JP6791816B2 - Voice section detection device, voice section detection method, and program - Google Patents

Description

The present invention relates to a voice section detection device, a voice section detection method, and a program.

With the spread of applications and devices that use voice, the importance of voice section detection technology that identifies the utterance time of voice buried in noise is increasing in a noisy environment. The technique of detecting a voice section is useful for recording by specifying a voice portion, voice enhancement for removing noise from voice, voice recognition, and the like.

Various techniques have been conventionally proposed as such voice section detection techniques. Patent Document 1 discloses a method of detecting a voice section by utilizing the fact that the power of voice is larger than the power of noise. Patent Document 2 discloses a method of using other physical information such as the direction of arrival of voice. Patent Document 3 discloses a technique for improving the performance of voice section detection by using a voice model learned from a large amount of voice in advance. Further, a configuration in which voice section detection is performed using the result of voice recognition is conceivable, but it is not realistic because the amount of calculation required for voice recognition is generally large and voice section detection must be performed at all times.

Japanese Unexamined Patent Publication No. 2005-321339 Japanese Unexamined Patent Publication No. 2012-048119 JP-A-2009-210647

The technology described in Patent Document 1 described above assumes that the power of voice is larger than the power of noise level, but the power of voice is not necessarily the power of voice as the use of voice in a real environment with high noise level increases. Is increasingly not expected to be greater than the power of noise. Further, the method described in Patent Document 2 requires a sensor for acquiring other physical information. Since the method disclosed in Patent Document 3 described above uses a voice model in which a large amount of voice is mixed, there is a problem that knowledge about voice is difficult to handle. In addition, it should be noted that due to the characteristics of voice interval detection, the required amount of calculation cannot be increased so much.

In view of the above background, it is an object of the present invention to propose a method for improving the detection performance of a voice section while suppressing an increase in the amount of calculation.

In the voice section detection device of the present invention, the voice feature amount is voice based on an input unit for inputting input voice including voice and noise, a feature amount calculation unit for obtaining voice feature amount from the input voice, and an acoustic model. An acoustic likelihood calculation unit that obtains the likelihood of any of the subsets, and a posterior probability calculation unit that calculates a posterior probability of voice or non-voice from the voice feature amount based on the voice interval detection model. It includes a posterior probability calculation unit that obtains posterior probability using the likelihood calculated by the likelihood calculation unit, and a determination unit that determines whether it is voice or non-voice based on the posterior probability. Here, the voice subset is a classification of phonemes such as "a", "i", and "u", or phonemes such as "vowel" and "consonant".

With this configuration, if the voice feature amount of the input voice is voice, the posterior probability of voice or non-voice is obtained in consideration of which subset it belongs to, so that the performance of voice section detection can be improved. Here, the acoustic model is a model for obtaining the likelihood that a sound characterized by a voice feature is a sound (for example, "a"). On the other hand, the voice section detection model is a model that learns the acoustic characteristics of voice and noise in order to classify them.

The voice section detection device of the present invention includes a dimension compression unit that compresses the dimension of the likelihood data calculated by the acoustic likelihood calculation unit, and the likelihood data whose dimension is compressed by the dimension compression unit. It may be input to the posterior probability calculation unit.

The dimension of voice likelihood data tends to be large, but by performing dimensional compression, the amount of calculation in the posterior probability calculation unit can be suppressed. As the dimensional compression method in the dimensional compression unit, for example, vector transformation, principal component analysis, neural network, or the like can be used.

The voice section detection device of the present invention may include a voice recognition unit that performs voice recognition using the likelihood data calculated by the acoustic likelihood calculation unit.

With this configuration, the data of the likelihood of the voice obtained for detecting the voice section can be used for the voice recognition, and the calculation cost of the voice recognition can be reduced.

The voice section detection device of another aspect of the present invention is based on an input unit for inputting input voice including voice and noise, a feature amount calculation unit for obtaining voice feature amount from the input voice, and a voice enhancement model. Based on the activation degree calculation unit that calculates the activation degree when the voice feature amount is separated into voice and noise, and the voice section detection model, the posterior probability of voice or non-voice is calculated from the voice feature amount. A post-probability calculation unit that calculates the post-probability, and a post-probability calculation unit that obtains the post-probability using the activation degree calculated by the activation-degree calculation unit, and a voice or non-voice based on the post-probability. It is provided with a determination unit for determining the above.

With this configuration, the posterior probability of voice or non-voice is obtained including the activation degree of voice and noise of the voice feature amount of the input voice, so that the performance of voice section detection can be improved.

In the voice section detection device of the present invention, the activation degree calculation unit factorizes the voice feature amount into a basis and an activation degree by non-negative matrix factorization, and obtains a basis of voice and a basis of noise. And may be obtained, and the degree of activation corresponding to each base of voice and noise may be obtained.

With this configuration, the degree of activation corresponding to the basis of speech and the basis of noise can be obtained by unsupervised learning.

The speech section detection device of the present invention includes a basis learning unit that learns the basis of speech based on clean speech, and the activation degree calculation unit uses the basis of speech obtained by learning to describe the speech feature. The quantity may be factor-decomposed into the basis and the degree of activation by non-negative matrix factor decomposition, and the degree of activation corresponding to each basis of speech and noise may be obtained.

By using the speech basis learned based on the clean speech that does not contain noise in this way, the accuracy of the basis and the degree of activation can be improved.

The voice section detection device of the present invention includes a data selection unit that selects clean voices based on the voice base calculated by the activation degree calculation unit, and the base learning unit is the selected clean voices. May be used to relearn the basis of speech.

With this configuration, a voice base close to the speaker and the content of the utterance is generated, so that the accuracy of voice section detection can be further improved depending on the user.

The voice section detection device of the present invention may include a noise suppression unit that suppresses noise using the basis and activation degree data calculated by the activation degree calculation unit.

With this configuration, the data of the basis of voice and noise obtained for detecting the voice section and the degree of activation thereof can be used for noise suppression, and the calculation cost of noise suppression can be reduced.

The voice section detection method of the present invention is a method of detecting a voice section by a voice section detection device, wherein the voice section detection device inputs an input voice including voice and noise, and the voice section detection device , The step of obtaining the voice feature amount from the input voice, the step of finding the likelihood of which subset of the voice the voice feature amount is based on the acoustic model, and the voice section detection device. However, there is a step of calculating the posterior probability of voice or non-voice from the voice feature amount based on the voice section detection model, and a step of obtaining the posterior probability using the likelihood that the voice feature amount is voice. The voice section detection device includes a step of determining whether the voice is voice or non-voice based on the posterior probability.

The voice section detection method of the present invention is a method of detecting a voice section by a voice section detection device, wherein the voice section detection device inputs an input voice including voice and noise, and the voice section detection device , The step of obtaining the voice feature amount from the input voice, and the voice section detection device calculates the activation degree when the voice feature amount is separated into voice and noise based on the voice enhancement model. A step and a step in which the voice section detection device calculates a voice or non-voice posterior probability from the voice feature amount based on the voice section detection model, and obtains the posterior probability using the activation degree as well. It includes a step and a step in which the voice section detection device determines whether the voice is voice or non-voice based on the post-probability.

The program of the present invention is a program for detecting a voice section, and is based on a step of inputting an input voice including voice and noise to a computer, a step of obtaining a voice feature amount from the input voice, and an acoustic model. The step of obtaining the likelihood of which subset of the voice the voice feature amount is, and the step of calculating the posterior probability of voice or non-voice from the voice feature amount based on the voice section detection model. The step of obtaining the posterior probability by using the likelihood that the voice feature amount is voice and the step of determining whether it is voice or non-voice based on the posterior probability are executed.

The program of the present invention is a program for detecting a voice section, which includes a step of inputting an input voice including voice and noise to a computer, a step of obtaining a voice feature amount from the input voice, and a voice enhancement model. Based on the step of calculating the activation degree of each when the voice feature amount is separated into voice and noise, and the voice section detection model, the post-probability of voice or non-voice is calculated from the voice feature amount. The step of calculating, the step of obtaining the posterior probability by using the activation degree, and the step of determining whether it is voice or non-voice based on the posterior probability are executed.

According to the present invention, the detection performance of the voice section can be improved.

It is a figure which shows the structure of the voice section detection apparatus of 1st Embodiment. It is a flowchart which shows the operation of the voice section detection apparatus of 1st Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 2nd Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 3rd Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 4th Embodiment. It is a flowchart which shows the operation of the voice section detection apparatus of 4th Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 5th Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 6th Embodiment. It is a figure which shows the structure of the voice section detection apparatus of 7th Embodiment.

Hereinafter, the voice section detection device 1 according to the embodiment of the present invention will be described with reference to the drawings.
(First Embodiment)
FIG. 1 is a diagram showing a configuration of a voice section detection device 1 according to the first embodiment. The input unit 10 has a function of accepting input in frame units of input voice to be detected in a voice section. In this document, the sound data to be detected in which voice and noise are mixed is referred to as "input voice", and the frame unit is referred to as "input voice frame". The feature amount calculation unit 11 calculates the voice feature amount of the input voice frame input by the input unit 10. In the present embodiment, the spectral feature amount v is used as the voice feature amount. The spectral feature amount v has a frequency component included in the input voice frame as a feature amount.

The posterior probability calculation unit 12 has a function of calculating the posterior probability of voice or noise based on the spectral feature amount v. The posterior probability calculation unit 12 is connected to the voice model storage unit 15 and the noise model storage unit 16. The voice model and the noise model are voice section detection models for detecting a voice section, the voice model is a model that learns the acoustic characteristics of voice, and the noise model is a model that learns the acoustic characteristics of noise.

The voice model and the noise model can be learned after the fact, but in the present embodiment, the voice model and the noise model are learned in advance. Considering the use in various situations, it is preferable to mix a wide variety of voices in the training data of the voice model. The noise model storage unit 16 has an arbitrary configuration and can be omitted.

The posterior probability calculation unit 12 calculates the posterior probability w based on the spectral feature amount v. When the conversion from the spectral feature amount v to the posterior probability w by the voice interval detection model is expressed by f (), w = f (v). The posterior probability calculation unit 12 inputs the obtained data of the posterior probability w into the determination unit 13.

The determination unit 13 determines whether the input audio frame is audio or non-audio based on the posterior probability w. For example, if w = [posterior probability of voice, posterior probability of noise] and the vector index starts from 1, when argmax (w) = 1, the input voice frame is voice, and when argmax (w) = 2. The frame can be determined to be noise.

The output unit 14 outputs the result of determining whether the input audio frame is audio or noise. The output unit 14 may output the data of the section determined to be voice by the determination unit 13.

The voice section detection device 1 of the present embodiment further includes an acoustic likelihood calculation unit 17. The feature amount calculation unit 11 inputs the obtained data of the spectrum feature amount v to the acoustic likelihood calculation unit 17. In the present embodiment, the same spectrum feature amount v as input to the posterior probability calculation unit 12 is input, but the feature amount input to the posterior probability calculation unit 12 and the acoustic likelihood feature unit is not necessarily the same. It may not be necessary, and different features may be input for each.

The acoustic likelihood calculation unit 17 applies the input spectrum feature amount v to the acoustic model to obtain the likelihood L in which the spectrum feature amount v is voice. The acoustic likelihood calculation unit 17 obtains the likelihood L in an appropriately selected state unit such as a monophone state, a triphone state, and a degenerate state. The acoustic model is stored in the acoustic model storage unit 18. As the acoustic model, a conventionally used mixed Gaussian distribution, a neural network, or the like can be used. The acoustic likelihood calculation unit 17 inputs the calculated likelihood L to the posterior probability calculation unit 12.

As described above, the posterior probability calculation unit 12 obtains the posterior probability w from the spectral feature amount v by the voice interval detection model, but in the present embodiment, the posterior probability w is further obtained by using the data of the likelihood L. That is, the posterior probability calculation unit 12 connects to the original spectral feature amount v and obtains it as w = f ([v; L]), or w = α using another model g ().・ It can be obtained as f (v) + β · g (L). Here, α and β are appropriate coefficients. The method given here is an example, and various joining methods are conceivable, not limited to the linear sum. Generally, a method of inputting to the integrated model h () such as w = h (f (v), g (L)) can be considered.

Although the configuration of the voice section detection device 1 of the first embodiment has been described above, examples of the hardware of the voice section detection device 1 include a CPU, RAM, ROM, hard disk, display, keyboard, mouse, communication interface, and the like. It is a computer equipped with. The voice section detection device 1 described above is realized by storing a program having a module that realizes each of the above functions in a RAM or ROM and executing the program by a CPU. Such programs are also included in the scope of the present invention.

FIG. 2 is a flowchart showing the operation of the voice section detection device 1 according to the first embodiment. First, the voice section detection device 1 accepts the input of the data of the input voice frame (S10), and obtains the spectral feature amount v of the input voice frame (S11). Subsequently, the voice interval detection device 1 obtains the acoustic likelihood L based on the spectral feature amount v (S12), and inputs the obtained data of the acoustic likelihood L to the posterior probability calculation unit 12. The posterior probability calculation unit 12 applies the spectral feature amount v to the voice model and the noise model, and obtains the posterior probability w of the voice or noise by using the data of the acoustic likelihood L (S13). The determination unit 13 determines whether the input audio frame is audio or noise based on the posterior probability w (S14), and detects the audio section. The voice section detection device 1 outputs the data of the obtained voice section (S15).

The configuration and operation of the voice section detection device 1 of the first embodiment have been described above. The voice section detection device 1 of the first embodiment obtains the posterior probability of voice or noise including the likelihood L that the spectral feature amount of the input voice frame is voice. Since the likelihood L reflects the utterance content, it is useful for detecting the voice section, and by adding it to the calculation of the posterior probability in the posterior probability calculation unit 12, the performance of the voice section detection can be improved.

(Second Embodiment)
FIG. 3 is a diagram showing the configuration of the voice section detection device 2 according to the second embodiment. The basic configuration of the voice section detection device 2 of the second embodiment is the same as that of the voice section detection device 1 of the first embodiment, but the voice section detection device 2 of the second embodiment is In addition to the configuration of the voice section detection device 1 of the first embodiment, the vector conversion unit 19 is provided, and the vector of the acoustic likelihood L obtained by the acoustic likelihood calculation unit 17 is converted to a lower dimension. Is different.

The likelihood L calculated by the acoustic likelihood calculation unit 17 tends to have a larger dimension than the spectral feature amount v. The voice section detection device 2 of the second embodiment is converted into a low-dimensional vector by the vector conversion unit 19. As an example, the conversion function realized by the vector conversion unit 19 is T, and the method of inputting by connecting to the original feature amount v described above is taken as an example, and w = f ([v; T (L)]). As a result, the burden of calculation processing by the posterior probability calculation unit 12 can be reduced.

In the present embodiment, the voice section detection device 2 of the second embodiment has described an example of performing vector conversion for the purpose of dimension compression, but even when it is not necessary to reduce the dimension, the length of the vector Conversions such as normalization and compensation of mean values may be performed.

In this embodiment, an example of performing vector conversion has been described, but as a method of compressing dimensions, principal component analysis or dimension compression by a neural network may be used.

(Third Embodiment)
FIG. 4 is a diagram showing the configuration of the voice section detection device 3 according to the third embodiment. The basic configuration of the voice section detection device 3 of the third embodiment is the same as that of the voice section detection device 2 of the second embodiment, but the voice section detection device 3 of the third embodiment is In addition to the configuration of the voice section detection device 2 of the second embodiment, the voice recognition unit 20 is provided, and voice recognition is performed using the sound likelihood L obtained by the sound likelihood calculation unit 17. Is different.

In this way, it is possible to perform voice recognition at a relatively low cost by combining the voice recognition unit 20. The output unit 14 also outputs the result of recognition by the voice recognition unit 20. In the present embodiment, the configuration in which the voice recognition unit 20 is added to the voice section detection device 2 of the second embodiment has been described, but the voice section detection device 1 of the first embodiment has voice. Of course, the recognition unit 20 may be added.

(Fourth Embodiment)
FIG. 5 is a diagram showing the configuration of the voice section detection device 4 according to the fourth embodiment. In the above-described embodiment, it has been shown that the voice section detection reflecting the utterance content can be performed by using the voice recognition mechanism, but in the voice section detection device 4 of the fourth embodiment, the voice can be detected. Using the enhancement mechanism, the characteristics of the speaker and the speech section detection using one or both of the utterance contents will be explained.

The voice section detection device 4 of the fourth embodiment includes an activation degree calculation unit 21 in place of the acoustic likelihood calculation unit 17 described in the first to third embodiments. The feature amount calculation unit 11 inputs the spectral feature amount obtained from the input voice frame to the activation degree calculation unit 21. The activation degree calculation unit 21 factorizes into the basis and the activation degree (activation) by using the non-negative matrix factorization. Specifically, the activation degree calculation unit 21 calculates the voice base and the noise base, calculates the activation degree corresponding to these, and restores the feature amount from the calculated base and the activation degree. The basis and activation degree are sequentially updated so as to be close to the original features. As a result, the activation degree calculation unit 21 can obtain the activation degree of each of the voice base and the noise base.

The basis and activation degree data obtained by the activation degree calculation unit 21 are dimensionally compressed via the vector conversion unit 19 and input to the posterior probability calculation unit 12. The posterior probability calculation unit 12 obtains the posterior probability w by using the input data of the basis and the degree of activation thereof.

FIG. 6 is a flowchart showing the operation of the voice section detection device 4 according to the fourth embodiment. The voice section detection device 4 first receives the input of the input voice frame (S20) and obtains the spectral feature amount v of the input voice frame (S21). Subsequently, the voice section detection device 4 obtains the basis of voice and noise and their respective activation degrees based on the spectral feature amount v (S22), and obtains the obtained basis and activation degree data in the posterior probability calculation unit 12. Enter in.

The posterior probability calculation unit 12 applies the spectral feature amount v to the voice model and the noise model, and obtains the posterior probability w of the voice or noise by using the data of the acoustic likelihood L (S23). The determination unit 13 determines whether the input audio frame is audio or noise based on the posterior probability w (S24), and detects the audio section. The voice section detection device 4 outputs the data of the obtained voice section (S25).

The configuration and operation of the voice section detection device 4 according to the fourth embodiment have been described above. The voice section detection device 4 of the present embodiment decomposes the input voice frame into the base of voice and noise and the respective activation degrees, and obtains the posterior probability of voice or noise using this information. The degree of activation is useful for detecting voice sections because it probably reflects the characteristics of the speaker and the content of the utterance, and by adding it to the calculation of posterior probabilities by the posterior probability calculation unit 12, the performance of voice section detection Can be enhanced.

(Fifth Embodiment)
FIG. 7 is a diagram showing a configuration of the voice section detection device 5 according to the fifth embodiment. The basic configuration of the voice section detection device 5 of the fifth embodiment is the same as the configuration of the voice section detection device 4 of the fourth embodiment, but the voice section detection device of the fifth embodiment. In addition to the configuration of the voice section detection device 4 of the fourth embodiment, the fifth has a base learning unit 22 and a clean voice storage unit 23 that stores clean voices used for learning in the base learning unit 22. Here, the clean voice is data consisting only of human voice acquired in a noise-free environment.

The base learning unit 22 can obtain a highly accurate voice base by learning the voice base using the clean voice data stored in the clean voice storage unit 23. In the fifth embodiment, the activation degree calculation unit 21 of the voice generated in advance by the base learning unit 22 based on the clean voice when factorizing the spectral features into the basis and the activation degree. Use the basis. As a result, the accuracy of non-negative matrix factorization can be improved, and the accuracy of speech interval detection can be improved.

(Sixth Embodiment)
FIG. 8 is a diagram showing the configuration of the voice section detection device 6 according to the sixth embodiment. The basic configuration of the voice section detection device 6 of the sixth embodiment is the same as the configuration of the voice section detection device 5 of the fifth embodiment, but the voice section detection device of the sixth embodiment. 6 further includes a data selection unit 24 for selecting clean voice data, in addition to the configuration of the voice section detection device 5 of the fifth embodiment.

The basis learning unit 22 first generates a speech basis by learning using clean speech, and this speech basis is a speech basis based on various speech contents of various speakers. In the present embodiment, the basis of the voice is learned from the input unit 10 based on the input voice frame. That is, when the activation degree calculation unit 21 obtains the voice basis and its activation degree from the spectral feature amount of the input voice frame, the data selection unit 24 determines the clean voice having a feature amount close to the obtained voice base. To sort out. Then, the base learning unit 22 relearns the base of the voice using the selected clean voice.

The voice section detection device 6 of the present embodiment can customize the voice base according to the characteristics of the speaker and the utterance content, and can improve the accuracy of the voice section detection. If the gender of the speaker and the content of the utterance (for example, when limited to instructions and commands to the in-vehicle device) are known in advance, the data selection unit 24 selects the clean voice based on the information. , May be passed to the basic learning unit 22. As described above, the voice section detection device 6 of the present embodiment has a configuration that makes it very easy to customize the user according to the characteristics and utterance contents of the speaker.

(7th Embodiment)
FIG. 9 is a diagram showing the configuration of the voice section detection device 7 according to the seventh embodiment. The basic configuration of the voice section detection device 7 of the seventh embodiment is the same as that of the voice section detection device 4 of the fourth embodiment, but the voice section detection device 7 of the seventh embodiment is , A noise suppression unit 25 is provided in addition to the configuration of the voice section detection device 4 of the fourth embodiment, and noise suppression is performed by using the voice base obtained by the activation degree calculation unit 21 and the activation degree thereof. The difference is that the processing of is performed.

In this way, by combining the noise suppression unit 25, it is possible to perform the noise suppression processing at a relatively low cost. The output unit 14 also outputs noise-suppressed voice data. In the present embodiment, an example in which the noise suppression unit 25 is added to the voice section detection device 4 of the fourth embodiment is given, but the fifth embodiment and the sixth embodiment are given. Of course, it is also possible to add the noise suppression unit 25 to the voice section detection devices 5 and 6.

Although the voice section detection device of the present invention has been described in detail with reference to the embodiments, the voice section detection device of the present invention is not limited to the above-described embodiment. In the above-described embodiment, the posterior probability calculation unit uses the acoustic likelihood L (first to third embodiments) and the basis and activation degree (fourth to third). Although the seventh embodiment) has been described, the acoustic likelihood L may be used in combination with the basis and the activation degree, and further improvement in performance can be expected.

The present invention is useful as a device for detecting a voice section of an input voice frame.

1-7 Voice section detection device 10 Input unit 11 Feature amount calculation unit 12 Posterior probability calculation unit 13 Judgment unit 14 Output unit 15 Voice model storage unit 16 Noise model storage unit 17 Acoustic likelihood calculation unit 18 Acoustic model 19 Vector conversion unit 21 Activation degree calculation unit 22 Basic learning unit 23 Clean voice storage unit 24 Data selection unit 25 Noise suppression unit

Claims (12)

Input section that inputs voice and noise, and an input unit that inputs voice,
A feature amount calculation unit that obtains a voice feature amount from the input voice,
Based on the acoustic model, the acoustic likelihood calculation unit that obtains the likelihood that the voice feature amount is voice , and
A posterior probability calculation unit that calculates the posterior probability of voice or non-voice from the voice feature amount based on the voice interval detection model, and the posterior probability using the likelihood calculated by the acoustic likelihood calculation unit. The posterior probability calculation unit that finds
A determination unit that determines whether it is voice or non-voice based on the posterior probability,
A voice section detection device comprising. A dimension compression unit for compressing the dimension of the likelihood data calculated by the acoustic likelihood calculation unit is provided.
The voice section detection device according to claim 1, wherein the likelihood data whose dimensions are compressed by the dimension compression unit is input to the posterior probability calculation unit. The voice section detection device according to claim 1 or 2, further comprising a voice recognition unit that performs voice recognition using the likelihood data calculated by the acoustic likelihood calculation unit. Input section that inputs voice and noise, and an input unit that inputs voice,
A feature amount calculation unit that obtains a voice feature amount from the input voice,
Based on the speech enhancement model, an activation degree calculation unit that calculates the activation degree when the voice feature amount is separated into voice and noise, and an activation degree calculation unit.
A posterior probability calculation unit that calculates posterior probabilities of voice or non-voice from the voice feature amount based on the voice interval detection model, and also uses the activation degree calculated by the activation degree calculation unit. The posterior probability calculation unit for finding the probability and
A determination unit that determines whether it is voice or non-voice based on the posterior probability,
A voice section detection device comprising. The activation degree calculation unit factorizes the voice feature amount into a basis and an activation degree by non-negative matrix factorization, obtains a voice base and a noise base from the obtained base, and obtains each of the voice and noise. The voice section detection device according to claim 4, wherein the degree of activation corresponding to the basis of the above is obtained. Equipped with a basis learning unit that learns the basis of speech based on clean speech
The activation degree calculation unit factorizes the voice features into bases and activation degrees by non-negative matrix factorization using the voice bases obtained by learning, and corresponds to each base of voice and noise. The voice section detection device according to claim 4, wherein the degree of activation is determined. A data selection unit that selects clean voice based on the voice base calculated by the activation degree calculation unit is provided.
The voice section detection device according to claim 6, wherein the base learning unit relearns the base of the voice by using the selected clean voice. The voice section detection device according to any one of claims 4 to 7, further comprising a noise suppression unit that suppresses noise using the basis and activation degree data calculated by the activation degree calculation unit. It is a method of detecting a voice section by a voice section detection device.
A step in which the voice section detection device inputs input voice including voice and noise,
A step in which the voice section detection device obtains a voice feature amount from the input voice,
A step in which the voice section detection device obtains a likelihood that the voice feature amount is voice based on an acoustic model.
The voice section detection device is a step of calculating the posterior probability of voice or non-voice from the voice feature amount based on the voice section detection model, and the posterior probability using the likelihood that the voice feature amount is voice is also used. Steps to find the probability and
A step in which the voice section detection device determines whether the voice is voice or non-voice based on the posterior probability.
A voice section detection method comprising. It is a method of detecting a voice section by a voice section detection device.
A step in which the voice section detection device inputs input voice including voice and noise,
A step in which the voice section detection device obtains a voice feature amount from the input voice,
A step of calculating the activation degree of each of the voice feature amounts when the voice feature amount is separated into voice and noise based on the voice enhancement model by the voice section detection device.
The voice section detection device calculates the posterior probability of voice or non-voice from the voice feature amount based on the voice section detection model, and obtains the posterior probability using the activation degree as well.
A step in which the voice section detection device determines whether the voice is voice or non-voice based on the posterior probability.
A voice section detection method comprising. A program for detecting audio sections, which can be applied to a computer.
Input including voice and noise Steps to input voice and
The step of obtaining the voice feature amount from the input voice and
Based on the acoustic model, the step of finding the likelihood that the voice feature is voice , and
A step of calculating the posterior probability of voice or non-voice from the voice feature amount based on the voice interval detection model, and a step of obtaining the posterior probability using the likelihood that the voice feature amount is voice.
The step of determining whether it is voice or non-voice based on the posterior probability,
A program that executes. A program for detecting audio sections, which can be applied to a computer.
Input including voice and noise Steps to input voice and
The step of obtaining the voice feature amount from the input voice and
Based on the speech enhancement model, the step of calculating the activation degree of each of the speech features when they are separated into speech and noise, and
A step of calculating the posterior probability of voice or non-voice from the voice feature amount based on the voice interval detection model, and a step of obtaining the posterior probability using the activation degree as well.
The step of determining whether it is voice or non-voice based on the posterior probability,
A program that executes.