JP2019028300A - Acoustic signal processing apparatus, method and program - Google Patents

Abstract

To provide an acoustic signal processing apparatus and the like that can detect at least either of a voice section or a non-voice section more accurately than before even when voice desired to recognize and noise have close characteristics.SOLUTION: An acoustic signal processing apparatus determines that a section of a time-sequence acoustic signal which is estimated as a section corresponding to a specific sound on the basis of detection time is a specific sound voice section, determines that a section of a time-sequence acoustic signal which is estimated not as the section corresponding to the specific sound on the basis of the detection time is a non-voice section, calculates a voice section feature amount being its feature amount from the time-sequence acoustic signal corresponding to the specific sound voice section, calculates a non-voice section feature amount being its feature amount from the time-sequence acoustic signal corresponding to the non-voice section, obtains a voice parameter indicating a feature of the voice section from the voice section feature amount, obtains non-voice parameter indicating a feature of the non-voice section from the non-voice section feature amount, and detects at least either of the voice section or the non-voice section from the time-sequence acoustic signal using the voice parameter and the non-voice parameter.SELECTED DRAWING: Figure 13

Description

  The present invention relates to an acoustic signal processing technique.

  In an application such as voice recognition, it is important to detect “in which section the voice is spoken”. However, for example, in a noisy environment, the difference in signal between utterances and non-utterances becomes small, and it is difficult to simply detect the speech section with the sound volume. For example, Patent Document 1 discloses a method for detecting a voice section. In Patent Document 1, the accuracy of section detection is improved by using a speech model to determine whether the sound signal is speech / non-speech.

JP 2009-63700 A

  However, for example, when the sound or music of the TV is in the background noise and the voice to be recognized and the noise have characteristics close to each other, the method of Patent Document 1 makes it difficult to detect the voice section.

  The present invention provides an acoustic signal processing apparatus and method capable of detecting at least one of a speech segment and a non-speech segment with higher accuracy than in the past even when the speech to be recognized and noise have close characteristics. And to provide a program.

  In order to solve the above problems, according to one aspect of the present invention, an acoustic signal processing device includes a time-series acoustic signal including a specific sound that is a predetermined sound emitted by a person, and a specification included in the time-series acoustic signal. Information indicating the detection time of the sound as an input, the section corresponding to the specific sound based on the detection time as a specific sound voice section that is estimated as the section corresponding to the specific sound, and in the section corresponding to the specific sound based on the detection time A specific sound voice interval calculation unit that determines that a time-series acoustic signal section that is estimated to be a non-speech section is calculated, and a voice section feature amount that is a feature amount is calculated from a time-series acoustic signal corresponding to the specific sound voice section. A feature amount calculation unit for calculating a feature amount of a non-speech segment that is a feature amount from a time-series acoustic signal corresponding to the non-speech segment; and a speech parameter indicating a feature of the speech segment from the speech segment feature amount; Non-sound from features Seeking non-voice parameters indicating the characteristics of the section, and a speech section detection unit for detecting at least one of the time-series audio signals and the voice section and the non-speech section using the speech parameters and non-speech parameters.

  In order to solve the above problems, according to another aspect of the present invention, an acoustic signal processing method includes: a time-series acoustic signal including a specific sound that is a predetermined sound emitted by a person; Using the information indicating the detection time of the specific sound included in the sequence acoustic signal, the section corresponding to the specific sound based on the detection time as the specific sound voice section, and based on the detection time A specific sound / voice section calculation step for determining a section of a time-series sound signal estimated not to be a section corresponding to a specific sound as a non-speech section, and a feature amount from the time-series sound signal corresponding to the specific sound / voice section A feature amount calculating step for calculating a speech segment feature value and calculating a non-speech segment feature value, which is a feature value from a time-series acoustic signal corresponding to the non-speech segment, and a voice indicating the feature of the speech segment from the speech segment feature value Parameters Obtaining a non-speech parameter indicating the feature of the non-speech segment from the non-speech segment feature quantity, and detecting at least one of the speech segment and the non-speech segment from the time-series acoustic signal using the speech parameter and the non-speech parameter A speech segment detection step.

  Advantageous Effects of Invention According to the present invention, there is an effect that it is possible to detect at least one of a speech segment and a non-speech segment with higher accuracy than in the past even when the speech to be recognized and noise have close characteristics. .

The block diagram for demonstrating the example of the acoustic signal processing apparatus of 1st embodiment. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 1 of 1st embodiment. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 2 of 1st embodiment. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 3 of 1st embodiment. The flowchart for demonstrating the example of the acoustic signal processing method. The block diagram for demonstrating the example of the acoustic signal processing apparatus of 2nd embodiment. The block diagram for demonstrating the example of the direction estimation part 22 of 2nd embodiment. The block diagram for demonstrating the example of the direction estimation part 22 of 2nd embodiment. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 2 of 2nd embodiment. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 3 of 2nd embodiment. The flowchart for demonstrating the example of the acoustic signal processing method. The block diagram for demonstrating the example of the directional sound collector of background art. The functional block diagram of the acoustic signal processing apparatus which concerns on 3rd embodiment. The figure which shows the example of the processing flow of the acoustic signal processing apparatus which concerns on 3rd embodiment. The functional block diagram of the audio | voice area detection information storage part which concerns on 3rd embodiment. The figure for demonstrating a specific sound audio | voice area and a non-voice area. The functional block diagram of the audio | voice area detection part which concerns on 3rd embodiment. The functional block diagram of the 1st acoustic signal analysis part which concerns on 3rd embodiment. The functional block diagram of the probability estimation part which concerns on 3rd embodiment. The functional block diagram of the audio | voice area detection part which concerns on the 1st modification of 3rd embodiment. The functional block diagram of the acoustic signal processing apparatus which concerns on the 3rd modification of 3rd embodiment, and a 4th modification. The figure which shows the example of the processing flow of the acoustic signal processing apparatus which concerns on the 3rd modification of 3rd embodiment, and a 4th modification. The block diagram for demonstrating the example of the acoustic signal processing apparatus of the modification 4 of 1st embodiment. The flowchart for demonstrating the example of the acoustic signal processing method of the modification 4 of 1st embodiment.

  In the drawings used for the following description, components having the same function and steps for performing the same process are denoted by the same reference numerals, and redundant description is omitted. In the following explanation, the symbol “^” etc. used in the text should be described immediately above the character immediately after it, but it is described immediately before the character due to restrictions on the text notation. In the formula, these symbols are written in their original positions. Further, the processing performed for each element of a vector or matrix is applied to all elements of the vector or matrix unless otherwise specified.

[Technical background]
The acoustic signal processing apparatus performs acoustic signal processing using information about the specific sound, assuming that information about the specific sound that is a predetermined sound is given. By using information about a specific sound given in advance, usable information increases, so that more accurate acoustic signal processing can be performed.

  Examples of acoustic signal processing are estimation of sound arrival direction, directional sound collection, target speech extraction, speech section detection, and speech recognition.

  For example, by detecting a keyword that is a specific sound for a user's specific utterance, the signal section of the target speech and the signal section of the noise can be accurately grasped, and can be utilized for subsequent processing.

  Further, when this property is used for speech section detection, the noise section and the speech section signals are respectively found, and therefore the parameters for speech / non-speech determination can be updated to values that more closely match the actually measured values.

  Also, when performing speech direction estimation as acoustic signal processing, the direction in which the specific sound is detected is regarded as the direction of the speech, so that the direction estimation is robust even if sound including speech comes from other than the original direction. Operate.

  In addition, when target speech extraction is performed as acoustic signal processing, signals in speech sections and non-speech sections can be obtained with high accuracy, so that a spatial correlation matrix for calculating a steering vector for speech separation is obtained more accurately. be able to.

  Also, when performing speech recognition as acoustic signal processing, the noise level can be obtained more accurately, so that accuracy can be improved by selecting an acoustic model.

  Hereinafter, each embodiment will be described with reference to the drawings.

[First embodiment]
The acoustic signal processing apparatus and method according to the first embodiment performs directional sound collection processing as acoustic signal processing.

  As shown in FIG. 11, the acoustic signal processing device includes, for example, a direction estimation unit 11, a specific sound detection unit 12, a direction storage unit 13, and a first directivity sound collection unit 14. The acoustic signal processing device may not include the specific sound detection unit 12.

  The acoustic signal processing method is realized, for example, by the acoustic signal processing apparatus performing the processing from step S11 to step S14 described below with reference to FIG.

  The direction estimation unit 11 estimates the arrival direction of sound from signals collected by a plurality of microphones (step S11). The direction estimation unit 11 estimates the arrival direction of sound at each time. The estimated direction of arrival of the sound at each time is output to the direction storage unit 13.

  The direction estimation method by the direction estimation unit 11 is arbitrary. The direction estimation unit 11 estimates the direction of arrival of sound using, for example, the direction estimation technique described in Patent Literatures 1 and 2. The direction of arrival of sound may be represented not by direction but by position.

  The specific sound detection unit 12 detects a specific sound that is a predetermined sound (step S12). Examples of predetermined sounds are voice, whistle, and clapping of a specific keyword. A predetermined sound other than the above example may be used as the predetermined sound.

  The direction storage unit 13 stores the arrival direction estimated by the direction estimation unit 11 at the time when the specific sound is detected by the specific sound detection unit 12. More specifically, the direction storage unit 13 stores the sound arrival direction at the time when the specific sound is detected by the specific sound detection unit 12 among the sound arrival directions at each time input from the direction estimation unit 11. .

  The first directivity sound collection unit 14 performs sound collection so that the sound from the direction of arrival read from the direction storage unit 13 is emphasized (step S14). The method of directivity sound collection by the first directivity sound collection unit 14 is arbitrary. The first directivity sound collection unit 14 performs directivity sound collection described in, for example, Japanese Patent Application Laid-Open No. 2009-44588.

  In this way, it is possible to collect sound with a high S / N ratio by discriminating the sound source from which the specific sound is emitted as the sound source to be collected and collecting the sound source with directional sound. The user can change the direction of the directivity by emitting a specific sound such as a specific keyword. Even when a sound source such as a TV is present, the user directs the directivity toward the user and then fixes it. be able to.

  In addition, when it takes time for the specific sound to be detected by the specific sound detection unit 12, a delay unit 15 that delays the specific sound by a time corresponding to the time may be placed after the direction estimation unit 11. In FIG. 1, the delay unit 15 is indicated by a broken line. The delay unit 15 delays the output from the direction estimation unit 11 by a time corresponding to the time of detection of the specific sound by the specific sound detection unit 12 and then inputs the delay to the direction storage unit 13. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[First Modification of First Embodiment]]
As illustrated in FIG. 2, the acoustic signal processing device may further include an estimated frequency measurement unit 16 and a selection unit 17. In this case, the direction estimation unit 11 may be capable of simultaneous estimation in a plurality of directions. That is, the direction estimating unit 11 may be able to estimate the directions of both sound sources when there is a noise source sound simultaneously with the specific sound. In this case, since it becomes impossible to determine which sound source has generated the specific sound, the estimation frequency measurement unit 16 performs the determination based on how much direction estimation has been performed in the past. That is, since the sound source such as the TV always outputs sound, the estimated frequency measuring unit 16 determines that many directions have been estimated in the past.

  The estimated frequency measuring unit 16 measures the frequency of the arrival direction estimated by the direction estimating unit 11 in a predetermined past time interval (step S16). That is, the estimated frequency measuring unit 16 measures how often the direction is estimated within a certain past time. Information about the measured frequency is output to the selection unit 17.

  For example, if the time during which the output of the direction estimation unit 11 is in the direction θ during the past T seconds is A (θ) seconds, the estimated frequency in the θ direction is the ratio D (θ) = A (θ ) / T. The estimated frequency measuring unit 16 obtains all the frequencies in each direction. If it is assumed that the noise source is a TV or a speaker for listening to music, the sound is emitted from the same direction with almost no silence for a long time. When such a sound source is in the θ direction, the estimated frequency D (θ) takes a large value close to 1.

  The selection unit 17 selects the arrival direction with the lowest frequency among the frequencies measured by the estimated frequency measurement unit 16. For example, when there are two estimated directions of the output of the direction estimation unit 11, the selection unit 17 selects the one with the smaller estimation frequency D (θ). The direction of arrival selected by the selection unit 17 at the time when the specific sound is detected by the specific sound detection unit 12 is stored in the direction storage unit 13.

  Thereafter, the first directivity sound collecting unit 14 performs sound collection so that the sound from the arrival direction read from the direction storage unit 13 is emphasized in the same manner as described above.

  Also in the first modification of the first embodiment, when it takes time for the specific sound to be detected by the specific sound detection unit 12, the delay unit 15 that delays by the time corresponding to the time is provided after the direction estimation unit 11. You may put in. In FIG. 2, the delay unit 15 is indicated by a broken line. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[Modification 2 of the first embodiment]]
As illustrated in FIG. 3, the acoustic signal processing device may further include a second directional sound collection unit 18.

  By performing the directional sound collection by the second directional sound collection unit 18 before the process of the specific sound detection unit 12, it is possible to detect the specific sound with higher accuracy.

  A signal obtained by delaying signals collected by a plurality of microphones is input to the second directivity sound collecting unit 18. This delay has a length of time corresponding to the time required for the arrival direction estimation processing by the direction estimation unit 11. This delay is performed by a delay unit 19 indicated by a broken line in FIG. Further, the arrival direction estimated by the direction estimation unit 11 is input to the second directivity sound collection unit 18.

  The second directivity sound collection unit 18 collects sound so that the sound from the direction of arrival estimated by the direction estimation unit 11 is emphasized (step S18). More specifically, the second directional sound collection unit 18 emphasizes the sound from the direction of arrival estimated by the direction estimation unit 11 using a signal obtained by delaying signals collected by a plurality of microphones. So that the sound is collected. The signal collected by the second directional sound collection unit 18 is output to the specific sound detection unit 12.

  The specific sound detection unit 12 detects a specific sound based on the signal collected by the second directional sound collection unit 18. Subsequent processing is the same as described above.

  In addition, as shown in FIG. 3, the some 2nd directivity sound collection part 18 may be provided in the acoustic signal processing apparatus. In this case, the same number of specific sound detection units 12 as the number of second directivity sound collection units 18 are provided in the acoustic signal processing device.

  In this case, when a plurality of arrival directions are estimated by the direction estimation unit 11, the specific sound detection unit 12 operates to emphasize each of the estimated plurality of arrival directions, and a plurality of outputs thereof are provided. The specific sound is input to the specific sound detection unit 12 and the specific sound is detected.

  Thereby, when a specific sound is detected by a plurality of specific sound detection units 12, it becomes possible to give priority.

  Also in the second modification of the first embodiment, when it takes time for the specific sound to be detected by the specific sound detection unit 12, the delay unit 15 that delays by the time corresponding to the time is included in the subsequent stage of the direction estimation unit 11. You may put in. In FIG. 2, the delay unit 15 is indicated by a broken line. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[Modification 3 of the first embodiment]]
As illustrated in FIG. 4, in Modification 2 of the first embodiment, the acoustic signal processing device may further include the estimation frequency measurement unit 16 and the selection unit 17 described in Modification 1 of the first embodiment. . In this case, the direction estimation unit 11 may be capable of simultaneous estimation in a plurality of directions. That is, the direction estimating unit 11 may be able to estimate the directions of both sound sources when there is a noise source sound simultaneously with the specific sound.

  The processes of the estimation frequency measurement unit 16 and the selection unit 17 are the same as those described in the first modification of the first embodiment.

  That is, the estimated frequency measuring unit 16 measures the frequency of the arrival direction estimated by the direction estimating unit 11 in a past predetermined time interval (step S16). That is, the estimated frequency measuring unit 16 measures how often the direction is estimated within a certain past time. Information about the measured frequency is output to the selection unit 17.

  For example, if the time during which the output of the direction estimation unit 11 is in the direction θ during the past T seconds is A (θ) seconds, the estimated frequency in the θ direction is the ratio D (θ) = A (θ ) / T. The estimated frequency measuring unit 16 obtains all the frequencies in each direction. If it is assumed that the noise source is a TV or a speaker for listening to music, the sound is emitted from the same direction with almost no silence for a long time. When such a sound source is in the θ direction, the estimated frequency D (θ) takes a large value close to 1.

  The selection unit 17 selects the arrival direction with the lowest frequency among the frequencies measured by the estimated frequency measurement unit 16. For example, when there are two estimated directions of the output of the direction estimation unit 11, the selection unit 17 selects the one with the smaller estimation frequency D (θ). The direction of arrival selected by the selection unit 17 at the time when the specific sound is detected by the specific sound detection unit 12 is stored in the direction storage unit 13.

  Thereafter, the first directivity sound collecting unit 14 performs sound collection so that the sound from the arrival direction read from the direction storage unit 13 is emphasized in the same manner as described above.

  Also in the first modification of the first embodiment, when it takes time for the specific sound to be detected by the specific sound detection unit 12, the delay unit 15 that delays by the time corresponding to the time is provided after the direction estimation unit 11. You may put in. In FIG. 4, the delay unit 15 is indicated by a broken line. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[Modification 4 of the first embodiment]]
As illustrated in FIG. 23, the acoustic signal processing device may include a third directional sound collecting unit 52 instead of the first directional sound collecting unit 14 and may further include a noise direction storage unit 51.

  The acoustic signal processing method is realized, for example, by the acoustic signal processing device performing the processing of FIG. 24 and step S31 described below.

  The noise direction storage unit 51 stores the arrival direction estimated by the direction estimation unit 11 excluding the time when the specific sound is detected by the specific sound detection unit 12. Here, excluding the time when the specific sound is detected may be a time before the time when the specific sound is detected or may be a time after the time series. It may be both the previous time and the later time. Needless to say, the delay unit 15 may be inserted before the noise direction storage unit 51 and after the direction estimation unit 11.

The third directivity sound collecting unit 52 collects the sound so that the sound from the direction of arrival read from the direction storage unit 13 is emphasized and the sound from the direction of arrival read from the noise direction storage unit 51 is suppressed. This is performed (step S52). The method of directivity sound collection by the third directivity sound collection unit 52 is arbitrary. As a method of directivity sound collection performed by the third directivity sound collecting unit 52, for example, the method described in Reference 5 may be used.
(Reference 5) Tadashi Asano, “Sound Array Signal Processing”, pp.82-85, Corona, 2011.

[Second Embodiment]
The acoustic signal processing apparatus and method according to the first embodiment performs directional sound collection processing as acoustic signal processing.

  As shown in FIG. 6, the acoustic signal processing device includes, for example, a specific sound detection unit 21, a direction estimation unit 22, and a first directivity sound collection unit 23. The acoustic signal processing device may not include the specific sound detection unit 12.

  The acoustic signal processing method is realized, for example, by the acoustic signal processing apparatus performing the processing from step S21 to step S23 described below with reference to FIG.

  The specific sound detection unit 21 detects a specific sound that is a predetermined sound (step S21). Examples of predetermined sounds are voice, whistle, and clapping of a specific keyword. A predetermined sound other than the above example may be used as the predetermined sound.

  The direction estimation unit 22 estimates the arrival direction of sound from signals collected by a plurality of microphones (step S22). At that time, the direction estimation unit 22 determines the arrival direction of the sound from the signals collected by the plurality of microphones, and the direction closer to the arrival direction estimated at the time when the specific sound is detected by the specific sound detection unit 21. It is estimated so that it is easy to be estimated.

  That is, in the direction estimation unit 22, the ease of detection in each direction is set according to the result of detection of the specific sound. In other words, in the direction estimation unit 22, the closer to the direction estimated at the time of detecting the specific sound, the easier the direction is detected, and the farther the direction is, the harder it is to detect. By doing so, the directivity is easily directed to the user who has emitted the specific sound, and the directivity is not easily directed to the noise source. Moreover, even if the user who emitted the specific sound moves, it can follow it.

  An example of the configuration of the direction estimation unit 22 is shown in FIG. As illustrated in FIG. 7, the direction estimation unit 22 includes a direction enhancement unit 221, a power calculation unit 222, a weight multiplication unit 223, a maximum power direction detection unit 224, and a weight determination unit 225.

  Each of the signals collected by the plurality of microphones is input to the direction enhancement unit 221.

  The direction emphasizing unit 221 performs direction emphasis processing on the signals collected by the plurality of microphones so as to emphasize each of the plurality of directions (step S221). For example, when N direction enhancing units 221 are provided, θ1, θ2,..., ΘN are set as directions different from each other, and the N direction enhancing units 221 have directions of θ1, θ2,. The direction emphasis process is performed so as to emphasize. The emphasized signal is output to the power calculation unit 222.

  The power calculation unit 222 calculates the power of the signal emphasized by the direction enhancement unit 221 (step S222). The calculated power is output to the weight multiplication unit 223.

  The weight multiplication unit 223 multiplies the power calculated by the power calculation unit 222 by the weight set by the weight setting unit 225 (step S223). The power after weighting is output to the maximum power direction detection unit 224. As will be described later, therefore, the weight multiplier 223 multiplies the power of the signal in which each arrival direction is emphasized by multiplying the power after weighting by increasing the weight as the arrival direction is closer to the selected arrival direction. Get.

  The maximum power direction detection unit 224 selects the arrival direction of the maximum power among the outputs of the weight multiplication unit 223. In other words, the maximum power direction detection unit 224 selects the arrival direction having the largest weighted power, and sets the selected arrival direction as the estimated arrival direction (step S224). The estimated arrival direction is output to the weight determination unit 225 and the first directivity sound collection unit 23 as a direction estimation result.

  The weight setting unit 225 determines a weight corresponding to the direction estimation result output by the maximum power direction detection unit 224 at the time when the specific sound is detected by the specific sound detection unit 21. The determined weight is output to the weight multiplier 223. In other words, the weight setting unit 225 sets a weight corresponding to the direction estimation result when the specific sound is detected.

  The weight corresponding to the direction estimation result is set such that the weight with respect to the estimated arrival direction increases and the weight decreases as the distance from the arrival direction increases. For example, the weight for the estimated direction of arrival is set to 1.0, and a weight obtained by multiplying a multiplier (for example, 0.8) less than 1.0 is set every time the estimated direction of arrival is deviated by 10 degrees.

  The first directivity sound collection unit 23 performs sound collection so that the sound from the direction of arrival estimated by the direction estimation unit 22 is emphasized (step S23). The method of directivity collection by the first directivity sound collection unit 23 is arbitrary. The first directivity sound collection unit 23 performs directivity sound collection described in, for example, Japanese Patent Application Laid-Open No. 2009-44588.

  In this way, it is possible to collect sound with a high S / N ratio by discriminating the sound source from which the specific sound is emitted as the sound source to be collected and collecting the sound source with directional sound. The user can change the direction of the directivity by emitting a specific sound such as a specific keyword. Even when a sound source such as a TV is present, the user directs the directivity toward the user and then fixes it. be able to.

  When it takes time to detect a specific sound by the specific sound detection unit 21, a delay unit 226 that delays by a time corresponding to the time may be placed after the maximum power direction detection unit 224. In FIG. 7, the delay unit 226 is indicated by a broken line. The delay unit 226 delays the output from the maximum power direction detection unit 224 by a time corresponding to the time of detection of the specific sound by the specific sound detection unit 21 and then inputs the delay to the weight setting unit 225. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[Modification 1 of the second embodiment]]
As illustrated in FIG. 8, the acoustic signal processing device may further include an estimated frequency measurement unit 227 and a selection unit 228.

  In this case, the maximum power direction detection unit 224 may be capable of simultaneous estimation in a plurality of directions by detecting all power directions exceeding a predetermined threshold. That is, the maximum power direction detection unit 224 detects the direction of the maximum power, and further detects the direction of the maximum power except for the detected direction. The maximum power direction detection unit 224 ends the maximum power detection when the preset maximum number of estimated directions is reached or the maximum power is equal to or less than a preset threshold value. The maximum power direction detection unit 224 may be capable of simultaneously estimating the directions of a plurality of sound sources by such a method, for example. Thereby, the maximum power direction detection unit 224 can estimate the directions of both sound sources when there is a noise source sound simultaneously with the specific sound.

  In this case, since it becomes impossible to determine which sound source has generated the specific sound, the estimation frequency measurement unit 227 performs the determination based on how much direction estimation has been performed in the past. That is, the estimated frequency measuring unit 227 determines that a large number of directions have been estimated in the past because sound is always output from a sound source such as a TV.

  The estimated frequency measuring unit 227 measures the frequency of the arrival direction estimated by the direction estimation unit 22 in a predetermined past time period, in other words, the frequency of the arrival direction selected by the maximum power direction detection unit 22 (step) S16). That is, the estimated frequency measuring unit 227 measures how often the direction is estimated within a certain past time. Information about the measured frequency is output to the selection unit 228.

  For example, if the time during which the output of the maximum power direction detector 224 is in the direction θ during the past T seconds is A (θ) seconds, the estimated frequency in the θ direction is the ratio D (θ) = A It is obtained by (θ) / T. The estimated frequency measuring unit 227 calculates all the frequencies for each direction. If it is assumed that the noise source is a TV or a speaker for listening to music, the sound is emitted from the same direction with almost no silence for a long time. When such a sound source is in the θ direction, the estimated frequency D (θ) takes a large value close to 1.

  The selection unit 228 selects the arrival direction with the lowest frequency among the frequencies measured by the estimated frequency measurement unit 227. For example, when there are two estimated directions of the output of the maximum power direction detection unit 22, the selection unit 228 selects the one having the smaller estimated frequency D (θ). The selected arrival direction is output to the weight setting unit 225.

  When it takes time to detect a specific sound by the specific sound detection unit 21, a delay unit 226 that delays by a time corresponding to the time may be placed after the maximum power direction detection unit 224. In FIG. 8, the delay unit 226 is indicated by a broken line. The delay unit 226 delays the output from the maximum power direction detection unit 224 by a time corresponding to the time of detection of the specific sound by the specific sound detection unit 21 and then inputs the delay to the weight setting unit 225. Thereby, it operates normally even if there is a delay in the detection of the specific sound.

[[Modification 2 of the second embodiment]]
As illustrated in FIG. 9, the acoustic signal processing device may further include a second directional sound collection unit 24.

  By performing the directional sound collection by the second directional sound collection unit 24 before the processing of the specific sound detection unit 21, it is possible to detect the specific sound with higher accuracy.

  A signal obtained by delaying signals collected by a plurality of microphones is input to the second directivity sound collecting unit 24. This delay has a length of time corresponding to the time required for the arrival direction estimation processing by the direction estimation unit 22. This delay is performed by the delay unit 25 indicated by a broken line in FIG. Further, the arrival direction estimated by the direction estimation unit 22 is input to the second directivity sound collection unit 24.

  The second directivity sound collecting unit 24 performs sound collection so that the sound from the direction of arrival estimated by the direction estimating unit 22 is emphasized (step S24). More specifically, the second directional sound collection unit 24 emphasizes the sound from the direction of arrival estimated by the direction estimation unit 22 using a signal obtained by delaying signals collected by a plurality of microphones. So that the sound is collected. The signal collected by the second directional sound collection unit 24 is output to the specific sound detection unit 21.

  The specific sound detection unit 21 detects a specific sound based on the signal collected by the second directivity sound collection unit 24. The subsequent processing is the same as described above.

  In addition, as shown in FIG. 9, the some 2nd directivity sound collection part 24 may be provided in the acoustic signal processing apparatus. In this case, the same number of specific sound detectors 21 as the number of second directivity sound collectors 24 are provided in the acoustic signal processing device.

  In this case, when a plurality of arrival directions are estimated by the direction estimation unit 22, the specific sound detection unit 21 operates to emphasize each of the estimated plurality of arrival directions, and a plurality of outputs thereof are provided. The specific sound is input to the specific sound detection unit 21 and the specific sound is detected.

  Thereby, when a specific sound is detected by a plurality of specific sound detectors 21, it is possible to give priority.

[[Modification 3 of the second embodiment]]
As illustrated in FIG. 10, in Modification 2 of the second embodiment, the acoustic signal processing device may further include an estimated frequency measurement unit 26 and a selection unit 27. In this case, the direction estimation unit 22 may be capable of simultaneous estimation in a plurality of directions. That is, the direction estimating unit 22 may be able to estimate the directions of both sound sources when there is a sound of a noise source simultaneously with the specific sound.

  The processes of the estimation frequency measurement unit 26 and the selection unit 27 are the same as those described in the first modification of the first embodiment.

  That is, the estimated frequency measuring unit 26 measures the frequency of the arrival direction estimated by the direction estimating unit 22 in a past predetermined time interval (step S26). That is, the estimated frequency measuring unit 26 measures how often the direction has been estimated within a certain past time. Information about the measured frequency is output to the selection unit 27.

  For example, if the time during which the output of the direction estimation unit 22 is in the direction θ during the past T seconds is A (θ) seconds, the estimated frequency in the θ direction is the ratio D (θ) = A (θ ) / T. The estimated frequency measuring unit 26 obtains all the frequencies in each direction. If it is assumed that the noise source is a TV or a speaker for listening to music, the sound is emitted from the same direction with almost no silence for a long time. When such a sound source is in the θ direction, the estimated frequency D (θ) takes a large value close to 1.

  The selection unit 27 selects the arrival direction with the lowest frequency among the frequencies measured by the estimated frequency measurement unit 26 (step S27). For example, when there are two estimated directions of the output of the direction estimating unit 22, the selecting unit 27 selects the one having the smaller estimated frequency D (θ). The arrival direction selected by the selection unit 27 at the time when the specific sound is detected by the specific sound detection unit 21 is output to the direction estimation unit 22 and is the arrival direction estimated by the direction estimation unit 22.

  Thereafter, the first directivity sound collection unit 23 collects sound so that the sound from the arrival direction estimated by the direction estimation unit 22 is emphasized in the same manner as described above.

[Third embodiment]
The acoustic signal processing apparatus and method according to the third embodiment detect a voice section as acoustic signal processing.

<Points of third embodiment>
In the present embodiment, by narrowing down the user's utterance content, information on the usage environment (noise, etc.) can be obtained more correctly. For example, the user is restricted to utter a specific word (keyword) before starting to speak. At that time, it is assumed that only the specific word speech can be detected with high accuracy, and that “the section is speech” and “the previous section is noise”. Then, the information for the determination of “voice / non-voice” is updated using the voice of the noise section and the voice section.

  By doing so, when determining a section of a target speech to be subsequently issued, information on “noise” and “speech” more suited to the actual usage environment can be used, and the accuracy of section detection is improved.

  Hereinafter, embodiments of the acoustic signal processing apparatus and method will be described. The acoustic signal processing apparatus is realized by a computer such as a dedicated machine configured by dedicated hardware or a general-purpose machine such as a personal computer. Here, description will be made on the case where it is realized by a computer (general-purpose machine).

  A hardware configuration example of the acoustic signal processing device will be described.

  The acoustic signal processing device consists of an input unit to which a keyboard, pointing device, etc. can be connected, an output unit to which a liquid crystal display, a CRT (Cathode Ray Tube) display, etc. can be connected, and a communication device that can communicate outside the acoustic signal processing device ( For example, a communication unit to which a communication cable, a LAN card, a router, a modem, or the like) can be connected and a CPU (Central Processing Unit) [DSP (Digital Signal Processor) may be used. A cache memory, a register, or the like may be provided. ] RAM, ROM, which is a memory, external storage devices such as hard disks, optical disks, semiconductor memories, etc., and the exchange of data between these input units, output units, communication units, CPU, RAM, ROM, external storage devices It has a bus that connects as possible. If necessary, the acoustic signal processing device may be provided with a device (drive) capable of reading and writing storage media such as a flexible disk, a CD-ROM (Compact Disc Read Only Memory), and a DVD (Digital Versatile Disc).

  The acoustic signal processing apparatus can be connected to an acoustic signal collecting means (for example, a microphone) that receives sound such as voice, music, and noise, and inputs an (analog) signal obtained by the microphone. A signal input unit for receiving and a sound output device (for example, a speaker) that outputs a reproduction signal as sound can be connected, and a signal for outputting a signal (a D / A converted version of the reproduction signal) input to the speaker A configuration in which an output unit is provided is also possible. In this case, a microphone is connected to the signal input unit, and a speaker is connected to the signal output unit.

  The external storage device of the acoustic signal processing device stores a program for detecting a voice section and data necessary for processing of the program [not limited to the external storage device, for example, the program is read using a read-only storage device. It may be stored in a certain ROM. ]. Further, data obtained by the processing of this program is appropriately stored in a RAM, an external storage device, or the like. Hereinafter, the storage means for storing the data, the address of the storage area, and the like will be simply referred to as “XX storage unit”.

  In this embodiment, an acoustic signal that is a discrete signal is stored in the main storage unit in order to acquire a signal in a section that is earlier in time series than the speech section included in the acoustic signal. This storage may be temporary storage such as a buffer.

<Configuration of acoustic signal processing apparatus>
FIG. 13 is a functional block diagram of the acoustic signal processing apparatus according to the third embodiment, and FIG. 14 shows the processing flow.

  The acoustic signal processing device includes a speech segment detection unit 320 and a speech segment detection information storage unit 330.

  The acoustic signal processing device receives the time-series acoustic signal collected by one microphone 310 and the output value of the specific speech section detection unit 340, and inputs the speech section and the non-speech section included in the time-series acoustic signal. At least one of them is detected and a detection result is output.

The specific voice section detection unit 340 detects the arrival of a predetermined sound (hereinafter also referred to as “specific sound”), and outputs information indicating the detection time of the specific sound. In the present embodiment, the specific sound is a predetermined sound uttered by a person, for example, a sound when a person utters a predetermined keyword. For example, the specific speech section detection unit 340 can be implemented using a technique such as “phrase spotting” as in Reference 1.
(Reference 1) “Sensory's Voice Technology Description” [online], 2010, [searched July 24, 2017], Internet <URL: http://www.sensory.co.jp/Parts/Docs /SensoryTechnologyJP1003B.pdf>
Note that the information indicating the detection time of the specific sound is information indicating at least the time when the specific sound (for example, a keyword) is finished, and (1-i) the time when the specific sound is finished may be output. (1-ii) The frame number of the time-series sound signal corresponding to the time when the specific sound is finished may be output, or (1-iii) it is detected at a frame time other than the time when the specific sound is finished. Information indicating that the specific sound has not been output (for example, “0”), and information indicating that the specific sound has been detected (for example, “1”) is output to indicate the time when the specific sound has been ended It may be information, or information indicating the time when the other specific sound is finished. Further, the information indicating the detection time of the specific sound may include information indicating the time when the specific sound is started, or (2-i) the time when the specific sound is started and the time when the specific sound is finished may be output. (2-ii) The time when the specific sound started and the frame number of the time-series sound signal corresponding to the time when the specific sound ended may be output, or (2-iii) the time when the specific sound started To output information (eg, “1”) indicating that it has been detected up to the time when it is finished, and to output information (eg, “0”) indicating that it has not been detected at other times. It may be information indicating the time when the user has finished speaking, or may be information indicating the time when the other specific sound is ended.

  Hereinafter, the processing content of each part is demonstrated.

<Audio section detection information storage unit 330>
The voice section detection information storage unit 330 receives information indicating the detection time of a specific sound and a time-series sound signal as inputs, and features a time-series sound signal corresponding to the specific sound voice section in units of frames and a non-voice section. The feature amount of the corresponding time-series acoustic signal is obtained (S330) and output. In each unit including the audio section detection information storage unit 330, each process is performed in units of frames.

  As shown in FIG. 15, the speech segment detection information storage unit 330 includes a speech storage unit 331, a specific sound speech segment calculation unit 332, and a feature amount calculation unit 333. Hereinafter, the processing content of each part is demonstrated.

(Voice storage unit 331)
The voice accumulation unit 331 receives and accumulates time-series acoustic signals to be detected as voice segments.

(Specific sound voice section calculation unit 332)
The specific sound speech section calculation unit 332 receives information indicating the detection time of the specific sound as an input, sets a section of a time-series acoustic signal estimated as a section corresponding to the specific sound based on the detection time as a specific sound sound section, and detects the detection time The time-series acoustic signal section estimated not to be a section corresponding to the specific sound is determined as a non-voice section, and information indicating the specific sound voice section and information indicating the non-voice section are output. For example, t ₁ second before the specific sound detection time (in this example, the time when the specific sound is finished) is defined as the specific sound voice section, and t ₂ seconds before the specific sound voice section is determined as the non-voice section. (See FIG. 16).

For example, if the information indicating the detection time of the specific sound includes only information indicating the frame time when the specific sound is finished (for example, t), t ₁ and t ₂ are set to predetermined values in advance. Then, the specific sound speech section (from tt ₁ to t) and the non-speech section (from tt ₁ -t ₂ to tt ₁ ) are obtained from the information indicating the detection time of the specific sound. As t ₁ , an average value of time taken when a specific sound is emitted may be used. In addition, when the information indicating the specific sound detection time includes information indicating the time when the specific sound is started and the time when the specific sound is ended (for example, t), the time when the specific sound is started is defined as tt ₁ The sound voice section is defined as from the time tt _{1 at} which the specific sound starts to the time t at which the specific sound ends. In addition, t ₂ is set to a predetermined value in advance, and a non-speech interval (from tt ₁ -t ₂ to tt ₁ ) is obtained from the predetermined value t ₂ and the time tt _{1 at} which the specific sound is started.

(Feature amount calculation unit 333)
The feature amount calculation unit 333 receives information indicating the specific sound voice interval and information indicating the non-voice interval from the specific sound voice interval calculation unit 332, and receives the time-series acoustic signal to be detected as the voice interval stored in the voice storage unit 331. receive. Then, the feature amount calculation unit 333 associates the time-series acoustic signal with the specific sound speech section, associates the time-series sound signal with the non-speech section, and features the time-series sound signal corresponding to the specific sound speech section. A speech segment feature value that is a quantity is calculated, a non-speech segment feature value that is a feature value is calculated from a time-series acoustic signal corresponding to the non-speech segment, and a speech segment feature value and a non-speech segment feature value are output. As the feature amount, for example, a log mel spectrum or a cepstrum coefficient can be used. However, it is good to set it as acoustic feature-values other than the acoustic feature-value (fundamental frequency) which the 2nd acoustic signal analysis part 322 uses. Any method may be used as the feature amount calculation method. For example, the method described in Reference 4 is used.
(Reference 4) JP 2009-63700 A

<Audio section detection unit 320>
The voice segment detection unit 320 receives a time-series acoustic signal from the microphone 310 and receives a voice segment feature value and a non-speech segment feature value from the feature value calculation unit 333. The speech segment detection unit 320 obtains a speech parameter indicating the feature of the speech segment from the speech segment feature value, obtains a non-speech parameter indicating the feature of the non-speech segment from the non-speech segment feature value, and obtains the speech parameter and the non-speech parameter. Using it, at least one of a speech section and a non-speech section is detected from the time-series acoustic signal (S320), and a detection result is output.

  For example, the speech segment detection unit 320 obtains a speech parameter, which is a parameter of an acoustic model used when estimating a speech segment, from the speech segment feature, and is a parameter of the acoustic model used when estimating a non-speech segment. The non-speech parameter is obtained from the non-speech segment feature.

  For example, the speech segment detection device of Reference 4 can be used for the speech segment detection unit 320. In this case, the voice parameter is a parameter of the voice GMM, and the non-voice parameter is a parameter of the non-voice GMM.

  As shown in FIG. 17, the speech section detection unit 320 includes a first acoustic signal analysis unit 321 that performs probability calculation on an input time-series acoustic signal using a parallel Kalman filter / parallel Kalman smoother, and a time-series acoustic signal. A second acoustic signal analysis unit 322 that performs probability calculation using the ratio of the periodic component to the non-periodic component, a weight calculation unit 323 that calculates the weight of each probability, and the calculated weight to A speech state / non-speech state synthesis probability ratio calculation unit 324 that calculates a synthesis probability that a sequence acoustic signal belongs to a speech state and a synthesis probability that belongs to a non-speech state and obtains a ratio between them, and a speech state / non-speech state synthesis probability ratio And a speech segment estimation unit 325 that performs speech / non-speech discrimination based on. In addition, about the structure other than the 1st acoustic signal analysis part 321, since it performs the process similar to the reference document 4, description is abbreviate | omitted.

  The time-series acoustic signal input to the first acoustic signal analysis unit 321 is an acoustic signal that is sampled at a sampling rate of, for example, 8,000 Hz and converted into a discrete signal. This acoustic signal is a sound in which a noise signal is superimposed on an audio signal that is a target signal. Hereinafter, an acoustic signal is referred to as an “input signal”, an audio signal is referred to as “clean audio”, and a noise signal is referred to as “noise”.

  The speech segment detection unit 320 receives the input signal, speech segment feature value, and non-speech segment feature value, and outputs a speech segment detection result. The speech section detection result is 1 if the acoustic signal in units of frames belongs to the speech state, and takes 0 if it belongs to the non-speech state. The voice segment detection unit 320 may output a signal obtained by multiplying the value of the voice segment detection result by the input signal. In other words, the value of the input signal of the frame belonging to the voice state is retained, and all the signal values are replaced with 0 in the frame belonging to the non-voice state.

<First acoustic signal analysis unit 321>
As shown in FIG. 18, the first acoustic signal analysis unit 321 receives an input signal, a speech segment feature value, and a non-speech segment feature value, and extracts a feature value calculation unit for extracting an acoustic feature value used for speech segment detection. 3211 and a probability estimation unit 3212 for estimating a probability model parameter and performing a probability calculation of an input signal using a probability model constituted by the obtained probability model parameter.

(Feature amount calculation unit 3211)
The feature amount calculation unit 3211 calculates the feature amount from the input signal and outputs it by the same method as the feature amount calculation unit 333. For example, a vector G _t = {gt _{, 0} ,..., Gt _{, φ} ,..., Gt _{, 23} } having a 24-dimensional log mel spectrum as an element is calculated and output. The vector G _t represents the acoustic feature amount in the frame whose start point is t. φ indicates the element number of the vector. Hereinafter, t is referred to as a frame time.

(Probability estimation unit 3212)
The 24-dimensional log mel spectrum, which is the output of the feature quantity calculation unit 3211, is input to the probability estimation unit 3212. The probability estimation unit 3212 applies a parallel nonlinear Kalman filter and a parallel Kalman smoother to the input frame to estimate a noise parameter. Using the estimated noise parameters, probabilistic models of non-speech (noise + silence) and speech (noise + clean speech) are generated, and the probability when the log mel spectrum is input to each probability model is calculated.

  As shown in FIG. 19, the probability estimation unit 3212 includes a forward estimation unit 3212-1, a backward estimation unit 3212-2, a GMM (Gaussian Mixture Model) storage unit 3212-3, and a parameter storage unit 3212-4. Note that the backward estimation unit 3212-2 performs the same processing as in the reference document 4, and thus the description thereof is omitted.

The GMM storage unit 3212-3 stores a silence GMM and a clean sound GMM, which are acoustic models of a silence signal and a clean sound signal prepared in advance. Hereinafter, the silent GMM and the clean voice GMM are simply referred to as GMM or the like. Since the GMM configuration method is a known technique, a description thereof will be omitted. Each GMM contains a plurality of normal distributions (for example, 32), and each normal distribution has a mixture weight w _{j, k} , mean μ _{S, j, k, φ} , variance Σ _{S, j, k, φ} Where j is the GMM type (j = 0: silent GMM, j = 1: clean speech GMM), and k is the number of each normal distribution. Each parameter becomes an input to the forward estimation unit 3212-1 and the backward estimation unit 3212-2.

  The parameter storage unit 3212-4 includes an initial noise model estimation buffer and a noise model estimation buffer.

[Forward estimation unit 3212-1]
The processing content in the forward estimation unit 3212-1 is different from that in Reference Document 4.

In Reference 4, the forward estimation unit obtains the noise model parameters ^ N _{t, j, k, φ} and ^ Σ _{N, t, j, k, φ} by sequential updating from the processing start time. The parameters of the non-speech / speech GMM are updated without determining whether the sound is voice or non-speech (noise). On the other hand, in the present embodiment, since the non-speech section and the speech section are known, the information is updated more actively to update the parameters. That is, the parameters of the non-speech GMM are updated using the speech feature amount of the non-speech segment, and the parameters of the speech GMM are updated using the speech feature amount of the speech segment. A processing example is shown below.

First, the forward estimation unit 3212-1 uses the feature quantities g _{t-t_1 -t_2, φ 1} ,..., G _{t-t_1, φ} from the frame times tt ₁ -t ₂ to tt ₁ corresponding to the non-speech section. The non-voice GMM (j = 0) parameter is updated. However, subscripts t_1 and t_2 mean t ₁ and t ₂ , respectively.

The forward estimation unit 3212-1 stores q speech non-speech segment feature amount g _t− out of non-speech segment feature amount (in this example, log mel spectrum g _{t, φ} ) in the initial noise model estimation buffer. _{t_1-t_2, φ} ,..., g _{t-t_1-t_2-1 + q-1, φ} are stored. However, q is an integer of 1 or more that does not exceed the length t ₂ of the non-speech interval, for example, q = 10.

The forward estimation unit 3212-1 extracts q frame feature quantities g _{t-t — 1 — t — 2} ,..., G _{t — t — 1 — t — 2 + q−1, φ} from the initial noise model estimation buffer. The initial noise model parameters N ^init _φ and Σ ^init _{N, φ} are estimated by the following equations and stored in the noise model estimation buffer.

Also, using the feature quantities g _{t-t_1-t_2 + q, φ} ,..., G _{t-t_1, φ} from the frame times tt ₁ -t ₂ + q to tt ₁ , the non-voice GMM (j = 0) Update parameters. The non-voice GMM parameter update method and update formula are the same as in Reference Document 4.

Next, the forward estimation unit 3212-1 uses the feature values g _{t-t_1 + 1, φ} ,..., G _{t, φ} from the frame times tt ₁ +1 to t corresponding to the speech section, to generate the speech GMM ( Update the parameter of j = 1). Note that the parameter updated using the last frame of the non-speech segment is the first parameter of the speech segment. That means

And Further, the parameters of the speech GMM (j = 1) are updated using the feature quantities g _{t-t_1 + 1, φ} ,..., G _{t, φ} . Note that the method and formula for updating the parameters of the voice GMM are the same as in Reference Document 4.

  Note that after the frame time t, the parameters of the voice / non-voice GMM are updated using the feature amount of the input signal, as in the prior art.

  The speech segment detection unit 320 uses the frame time t based on the non-speech GMM parameter updated using the speech feature value of the non-speech segment and the speech GMM parameter updated using the speech feature value of the speech segment. Thereafter, the voice / non-voice GMM parameters are updated using the feature amount of the input signal, and the voice / non-voice is determined using the parameters obtained as a result. Therefore, the determination accuracy can be improved as compared with the prior art in which the parameters of the non-voice / voice GMM are updated without determining whether the voice or non-voice (noise).

Note that the above-described processing may be performed only when the speech segment feature amount and the non-speech segment feature amount are first received from the feature amount calculation unit 333, or the speech segment feature amount and the non-speech function from the feature amount calculation unit 333. You may carry out whenever it receives an area feature-value. In addition, every time the speech section feature quantity and the non-speech section feature quantity are received from the feature quantity calculation unit 333, (a) a process for ^obtaining initial noise model parameters N ^init _φ and Σ ^init _{N, φ} ( b) All the processes including the process of using the parameter updated using the last frame of the non-speech section as the first parameter of the speech section may be repeated, and in the second and subsequent processes, the above (a) and Features from the frame times tt ₁ -t ₂ to tt ₁ corresponding to the non-speech segment using the parameters at the time of receiving the speech segment feature and the non-speech segment feature without performing the process of (b) Update the parameters of the non-speech GMM (j = 0) using the quantities g _{t-t_1-t_2, φ} , ..., g _{t-t_1, φ,} and the frame times tt ₁ +1 to t corresponding to the speech interval The parameters of the speech GMM (j = 1) may be updated using the feature quantities g _{t-t_1, φ} ,..., G _{t, φ} .

<Effect>
With the above configuration, using the result of keyword detection for a specific utterance of the target person (user), information about the surrounding acoustic environment including the target voice can be known more accurately, and voice segment detection The signal processing becomes robust. In particular, even when the speech to be recognized and the noise are close to each other, it is possible to detect at least one of the speech segment and the non-speech segment with higher accuracy than in the past.

  Note that one microphone 310 and the specific voice section detection unit 340 may be part of the acoustic signal processing device. In this embodiment, the GMM is used as the acoustic model used when estimating the speech section and the non-speech section. However, another acoustic model such as an HMM (Hidden Markov Model) may be used. Even in this case, as in the present embodiment, the speech parameter and the non-speech parameter may be obtained from the speech segment feature value and the non-speech segment feature value, respectively.

<First Modification of Third Embodiment>
A description will be given centering on differences from the third embodiment.

  In the third embodiment, a log mel spectrum, a cepstrum coefficient, or the like is used as the feature quantity, but other feature quantities may be used. In this modification, the case where the level of sound is used for determination will be considered more simply.

  In the present embodiment, average power is used as the feature amount. Therefore, the feature amount calculation unit 333 calculates the average power from the time-series acoustic signal corresponding to the specific sound speech section and outputs the average power as the speech section feature amount, and calculates the average power from the time-series acoustic signal corresponding to the non-speech section. Calculate and output as non-speech segment feature.

<Audio section detection unit 320>
The voice segment detection unit 320 receives the time-series acoustic signal to be detected by the voice segment stored in the voice storage unit 331, and receives the voice segment feature value and the non-speech segment feature value from the feature value calculation unit 333. The speech segment detection unit 320 obtains a speech parameter indicating the feature of the speech segment from the speech segment feature value, obtains a non-speech parameter indicating the feature of the non-speech segment from the non-speech segment feature value, and obtains the speech parameter and the non-speech parameter. Using it, at least one of a speech section and a non-speech section is detected from the time-series acoustic signal (S320), and a detection result is output.

  As shown in FIG. 20, the voice section detection unit 320 includes a voice power calculation unit 326, a voice / non-voice determination unit 327, a non-voice level storage unit 328, and a voice level storage unit 329.

  The voice power calculation unit 326 receives the time-series acoustic signal to be detected in the voice section accumulated in the voice accumulation unit 331, calculates the average power P (n) for each frame n of the time-series acoustic signal, and outputs it.

For example,
P (n)> γV and P (n)> δN
If the condition is satisfied, a method of determining the section as a speech section is conceivable. n is an index representing a frame time, N and V are non-voice level power thresholds stored in the non-voice level storage unit 328 and the voice level storage unit 329, power thresholds of the voice period, γ is 0 or more and 1 or less, δ is a real number of 1 or more. If it is greater than a value (γV) close to the level of the signal in the speech segment to some extent and greater than a value (ΔN) sufficiently greater than the level of the signal in the non-speech segment (for example, noise), it is determined that it is a speech segment. In this case, it does not operate correctly when the non-speech and speech information (V, N) stored in advance differs from the actual speech segment and non-speech segment signal levels. It is also possible to update each information (V, N) sequentially according to the time-series acoustic signal, but the value is updated in the wrong direction because it is updated without knowing which section is non-speech or speech. There is a risk of being.

  In the present embodiment, the power thresholds V and N are changed using the speech segment feature value (average power of the speech segment) and the non-speech segment feature value (average power of the non-speech segment).

  The voice / non-voice determination unit 327 extracts the power thresholds V and N from the non-voice level storage unit 328 and the voice level storage unit 329, receives the average power P (n) from the voice power calculation unit 326, and receives the feature amount calculation unit. From 333, the average power Pv of the time-series acoustic signal corresponding to the specific sound speech section and the average power Pn of the time-series acoustic signal corresponding to the non-speech section are received.

The voice / non-voice determination unit 327 replaces the power thresholds V and N with the power thresholds V ′ and N ′ considering the average powers Pv and Pn, respectively, by the following equations.
N '= (1-α) N + αPn
V '= (1-β) V + βPv
Α and β represent parameters (0 <α <1, 0 <β <1) for determining the contribution ratio of the detected speech / non-speech interval. The voice / non-voice determination unit 327
P (n)> γV 'and P (n)>δN'
If the condition is satisfied, the section corresponding to the frame n is detected as a speech section. If not satisfied, the section corresponding to the frame n is detected as a non-speech section, and the detection result is output.

  In the present embodiment, V ′ corresponds to a speech parameter indicating a feature of a speech segment, and N ′ corresponds to a non-speech parameter indicating a feature of a non-speech segment.

<Effect>
With the above configuration, level determination can be performed in accordance with the actual situation, and the same effect as in the third embodiment can be obtained.

<Second Modification of Third Embodiment>
A description will be given centering on differences from the third embodiment.

  FIG. 13 is a functional block diagram of the acoustic signal processing apparatus according to the third embodiment, and FIG. 14 shows the processing flow.

  The acoustic signal processing device includes a speech segment detection unit 320, a speech segment detection information storage unit 330, and a preprocessing unit 350.

<Pre-processing unit 350>
The pre-processing unit 350 receives the time-series acoustic signal as input, performs a process (speech enhancement process) for enhancing speech included in the time-series acoustic signal (S350), and outputs the enhanced time-series acoustic signal. Any method may be used as the speech enhancement processing. For example, the noise suppression method described in Reference 2 is used.
(Reference Document 2) Japanese Patent Laid-Open No. 2009-11001

<Effect>
With the above configuration, the same effect as that of the third embodiment can be obtained. Furthermore, the detection accuracy can be improved by performing subsequent processing (S330, S320) using the time-series acoustic signal subjected to the speech enhancement processing.

<Third Modification of Third Embodiment>
A description will be given centering on differences from the third embodiment.

  The acoustic signal processing apparatus includes M time-series acoustic signals respectively collected by M microphones 310-m (m = 1, 2,..., M, and M is an integer of 2 or more). , L (L is any integer greater than or equal to 2) output values of the specific speech section detection unit 340 is input, and at least one of speech sections and non-speech sections included in the time-series acoustic signal is detected. The detection result is output.

  FIG. 21 is a functional block diagram of the acoustic signal processing apparatus according to the third modification, and FIG. 22 shows the processing flow.

  The acoustic signal processing device includes a beam forming unit 360, a speech segment detection unit 320, and a speech segment detection information storage unit 330.

<Beam forming unit 360>
The beam forming unit 360 receives M time-series acoustic signals as input, and M time-series acoustic signals are L time-series signals (time-series acoustic signals that have increased directivity in L directions, respectively. For example, it is converted into a beamforming output signal) (S360), and is output to the specific speech segment detection unit 340, the speech segment detection information storage unit 330, and the speech segment detection unit 320. For example, it is converted into L time-series beamforming output signals using a beamforming technique. Any method may be used as the beam forming technique. For example, the method described in Reference 3 is used.
(Reference 3) Japanese Patent Application Laid-Open No. 2017-107141

  The specific speech section detection unit 340 detects that a specific sound has arrived for each of the L time-series signals, and outputs information indicating the detection time of the specific sound to the speech section detection information storage unit 330. Note that it is detected that a specific sound has arrived in at least one time-series signal of L time-series signals, and information indicating the detection time of the specific sound is information indicating one or more detected channels. And information indicating detection times of one or more specific sounds respectively corresponding to the detected one or more channels. Information indicating the detection time of each specific sound is as described in the third embodiment.

<Audio section detection information storage unit 330>
The voice segment detection information storage unit 330 receives information indicating the detection time of a specific sound and L time-series signals as input, and obtains voice segment feature values and non-speech segment feature values of a channel in which the specific sound is detected. (S330) and output. It should be noted that feature amounts are obtained for all channels in which the specific sound is detected.

<Audio section detection unit 320>
The speech segment detection unit 320 receives L time-series signals, and receives the speech segment feature amount and the non-speech segment feature amount of the channel in which the specific sound is detected from the feature amount calculation unit 333. The voice section detection unit 320 obtains one voice parameter indicating the characteristics of the voice section from the voice section feature values of all the channels in which the specific sound is detected, and the non-voice section feature values of all the channels in which the specific sound is detected. One non-speech parameter indicating the characteristics of the speech segment is obtained, and at least one of the speech segment and the non-speech segment is detected from each of the L time-series signals using the speech parameter and the non-speech parameter (S320). The detection result is output. The detection method is as described in the third embodiment. In this modification, one (common) speech parameter and one (common) non-speech parameter are used for L time-series signals.

<Effect>
With such a configuration, the same effect as that of the third embodiment can be obtained. The beam forming unit 360 may be a separate device, and the acoustic signal processing device may be configured to receive L time-series signals. In addition, each of the L directional microphones 310-m (m = 1, 2,..., L, where L is an integer greater than or equal to 2) with increased directivity in the L directions, respectively, collects sound. A configuration may be adopted in which the L time-series acoustic signals are input and the beam forming unit 360 is not used.

<Fourth Modification of Third Embodiment>
A description will be given centering on differences from the third modification.

<Audio section detection unit 320>
The speech segment detection unit 320 receives L time-series signals, and receives the speech segment feature amount and the non-speech segment feature amount of the channel in which the specific sound is detected from the feature amount calculation unit 333. The speech segment detection unit 320 obtains one speech parameter indicating the feature of the speech segment from the speech segment feature value of one channel where the specific sound is detected, and the non-speech segment feature value of one channel where the specific sound is detected 1 non-speech parameter indicating the characteristics of the non-speech interval is obtained from the time-series acoustic signal from which the specific sound is detected using the sound parameter and the non-speech parameter obtained for each channel in which the specific sound is detected. At least one of the section and the non-voice section is detected (S320), and the detection result is output. The detection method is as described in the third embodiment.

  In this modification, L speech parameters and L non-speech parameters respectively corresponding to L time-series signals are used. Note that the speech segment detection unit 320 receives the speech segment feature amount and the non-speech segment feature amount of the channel in which the specific sound is detected, and obtains only the non-speech parameter and the speech parameter of the channel. For a channel for which a specific sound has not been detected, the non-speech parameter and the speech parameter are not obtained, but the non-speech parameter and the speech parameter corresponding to the channel are obtained at the timing when the specific sound is detected.

<Effect>
With such a configuration, it is possible to obtain the same effects as in the third embodiment, and to obtain detailed audio parameters and non-audio parameters for each channel.

[Supplement]
The acoustic signal processing device receives an acoustic signal including a specific sound, which is a predetermined sound, as an input, and an acoustic signal obtained by removing an acoustic signal corresponding to the specific sound from the acoustic signal as a noise acoustic signal. And an acoustic signal processing unit that performs acoustic signal processing in association with the acoustic signal corresponding to the specific sound.

  Alternatively, the acoustic signal processing device receives an acoustic signal including a specific sound that is a predetermined sound as an input, uses the acoustic signal corresponding to the specific sound as a target acoustic signal, and uses the target acoustic signal and the acoustic signal as described above. It can be said that the apparatus includes an acoustic signal processing unit that performs acoustic signal processing associated with acoustic signals excluding the target acoustic signal.

  Alternatively, the acoustic signal processing device receives an acoustic signal including a specific sound, which is a predetermined sound, and an acoustic signal obtained by removing an acoustic signal corresponding to the specific sound from the acoustic signal as a noise acoustic signal. It can be said that an acoustic signal processing unit that performs acoustic signal processing in which the target acoustic signal is associated with the noise acoustic signal using the acoustic signal corresponding to the sound as the target acoustic signal is provided.

  An example of the acoustic signal processing unit is the third directional sound collecting unit 52 of Modification 4 of the first embodiment. In this case, the target acoustic signal is a sound signal from the arrival direction read from the direction storage unit 13, and the noise acoustic signal is a sound signal from the arrival direction read from the noise direction storage unit 51.

  Other examples of the acoustic signal processing unit are the speech segment detection information storage unit 330 and the speech segment detection unit 320 of the third embodiment. In this case, the target acoustic signal is a time-series acoustic signal corresponding to the specific sound speech section, and the noise acoustic signal is a time-series acoustic signal corresponding to the non-speech section.

[Program and recording medium]
When the processing in each unit of each acoustic signal processing device is realized by a computer, the processing contents of the functions that each unit of these devices should have are described by a program. Then, by executing this program on a computer, the processing of each part is realized on the computer.

  The program describing the processing contents can be recorded on a computer-readable recording medium. As the computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used.

  The processing of each unit may be configured by executing a predetermined program on a computer, or at least a part of these processing may be realized by hardware.

  Needless to say, other modifications are possible without departing from the spirit of the present invention.

Claims (9)

The input is a time-series acoustic signal including a specific sound that is a predetermined sound emitted by a person, and information indicating a detection time of the specific sound included in the time-series acoustic signal,
The time series sound signal section estimated as the section corresponding to the specific sound based on the detection time is a specific sound voice section, and the time is estimated not to be a section corresponding to the specific sound based on the detection time A specific sound speech section calculating unit that determines a section of a sequence acoustic signal as a non-speech section;
A speech segment feature amount that is a feature amount is calculated from a time-series acoustic signal corresponding to the specific sound speech segment, and a non-speech segment feature amount that is a feature amount is calculated from a time-series acoustic signal corresponding to the non-speech segment. A feature amount calculation unit to
Obtaining a speech parameter indicating a feature of a speech segment from the speech segment feature value, obtaining a non-speech parameter indicating a feature of a non-speech segment from the non-speech segment feature value, and using the speech parameter and the non-speech parameter Including a speech segment detection unit that detects at least one of a speech segment and a non-speech segment from a time-series acoustic signal,
Acoustic signal processing device. The acoustic signal processing device according to claim 1,
The specific sound voice interval calculation unit determines t1 seconds before the detection time as a specific sound voice interval, and determines t2 seconds before the specific sound voice interval as a non-voice interval,
Acoustic signal processing device. The acoustic signal processing device according to claim 1 or 2,
The speech section detection unit obtains a speech parameter, which is a parameter of an acoustic model used when estimating a speech section, from the speech section feature amount, and is a parameter of an acoustic model used when estimating a non-speech section. A speech parameter is obtained from the non-speech segment feature.
Acoustic signal processing device. The acoustic signal processing device according to claim 1 or 2,
The feature amount is an average power, the speech parameter is a power threshold considering the average power Pv of the time-series acoustic signal corresponding to the specific sound speech section, and the non-speech parameter is the time-series acoustic signal corresponding to the non-speech section. It is a power threshold considering the average power Pn,
The voice interval detection unit is based on the magnitude relationship between the voice parameter and the average power of the time-series acoustic signal, and the magnitude relationship between the non-voice parameter and the average power of the time-series acoustic signal. Detect at least one of non-speech intervals;
Acoustic signal processing device. The acoustic signal processing device according to any one of claims 1 to 4,
Including a pre-processing unit that emphasizes speech included in the time-series acoustic signal,
Acoustic signal processing device. The acoustic signal processing device according to any one of claims 1 to 5,
L time-series acoustic signals each having increased directivity in L directions including a specific sound that is a predetermined sound in at least one channel, and a detection time of the specific sound included in the time-series acoustic signal And the information indicating
The feature amount calculation unit calculates a speech section feature amount that is a feature amount from a time-series acoustic signal corresponding to the specific sound speech section of the channel in which the specific sound is detected, and the channel of the channel in which the specific sound is detected Calculate the feature amount of the non-speech segment that is the feature amount from the time-series acoustic signal corresponding to the non-speech segment,
The speech section detection unit obtains one speech parameter indicating a feature of a speech section from the speech section feature amount of all channels in which the specific sound is detected, and the non-speech section feature amount of all channels in which the specific sound is detected. Obtaining one non-speech parameter indicating the characteristics of the non-speech segment from, and detecting at least one of the speech segment and the non-speech segment from the time-series acoustic signal using the speech parameter and the non-speech parameter;
Acoustic signal processing device. The acoustic signal processing device according to any one of claims 1 to 5,
L time-series acoustic signals each having increased directivity in L directions including a specific sound that is a predetermined sound in at least one channel, and a detection time of the specific sound included in the time-series acoustic signal And the information indicating
The feature amount calculation unit calculates a speech section feature amount that is a feature amount from a time-series acoustic signal corresponding to the specific sound speech section of the channel in which the specific sound is detected, and the channel of the channel in which the specific sound is detected Calculate the feature amount of the non-speech segment that is the feature amount from the time-series acoustic signal corresponding to the non-speech segment,
The speech section detection unit obtains one speech parameter indicating a feature of a speech section from the speech section feature amount of one channel where a specific sound is detected, and the non-speech section of one channel where the specific sound is detected One non-speech parameter indicating a feature of a non-speech section is obtained from the feature amount, and the time-series sound of the channel in which the specific sound is detected using the sound parameter and the non-speech parameter of the channel in which the specific sound is detected Detecting at least one of a speech segment and a non-speech segment from the signal;
Acoustic signal processing device. The acoustic signal processing device uses a time-series acoustic signal including a specific sound that is a predetermined sound emitted by a person, and information indicating a detection time of the specific sound included in the time-series acoustic signal,
The time series sound signal section estimated as the section corresponding to the specific sound based on the detection time is a specific sound voice section, and the time is estimated not to be a section corresponding to the specific sound based on the detection time A specific sound speech section calculating step for determining a section of the sequence acoustic signal as a non-speech section;
A speech segment feature amount that is a feature amount is calculated from a time-series acoustic signal corresponding to the specific sound speech segment, and a non-speech segment feature amount that is a feature amount is calculated from a time-series acoustic signal corresponding to the non-speech segment. A feature amount calculating step,
Obtaining a speech parameter indicating a feature of a speech segment from the speech segment feature value, obtaining a non-speech parameter indicating a feature of a non-speech segment from the non-speech segment feature value, and using the speech parameter and the non-speech parameter A voice section detection step for detecting at least one of a voice section and a non-voice section from a time-series acoustic signal,
Acoustic signal processing method.   A program for causing a computer to function as the acoustic signal processing apparatus according to claim 1.

Images