regarding the sound and noise, respectively, when learning several HMMs and detecting the beginning of the start point of the sound interval, the likelihood of each HMM is calculated, and the frame with the highest HMM likelihood calculated in 12 frames (120 milliseconds) is calculated. When it is 4 or more frames, the front frame of the above 12 frames is detected as the beginning of the sound section.

[advance technical documents] [Patent Document]

[Patent Document 1] Japanese Patent Publication No. 2001-343983

However, in the technique disclosed in the above Patent Document 1, since the sound interval detection is performed using the feature amount representing the spectral characteristics of the input signal, it is possible to suppress erroneous detection of noise different from the spectral characteristics of the sound as sound, but no murmur (p, t, Since k, s, sh, h, f) and the like are similar to the spectral characteristics, there is a problem that the above-mentioned silent sub-tones and the like are not correctly recognized.

The present invention has been made to solve the above problems, and it is intended to suppress the erroneous detection of unsteady noise as a sound, and to improve the accuracy of detecting the unvoiced consonants of the head and the ending of the voice.

Sound interval detecting device according to the present invention, including the first feature The calculation unit calculates a first feature amount for displaying a spectral feature based on the input signal, and the second feature amount calculation unit calculates a second feature amount for displaying a sound feature amount different from the first feature amount based on the input signal, and a sound interval detecting unit. Using the identification model for recognizing the sound and the noise included in the input signal, the first feature amount calculated by the first feature amount calculation unit detects the start time and the instruction end point of the start point of the sound section included in the instruction input signal. And the terminal end correction unit corrects the start time and the terminal time detected by the sound section detection unit based on the comparison between the second feature amount calculated by the second feature amount calculation unit and the threshold value.

According to the present invention, it is possible to suppress the erroneous detection of the unstable noise as the sound interval, and it is also possible to improve the detection accuracy of the silent voice of the head and the end of the voice.

1‧‧‧1st feature quantity calculation unit

2‧‧‧2nd feature quantity calculation unit

3‧‧‧ Pattern Identification Model Accumulation Department

4‧‧‧Sound interval detection department

5‧‧‧ Always correcting department

5a‧‧‧ Always correcting department

5b‧‧‧ Always correcting department

6‧‧‧Threshold value calculation unit

10‧‧‧Sound interval detection device

10a‧‧‧Sound interval detection device

10b‧‧‧Sound interval detection device

20‧‧‧ processor

30‧‧‧ memory

[Fig. 1] is a block diagram showing a configuration of a sound section detecting device of the first embodiment; [Fig. 2] is a hardware configuration diagram showing a sound section detecting device of the first embodiment; [Fig. 3A] is a display The operation flowchart of the sound section detecting device of the first embodiment; [Fig. 3B] is a flowchart showing the operation of the sound section detecting apparatus of the first embodiment; [Fig. 4] shows the sound section according to the first embodiment. a search interval map generated by the always-end correction portion of the detecting device; [Fig. 5] is a block diagram showing a configuration of a sound section detecting device of the second embodiment; [Fig. 6] showing a search section generated by the constant end correcting section of the sound section detecting apparatus according to the second embodiment, according to a critical The threshold value calculation interval map generated by the value calculation unit; [FIG. 7A] shows an operation flowchart of the sound interval detection device of the second embodiment; [FIG. 7B] shows the operation of the sound interval detection device of the second embodiment. [Fig. 8] is a block diagram showing the configuration of the sound section detecting device of the third embodiment; [Fig. 9A] is a flowchart showing the operation of the sound section detecting apparatus of the third embodiment; and [Fig. 9B] The flowchart showing the operation of the sound section detecting device of the third embodiment is shown.

Hereinafter, in order to explain the present invention in more detail, the embodiments for carrying out the invention will be described with reference to the accompanying drawings.

[First Embodiment]

Fig. 1 is a block diagram showing the configuration of the sound section detecting device 10 of the first embodiment.

The sound section detecting device 10 is composed of a first feature amount calculating unit 1, a second feature amount calculating unit 2, a pattern identifying model storing unit 3, a sound section detecting unit 4, and a permanent end correcting unit 5.

The first feature amount calculation unit 1 performs acoustic analysis of an input signal that is externally input, and expresses a time series of a feature amount of the spectral feature (hereinafter referred to as a first feature amount). The first feature amount is, for example, data of 1 to 12 dimensions of MFCC (Meier Cepstral Coefficient). Further, for simplification of the description below, the data of the MFCC from 1 to 12 dimensions is simply referred to as MFCC.

The second feature amount calculation unit 2 is configured to calculate the feature amount different from the first feature amount converted by the first feature amount calculation unit 1 and is suitable for detecting the feature amount of the sound that is difficult to distinguish among the first feature amount (hereinafter referred to as The time series of the second feature quantity). For example, a time series suitable for detecting a feature amount such as a silent consonant of a sound that is difficult to distinguish between noise in the first special trace is calculated. Here, the silent sub-tones p, t, k, s, sh, h, f, and the like. In general, the unvoiced consonant is calculated such that the high-frequency power is emphasized as the second feature amount because the sound is concentrated at a high frequency.

The pattern recognition model accumulation unit 3 accumulates a pattern recognition model for discriminating sound and noise in the input signal. In the first embodiment, a case of using a GMM (Gaussian Mixture Model) will be described as an example. Specifically, a GMM (hereinafter referred to as a sound GMM) that models sound, and one GMM (hereinafter referred to as a noise GMM) that models noise are configured to form a pattern recognition model. The parameters of the sound GMM and the noise GMM are first obtained by, for example, learning using the most likelihood estimation method. The parameter learning of the sound GMM is performed using the MFCC of various sounds, and the parameters of the noise GMM are learned using the MFCC of various noises.

The voice section detecting unit 4 refers to the parameter identification model accumulated in the pattern recognition model accumulation unit 3, performs pattern matching of the first feature amount calculated by the first feature quantity calculation unit 1, and detects the start point of the sound section in the instruction input signal. The tentative start time (hereinafter referred to as the temporary start time) and the tentative end The terminal time of the point (hereinafter referred to as the temporary terminal time). The always-end correction unit 5 corrects the temporary start time and the temporary terminal time detected by the sound section detecting unit 4 based on the second feature amount, and determines the start time and the terminal time. The always-end correction unit 5 outputs the obtained start time and terminal time as time information of the sound interval in the input signal.

Fig. 2 is a view showing a hardware configuration of the sound section detecting device 10 of the first embodiment.

The processor 20 executes the program stored in the memory 30 to realize the first feature amount calculation unit 1, the second feature quantity calculation unit 2, the sound section detection unit 4, and the permanent correction unit 5 of the voice section detecting device 10. The pattern recognition model accumulation unit 3 constitutes the memory 30. Further, the processor 20 and the plurality of memories 30 configured in plural may perform the above functions in combination.

Next, the operation of the sound section detecting device 10 will be described.

3A and 3B are flowcharts showing the operation of the sound section detecting device 10 of the first embodiment.

When the signal is input (step ST1), the first feature amount calculation unit 1 divides the time interval (hereinafter referred to as a frame) for setting the input signal, and converts the input signal for each divided frame to calculate the first feature amount (step ST2). ). Further, in the division of the frame, the time interval between adjacent frames may be repeated. For example, the time interval of the frame is 30 milliseconds long, and the frame is shifted every 10 milliseconds, and the input signal is converted to calculate the first feature amount. The first feature amount is MFCC as described above. In other words, in the process of step ST2, the first feature amount calculation unit 1 calculates the time sequence of the MFCC at intervals of 10 msec and outputs it.

The second feature quantity calculation unit 2 is in phase with the first feature quantity calculation unit 1 At the same frame interval, the input signal is divided, and the input signal is converted for each divided frame, and the second feature amount is calculated (step ST3). In addition, in step ST3, the power of the high-frequency emphasis is calculated as the second feature quantity, and the following description will be made. The second feature amount calculation unit 2 regards the first K-frame (for example, K=10) of the input signal as a noise section in which the sound does not exist, and calculates the power average of the sound in the section of the K-frame as the noise level (step ST4). ). In addition, the second feature amount calculation unit 2 calculates the high frequency emphasis difference power by subtracting the noise level calculated in step ST4 from the power of the emphasized high frequency calculated in step ST3 in each frame (step ST5). In the process of step ST5, the second feature quantity calculation unit 2 calculates a time series of the high frequency emphasis difference power at intervals of 10 msec, and outputs the time series.

The voice interval detecting unit 4 inputs the time-series sequence of the MFCC, which is the first feature amount calculated in step ST2, and refers to the pattern recognition model accumulated in the pattern recognition model accumulation unit 3, and calculates the likelihood Ls and the noise GMM of each frame sound GMM. Log likelihood Ln (step ST6). The sound interval detecting unit 4 calculates the log likelihood difference S based on the likelihood Ls of the sound GMM and the log likelihood Ln of the noise GMM calculated in step ST6 based on the following equation (1) (step ST7).

S=Ls-Ln (1)

The sound section detecting unit 4 is a section in which the log likelihood difference S calculated in the forward direction search step ST7 of the time axis is equal to or greater than the threshold value Th_T1 of the set number of frames (step ST8). In the section searched for step ST8, the sound section detecting unit 4 acquires the time frame in which the log likelihood difference S first becomes the threshold value Th_S in the forward direction of the time axis, and serves as the temporary start time Tb' of the voice section (step ST9).

Next, the sound interval detecting unit 4 searches for the direction of the time axis The frame in which the log likelihood difference S calculated in step ST7 does not reach the set threshold value Th_S is a continuous section having a threshold value Th_T2 or more of the set number of frames (step ST10). In the section searched for step ST10, the sound section detecting unit 4 acquires a time frame in which the log likelihood difference S does not reach the threshold value Th_S in the forward direction of the time axis as the temporary terminal time Te' of the voice section (step ST11). . Further, the above-described search processing of step ST8 and step ST10 continues until the target frame is searched.

The time-series of the high-frequency emphasis difference power calculated in step ST5 with reference to step ST5, from the time Tb1 of the frame b1 in front of the time series of the temporal start time Tb' detected in step ST9, to the sound located in the sound In the section of the time Tb2 of the frame b2 after the time series of the temporary start time Tb', the frame in which the high-frequency emphasis difference power is the threshold value Th_P1 or more in the forward direction of the time axis is a continuous interval of the set number of frames Th_T1 or more. (Step ST12). The always-end correction unit 5 determines whether or not the section has been searched by the processing of step ST12 (step ST13). When the section has been searched (step ST13; YES), the constant-end correction unit 5 acquires the time at which the high-frequency emphasized difference power first becomes the threshold value Th_P1 or more in the forward direction of the time axis in the searched section. The start time Tb (step ST14). On the other hand, when the section is not searched (step ST13; NO). The end correction unit 5 uses the temporary start time Tb' detected in step ST9 as the start time Tb (step ST15).

Next, the time series of the high-frequency emphasis difference power calculated in step ST5 with reference to step ST5 is located from the time Te2 of the frame e2 behind the time series of the temporary terminal time Te' of the sound section detected in step ST11. The temporary terminal time of the sound, the time of the frame e1 in front of the time series Te1 In the section, the frame in which the high-frequency-emphasized difference power is the threshold value Th_P1 or more in the reverse direction of the time axis is a continuous section of the set threshold number Th_T1 or more (step ST16). The always-end correction unit 5 determines whether or not the section has been searched by the processing of step ST16 (step ST17). When the section has been searched (step ST17; YES), the constant-end correction unit 5 acquires the time when the high-frequency-emphasized difference power first becomes the threshold value Th_P1 or more in the reverse direction of the time axis in the searched section. The terminal time Te (step ST18). On the other hand, when the section is not searched (step ST17; NO). The terminal correction unit 5 uses the temporary terminal time Te' detected in step ST11 as the terminal time Te (step ST19).

The always-end correction unit 5 outputs the start time Tb obtained in step ST14 or step ST15 and the terminal time Te acquired in step ST18 or step ST19 as time information of the sound section (step ST20), and ends the process.

Further, the threshold value Th_S, the threshold value Th_P1, the threshold value Th_T1, and the threshold value Th_T2 are constants of 0 or more set in advance.

Fig. 4 is a view showing a search interval generated by the always-end correction unit 5 of the sound section detecting device 10 according to the first embodiment.

In Fig. 4, the horizontal axis indicates time, and the vertical axis indicates the intensity of the log likelihood difference S between the sound GMM and the noise GMM. In Fig. 4, the time Tb' is the temporary start time Tb' calculated in step ST9. The time Te' is the temporary terminal time Te' calculated in step ST11. The section A displays a section from the time Tb1 of the frame b1 in front of the time series of the temporary start time Tb' to the time Tb2 of the frame b2 located at the rear, and instructs the always-end correction unit 5 to perform a search interval for the search for the start time correction. . The arrow B indicates the search direction when the always-end correction unit 5 searches for the section A, and displays the forward direction search to the time axis.

Further, the section C displays the section from the time Te2 of the frame e2 located behind the time series of the temporary terminal time Te' to the time Te1 of the frame e1 located ahead, and instructs the always-end correction section 5 to perform the search section for the terminal time correction search. . The arrow D indicates the search direction when the always-end correction unit 5 searches for the section C, and displays the reverse direction search for the time axis.

When the specific example is displayed, for example, from the temporary start time Tb'25 to the frame front setting time Tb1, from the temporary start time Tb'10 to the frame rear setting time Tb2, and from the temporary terminal time Te'10 to the frame front setting time Te1, from the temporary starting time Set Te2 at the rear of the Te'30 frame. Further, the frame setting time Tb2 may be set from the temporary start time Tb'0, and the frame setting time Te1 may be set from the temporary terminal time Te'0 to avoid correction in front of the sound section detected by the first feature amount.

As described above, according to the first embodiment, the first feature amount calculation unit 1 is configured to calculate the first feature amount of the input signal, and the second feature amount calculation unit 2 calculates the first feature suitable for detection based on the input signal. The second feature amount of the sound that is difficult to distinguish from the noise; the sound segment detecting unit 4 uses the pattern recognition method for the first feature amount, determines the sound and the noise, calculates the temporary start time and the temporary terminal time, and the permanent end correction unit 5, By using the second feature amount, the temporary start time and the temporary terminal time are corrected, and the time information of the sound section is obtained. According to the processing of the sound section detecting unit 4, the unsteady noise having the difference in the detected spectral feature amount is suppressed as the sound section, and the correction is performed based on the end point. The processing of the section 5 can suppress the detection of a sound that is difficult to distinguish with the noise by the spectral feature amount, and improve the detection accuracy of the sound section.

Further, according to the first embodiment, since the second feature amount is calculated The output unit 2 calculates a high-frequency emphasis difference power suitable for detecting a silent sub-tone that is difficult to discriminate based on the spectral feature amount as the second feature amount, and the constant-side correction unit 5 corrects the temporary start time and the time-series using the high-frequency emphasized difference power. At the temporary terminal time, the time information of the sound section is obtained, and it is possible to suppress the omission of the unvoiced consonant and improve the detection accuracy of the sound section.

Further, in the first embodiment, the parameter learning of the sound GMM and the noise GMM of the pattern recognition model accumulated in the pattern recognition model accumulation unit 3 is exemplified as the case of using the most likelihood estimation method, but the positive discrimination sound may be applied. Learning with parameters of noise, such as mutual information maximization.

Further, in the first embodiment described above, a configuration is adopted in which one sound GMM and noise GMM are used as the GMM constituting the pattern recognition model accumulated in the pattern recognition model accumulation unit 3, and a plurality of GMMs may be used. In this case, the log likelihood of the sound GMM may be a maximum value or a weighted average value of the log likelihood of the plurality of sound GMMs. Similarly, the log likelihood of the noise GMM is assumed to be the maximum or weighted average of the log likelihood of a plurality of noise GMMs.

Further, in the first embodiment described above, the GMM is used as the pattern recognition model accumulated by the pattern recognition model accumulation unit 3. However, the HMM may be used. Further, pattern recognition methods such as a logistic regression model, a support vector machine, and a neural network may be used.

Further, in the first embodiment, the second feature amount calculation unit 2 calculates the high-frequency emphasis difference power as a feature amount suitable for detecting the unvoiced consonant, but is suitable for detecting the feature amount of the unvoiced consonant, that is, the silent state. If there is a characteristic feature quantity in the consonant, an arbitrary feature quantity can be applied. For example, calculate the output per band The power of the incoming signal, the power of the frequency band less than 2 KHz (kilohertz) and the power of 2 KHz or more, the ratio of the two powers can be applied as the feature quantity.

[Second embodiment]

In the first embodiment described above, the constant-end correction unit 5 displays a configuration using the preset threshold value Th_P1 when the high-frequency emphasis difference power and the threshold value are compared. However, in the second embodiment, the display uses the high-frequency emphasis differential power. The standard deviation is calculated as a threshold value of the comparison target of the high frequency emphasis difference power.

Fig. 5 is a block diagram showing the configuration of the sound section detecting device 10a of the second embodiment.

The sound section detecting device 10a of the second embodiment additionally adds the threshold value calculating unit 6 to the sound section detecting device 10 shown in the first embodiment.

Fig. 6 is a view showing a search section generated by the constant end correction unit 5a of the sound section detecting device 10a according to the second embodiment and a threshold value calculation section generated based on the threshold value calculation unit 6.

In the following, the same or equivalent components as those of the voice section detecting device 10 of the first embodiment are denoted by the same reference numerals as those used in the first embodiment, and the description thereof will be omitted or simplified.

The threshold value calculation unit 6 calculates the second-order feature quantity calculated by the second feature quantity calculation unit 2, that is, the time-series of the high-frequency-emphasized difference power and the temporary start time Tb' detected by the sound section detection unit 4, and calculates the constant-end correction unit. The critical value of 5a reference. When the description is made with reference to Fig. 6, the threshold value calculation unit 6 starts the time Tb0 at the time Tb1 of the frame b1 in front of the time series in the time series of the temporary start time Tb', and the interval E from the time Tb0 to the time Tb1. In the middle, the standard deviation sd of the high frequency emphasis difference power is calculated according to the following formula (2).

In the equation (2), the average value of the high-frequency difference power of the section E of the mp-based time Tb0 to the time Tb1, the high-frequency difference power in the time i of the p _i system, and the sqrt() system show a function of taking the square root. Further, the frame number Tv is a predetermined constant, and is, for example, 50 frames.

The threshold value calculation unit 6 calculates the constant value correction threshold value Th_P2 based on the following equation (3), using the standard deviation sd of the high frequency emphasis difference power calculated by the equation (2).

Th_P2=α*sd+β (3)

In the formula (3), α and β are predetermined constants of 0 or more. The constant end correction threshold value Th_P2 calculated by the threshold value calculation unit 6 is output to the constant end correction unit 5a.

Next, the operation of the sound section detecting device 10a will be described.

7A and 7B are flowcharts showing the operation of the sound section detecting apparatus 10a of the second embodiment.

In the following, the same or corresponding steps as those of the voice section detecting device 10 of the first embodiment are denoted by the same reference numerals as those used in the third and third embodiments, and the description thereof will be omitted or simplified.

When the sound section detecting unit 4 detects the temporary terminal time Te' of the sound in the step ST11, the threshold value calculating unit 6 calculates the time of the frame b1 in front of the time series from the temporary start time Tb' of the sound detected in step ST9. The time at which the Tb1 starts to trace back the frame number Tv is the time Tb0 (step ST31). The threshold value calculation unit 6 starts the region to the time Tb1 at the time Tb0 calculated in step ST31. The standard deviation sd of the high frequency emphasis difference power is calculated based on the above equation (2) (step ST32). In addition, the threshold value calculation unit 6 calculates the constant value correction threshold value Th_P2 based on the above equation (3), using the standard deviation sd of the high frequency emphasis difference power calculated in step ST32 (step ST33).

The all-time correction unit 5a refers to the time series of the high-frequency emphasis difference power calculated in step ST5, and starts from the time Tb1 of the frame b1 in front of the time series of the temporary start time Tb' of the sound detected in step ST9 to the temporary position of the sound. In the section from the time Tb2 of the frame b2 after the time series of the start time Tb', the frame in which the high-frequency emphasis difference power is searched for in the forward direction of the time axis is the frame of the set value of the constant end correction threshold Th_P2 calculated in step ST33. The critical value of the number Th_T1 is equal to or longer than the continuous interval (step ST34).

The always-end correction unit 5a determines whether or not the search has been performed in the process of step ST34 (step ST35). When the section is searched for (step ST35; YES), the terminal correction unit 5a obtains the time when the high-frequency emphasis difference power first becomes the frame of the constant-end correction threshold value Th_P2 or more as the start time Tb in the searched section (step ST36). ). On the other hand, when the section is not searched (step ST35; NO), the temporary end correction unit 5A sets the temporary start time Tb' detected in step ST39 as the start time Tb (step ST15).

Next, the all-in-time correction unit 5a refers to the time series of the high-frequency emphasis difference power calculated in step ST5, and starts at the time Te2 of the frame e2 located behind the time series of the temporary terminal time Te' of the sound detected in step ST11. In the range from the time Te1 of the frame e1 in front of the time series of the temporary terminal time Te' of the sound, the frame in which the high-frequency emphasis difference power is the upper limit correction threshold Th_P2 or more is set in the frame in the reverse direction of the time axis. Criticality The value is a continuous interval of Th_T1 or more (step ST37). The always-end correction unit 5a determines whether or not the section has been detected by the processing of step ST37 (step ST38). When the section has been searched (step ST38; YES), the terminal correction unit 5a obtains the time when the high-frequency emphasized difference power first becomes the frame of the constant-end correction threshold Th_P2 or more as the terminal time. Te (step ST39). On the other hand, when the section is not searched (step ST38; NO). The terminal correction unit 5a uses the terminal time Te' detected in step ST11 as the terminal time Te (step ST19).

The always-end correction unit 5 outputs the start time Tb obtained in step ST36 or step ST15 and the terminal time Te acquired in step ST39 or step ST19 as time information of the sound section (step ST20), and ends the process.

As described above, according to the second embodiment, the configuration including the threshold value calculation unit 6 is used as the time Tb0 at the time Tb0 at the time Tb1 of the frame b1 located in front of the time series of the temporary start time Tb' as the time Tb0. The standard deviation sd of the high-frequency-emphasized difference power calculated in the interval from the time Tb0 to the time Tb1, the constant-end correction threshold value Th_P2 is calculated, and the constant-end correction unit 5a is based on the calculated constant-end correction threshold value Th_P2 and the high-frequency emphasis The time series of the difference power, the temporary start time and the temporary terminal time are corrected, and the time information of the sound interval is obtained; the standard deviation value of the high frequency emphasis difference power is small, and for a stable noise environment, the threshold value for the low terminal correction can be set, and the threshold value can be set. The detection performance of the weak silent sub sound is improved. On the other hand, the high-frequency emphasis standard power of the differential power is large, and in the unstable noise environment, the threshold value for the high-end correction can be set, and the erroneous detection of the noise can be suppressed.

[Third embodiment]

In the third embodiment, the time series of the high-frequency emphasis difference power calculated by the second feature quantity calculation unit 2 is also considered, and the time series of the log likelihood difference S detected by the sound section detection unit 4 is also considered. Always at the moment.

Fig. 8 is a block diagram showing the configuration of the sound section detecting device 10b of the third embodiment.

The sound section detecting device 10b of the third embodiment has the same configuration as that of the sound section detecting device 10a shown in the second embodiment. In the following, the same or corresponding components as those of the sound section detecting device 10a of the second embodiment are denoted by the same reference numerals as those used in the second embodiment, and the description is omitted or simplified.

Similarly to the first embodiment and the second embodiment, the voice interval detecting unit 4 outputs the temporary start time Tb' and the temporary terminal time Te' to the constant end correcting unit 5b. Further, the sound section detecting unit 4 outputs the logarithmic likelihood S of the sound GMM and the noise GMM calculated from the above equation (1), that is, the time series of the log likelihood difference S to the constant end correcting unit 5b. In the same manner as in the second embodiment, the threshold value calculation unit 6 calculates the end-time correction based on the time series of the high-frequency emphasis difference power input by the second feature quantity calculation unit 2 and the temporary start time Tb′ detected by the sound section detection unit 4 . The threshold value for the threshold value referenced by the portion 5b is corrected by the threshold value Th_P2.

The time-series sequence of the high-frequency emphasis difference power input by the second feature quantity calculation unit 2, the time series of the log likelihood difference S input by the voice section detection unit 4, and the terminal end of the threshold value calculation unit 6 are input from the terminal correction unit 5b. The correction threshold value Th_P2 is used to correct the temporary start time Tb' and the temporary terminal time Te' detected by the sound section detecting unit 4, and acquire the start time Tb and the terminal time Te.

Next, the operation of the sound section detecting device 10b will be described.

9A and 9B are diagrams showing the sound section detecting device 10b of the third embodiment. Action flow chart.

In the following, the same steps as those of the sound interval detecting device 10a of the second embodiment are denoted by the same reference numerals as those used in the seventh and seventh embodiments, and the description thereof will be omitted or simplified.

In the step ST33, when the threshold value calculation unit 6 calculates the constant value correction threshold value Th_P2, the terminal correction unit 5b refers to the time series of the high frequency emphasis difference power calculated in step ST5 and the log likelihood difference S calculated in step ST7. The sequence is from the time Tb1 of the frame b1 in front of the time series of the temporary start time Tb' of the sound section detected in step ST9 to the time Tb2 of the frame b2 located behind the time series of the temporary start time Tb' of the sound, The high-frequency emphasis difference power in the forward direction of the time axis is equal to or greater than the constant-end correction threshold value Th_P2 calculated in step 33, and the log likelihood difference S is equal to or greater than the critical value Th_T1 of the set number of frames. The interval (step ST41).

Here, the threshold value Th_S2 is a constant of 0 or more which is set in advance, and is smaller than the threshold value Th_S.

The always-end correction unit 5b determines whether or not the section has been searched by the processing of step ST41 (step ST42). When the section has been searched (step ST42; YES), the constant-end correction unit 5b obtains the high-frequency emphasis difference power in the searched section, and initially becomes the constant-end correction threshold value Th_P2 or more, and the log-likelihood difference S is The time of the frame of the threshold value Th_S2 or more is referred to as the start time Tb (step ST43). On the other hand, when the section is not searched (step ST42; NO). The all-end correction unit 5b uses the temporary start time Tb' detected in step ST9 as the start time Tb (step ST15).

Next, the all-end correction unit 5b refers to the time series of the high-frequency emphasis difference power calculated in step ST5 and the time series of the log likelihood difference S calculated in step 7, from the temporary terminal time Te' of the sound detected in step ST11. At the time Te2 of the frame e2 at the rear of the time series, to the time zone Te1 of the frame e1 in front of the time series of the temporary terminal time Te' of the sound, the high-frequency emphasis difference power is searched in the reverse direction of the time axis as the critical value for the terminal correction The frame having a value of Th_P2 or more and a logarithm likelihood difference S of the set threshold value Th_S2 or more is a continuous section of the set threshold number Th_T1 or more (step ST44).

The end correction unit 5b performs a determination as to whether or not the processing in the step ST44 has been detected (step ST45). When the section has been searched (step ST45; YES), the constant-end correction unit 5b obtains the high-frequency emphasis difference power in the searched section, and initially becomes the constant-end correction threshold value Th_P2 or more, and the log-likelihood difference S is The timing of the frame having the threshold value Th_S2 or more is referred to as the terminal time Te (step ST46). On the other hand, when the section is not searched (step ST45; NO). The terminal correction unit 5b uses the temporary terminal time Te' detected in step ST11 as the terminal time Te (step ST19).

The always-end correction unit 5b outputs the start time Tb obtained in step ST43 or step ST15 and the terminal time Te acquired in step ST46 or step ST19 as time information of the sound section (step ST20), and ends the process.

As described above, by setting the threshold value Th_S2 to a value smaller than the threshold value Th_S, it is easy to detect a weak unvoiced consonant or the like which cannot be detected at the time of detection of the temporary start time Tb' and the temporary terminal time Te'. Further, the time series of the high-frequency-emphasized difference power is not used, and only the time series of the log likelihood difference S is used, and the threshold value Th_S2 is set to a value smaller than the threshold value Th_S, and the noise can be detected when performing the search processing. The sex becomes large, but only when the high frequency emphasizes the time series of the difference power and the time series of the log likelihood difference S, when the feature quantities of both are above the critical value, by correcting the temporary start time Tb' and the temporary terminal time Te ', can improve the correction accuracy.

The always-end correction unit 5 b corrects the constant-end time in addition to the high-frequency emphasis difference power plus the log likelihood difference, thereby making it possible to easily detect a weak silent sub-tone that cannot be detected at the temporary start time detection. Wait. However, when the constant end time is corrected using only the log likelihood difference setting low threshold value, the possibility of erroneously detecting noise as a sound becomes high. Therefore, when the log likelihood difference and other feature quantities are used at the same time, only when the feature quantities of both are above the critical value, the configuration of the correction end time is made, and the correction accuracy is improved.

As described above, according to the third embodiment, the time-series of the high-frequency emphasis difference power calculated by the second feature amount calculation unit 2 and the log likelihood detected by the sound section detection unit 4 are included in the third-end feature calculation unit 2. The difference time sequence and the threshold value for the constant end correction input from the threshold value calculation unit 6 can correct the temporary start time and the temporary terminal time detected by the sound interval detecting unit 4, thereby suppressing the erroneous detection of noise and correcting the sound, and improving The correction accuracy of the start point of the sound and the end point of the sound.

Further, according to the third embodiment, since the set threshold value Th_S2 is set to a value smaller than the threshold value Th_S, it is possible to easily detect the weak silence that cannot be detected at the time of detecting the temporary start time Tb' and the temporary terminal time Te'. Consonant and so on.

Further, in the third embodiment, the configuration of the application end correction unit 5b in the sound section detecting device 10a shown in the second embodiment is displayed. However, the application of the constant end correction unit 5b may be employed as shown in the first embodiment. Sound interval Detection device 10.

In the first to third embodiments described above, the sound having difficulty in identifying the noise is detected by the first feature amount, and the detection of the unvoiced consonant is taken as an example. However, the unvoiced consonant may be configured to perform the unvoiced vowel. Check out. In addition, when the vocal sound is detected or the vowel is detected, and the vocalization is unclear, when the first feature amount is difficult to distinguish from the noise, the sound can be detected and predicted.

Further, the present invention may be a free combination of the embodiments or a modification of any constituent elements of the respective embodiments within the scope of the invention, or any constituent elements may be omitted in the respective embodiments.

[Industry use possibility]

The sound section detecting apparatus according to the present invention can be applied to a device that requires sound interval detection, such as a voice recognition device, to prevent erroneous detection of unsteady noise as sound, and to improve the detection accuracy of silent heads at the head and the end.

1‧‧‧1st feature quantity calculation unit

2‧‧‧2nd feature quantity calculation unit

3‧‧‧ Pattern Identification Model Accumulation Department

4‧‧‧Sound interval detection department

5‧‧‧ Always correcting department

10‧‧‧Sound interval detection device

Claims (5)

A sound segment detecting device includes: a first feature amount calculating unit that calculates a first feature amount for displaying a spectral feature based on an input signal; and a second feature amount calculating unit that calculates a sound different from the first feature amount based on the input signal The second feature amount of the feature amount; the voice segment detecting unit detects the input based on the first feature amount calculated by the first feature amount calculating unit, using an identification model for identifying the sound and the noise included in the input signal. The start time of the start point of the sound section included in the signal and the terminal time of the instruction end point; and the constant end correction unit corrects the sound section based on the comparison between the second feature quantity calculated by the second feature quantity calculation unit and the threshold value The start time and the terminal time detected by the detection unit. The sound section detecting device according to the first aspect of the invention, comprising: a threshold value calculating unit that calculates a criterion of the second feature amount from a start time of the sound interval detecting unit that is traced back to a predetermined time period The deviation is calculated based on the standard deviation of the second feature amount described above. The sound section detecting device according to the first aspect of the invention, wherein the sound section detecting unit calculates a likelihood between a sound model that models the sound and a noise model that models the noise, by referring to the identification model; And the above-described constant-end correction unit, in addition to the comparison of the second feature amount and the threshold value, corrects the detected by the sound section detecting unit based on the comparison of the likelihood difference calculated by the sound section detecting unit and the threshold value At the beginning Engraved and terminal moments. The sound section detecting device according to the first aspect of the invention, wherein the second feature amount calculating unit calculates the second feature amount of a feature of displaying a silent sound among the sounds included in the input signal. A sound section detecting method includes the following steps: a first feature amount calculating step that calculates a first feature amount for displaying a spectral feature based on an input signal, and a second feature amount calculating step that is based on a second feature amount calculating unit The input signal calculates a second feature amount for displaying a sound feature amount different from the first feature amount; and the detecting step, the sound segment detecting unit uses an identification model for recognizing sound and noise included in the input signal, according to the above a feature quantity, a detection start time indicating a start time of a start point of the sound section included in the input signal, and a terminal time of the instruction end point; and a correction step, the always-end correction unit correcting based on the comparison of the second feature quantity and the threshold value The start time and the terminal time.