JP2013125085A - Target sound extraction device and target sound extraction program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To prevent missing of target sound that may be caused by a difference in utterance speed, which cannot be covered only by long-term average processing, to further reduce sound interruption, when preventing missing of a small amplitude portion in a target sound section during voice-switch processing by subjecting a target sound detection result value to the long-term average processing.SOLUTION: A target sound extraction device calculates a coherence value on the basis of a plurality of directive signals each having a dead angle in a predetermined direction, calculates a long-term average value of a detection result value of a present frame by weighting average processing using a detection result value obtained by determining whether or not the present analysis value corresponds to a target sound section on the basis of the coherence value, and a value using the detection result value obtained in a past frame, and controls a gain for an inputted signal on the basis of the long-term average value. The device comprises weighting coefficient control means for controlling a weighting coefficient related to the weighting average processing on the basis of an utterance speed of target sound estimated from the coherence value.

Description

  The present invention relates to a target sound extraction device and a target sound extraction program, and can be applied to a voice communication device used for voice communication such as a telephone or a video conference.

  One technique for extracting a desired sound from an input signal is a technique called a voice switch. This is to detect the section where the speaker is speaking (target speech section) from the input signal using the target speech section detection function, output without processing for the target speech section, and amplitude for the non-target speech section. It is a process of attenuating.

  FIG. 2 is a flowchart showing voice switch processing. In FIG. 2, when the input signal input is received (S901), it is determined whether or not the target speech section detection unit is the target speech section (S902).

  At this time, if the input is the target voice section, the voice switch gain VS_GAIN is set to “1.0” (S903), and if the input is the non-target voice section, VS_GAIN is “α” (α: arbitrary) The positive value of 0.0 ≦ α <1.0 is set (S904). Then, VS_GAIN is multiplied by input to obtain an output signal output (S905).

  This voice switch process can be applied to, for example, a voice communication device such as a video conference apparatus and a mobile phone. By performing this voice switch process, a non-target voice section (noise) is suppressed, and a voice quality is improved. Can be increased.

  By the way, the non-target voice is divided into “interfering voice” which is a human voice other than the speaker and “background noise” such as office noise and road noise.

  When the non-target speech section is only background noise, the target speech section detection unit can accurately determine whether or not the target speech section is, while the non-target speech section has interference noise superimposed on it. Since the target voice section detection unit regards the disturbing voice as the target voice, an erroneous determination may occur. As a result, the voice switch cannot suppress the disturbing voice and cannot provide sufficient call sound quality.

  This problem can be improved by changing the input signal level that has been used so far to the coherence as the feature amount referred to by the target speech section detection unit.

  Here, the coherence is a feature quantity that means the arrival direction of the input signal, simply speaking. For example, assuming use of a mobile phone, the voice of the speaker (target voice) comes from the front, and the disturbing voice tends to come from other than the front. It is possible to distinguish between the target voice and the disturbing voice.

  FIG. 3 is a block diagram showing a functional configuration of the voice switch 90A when coherence is used for the target voice detection function.

  In FIG. 3, input signals s1 (n) and s2 (n) are given to the FFT unit 91 from the microphones m1 and m2 via an AD converter (not shown).

Note that n is an index indicating the input order of samples, and is expressed as a positive integer. In the text, it is assumed that the smaller n is the older input sample, and the larger n is the newer input sample.

  The FFT unit 91 receives the input signal series s1 and s2 from the microphone m1 and the microphone m2, and performs fast Fourier transform (or discrete Fourier transform) on the input signals s1 and s2. Thereby, the input signals s1 and s2 can be expressed in the frequency domain. In performing the Fast Fourier Transform, analysis frames FRAME1 (K) and FRAME2 (K) composed of predetermined N samples are configured from the input signals s1 (n) and s2 (n). An example of configuring FRAME1 from the input signal s1 will be described below.

FRAME1 (1) = {s1 (1), s1 (2), ..., s1 (i), ... s1 (N)}
・
・
FRAME1 (K) = {s1 (N × K + 1), s1 (N × K + 2), ..., s1 (N × K + i), ... s1 (N × K + N)}
K is an index indicating the order of frames, and is expressed as a positive integer. In the text, the smaller the K, the older the analysis frame, and the larger the K, the newer the analysis frame. In the following description of the operation, it is assumed that the index representing the latest analysis frame to be analyzed is K unless otherwise specified.

In the FFT unit 91, by performing fast Fourier transform processing for each analysis frame, the frequency domain signal X1 (f, K) obtained by Fourier transform to the analysis frame FRAME1 (K) configured from the input signal s1, and the input signal The frequency domain signal X2 (f, K) obtained by Fourier transforming the analysis frame FRAME2 (K) composed of s2 is given to the first directivity forming unit 92 and the second directivity forming unit 93. . Note that f is an index representing a frequency. X1 (f, K) is not a single value,
X1 (f, K) = {X1 (f1, K), X1 (f2, K), ... X1 (fi, K) ..., X1 (fm, K)}
Thus, it is supplemented that it is composed of spectral components of a plurality of frequencies f1 to fm. The same applies to X2 (f, K) and B1 (f, K) and B2 (f, K) appearing in the directivity forming section at the subsequent stage.

  The first directivity forming unit 92 receives the frequency domain signals X1 (f, K) and X2 (f, K) from the FFT unit 91, and receives a signal B1 (f, K) having strong directivity in a specific direction. The signal B 1 (f, K) is provided to the coherence calculator 94.

  The second directivity forming unit 93 receives the frequency domain signals X1 (f, K) and X2 (f, K) from the FFT unit 91, and receives a signal B2 (f, K) having strong directivity in a specific direction. Then, the signal B 2 (f, K) is given to the coherence calculator 94.

  Here, as a method of forming a signal having strong directivity in a specific direction by the first directivity forming unit 92 and the second directivity forming unit 93, a method of an existing technique can be applied. A method of obtaining by calculation according to (1) and equation (2) can be applied.

The first directivity forming unit 92 performs a calculation according to the equation (1), and obtains a signal B1 (f, K) having strong directivity in a specific direction (right direction) of the sound source direction as will be described later. The second directivity forming unit 93 performs calculation according to the equation (2), and calculates a signal B2 (f, K) having strong directivity in a specific direction (left direction) of the sound source direction, as will be described later. (The frame index K is not included in the calculation formula because it is not involved in the calculation).

  The meanings of Expression (1) and Expression (2) will be described with reference to FIGS. In FIG. 4A, it is assumed that the microphone m1 and the microphone m2 are separated by a distance l. Sound waves arrive at the microphones m1 and m2. This sound wave is assumed to come from the direction of the angle θ with respect to the front direction of the plane passing through the microphone m1 and the microphone m2.

  At this time, there is a time difference until the sound wave reaches the microphone m1 and the microphone m2. This arrival time difference τ is given by the equation (2-1) because d = 1 × sin θ, where d is the sound path difference.

τ = 1 × sin θ / c (c: speed of sound) (2-1)
By the way, it can be said that the signal s1 (n−τ) obtained by delaying the input signal s1 (n) by the arrival time difference τ is the same signal as s2 (n).

  Therefore, the signal y (n) = s2 (n) −s1 (n−τ) taking the difference between them is a signal from which the sound coming from the θ direction is removed. As a result, the microphone array has a directivity characteristic as shown in FIG.

  In the above description, the calculation in the time domain is described. However, the same effect can be obtained even if the calculation is performed in the frequency domain. Expressions (1) and (2) are examples of arithmetic expressions for the frequency domain.

  Here, when the arrival direction θ = 90 degrees, the directivity characteristics as shown in FIGS. 5A and 5B are obtained. For the directivity, forward, backward, right, and left directions are defined as shown in FIG. Then, as shown in FIG. 5A, the directivity formed by the first directivity forming unit 92 is strong in the left direction, and the directivity formed by the second directivity forming unit 93 is As shown in FIG. 5 (B), it becomes strong in the right direction.

  In the following description, for convenience of explanation, the operation will be described assuming that θ = 90 degrees. However, the present invention is not limited to this setting.

The signals B1 (f, K) and B2 (f, K) obtained as described above are given to the coherence calculator 94. The coherence calculation unit 94 obtains coherence COH by performing calculations according to the following equations (3) and (4) (again, since the frame index K is not involved in the calculation, it is not described in the equations).

  Next, the target speech segment detection unit 95 compares the coherence COH (K) with the target speech segment determination threshold Θ, and if the coherence COH (K) is larger than the target speech segment determination threshold Θ, the target speech segment detection unit 95 regards it as the target speech segment. If 1.0 is substituted into the storage variable VAD_RES (K) and the coherence COH (K) is smaller than the target speech segment determination threshold Θ, it is regarded as a non-target speech segment (interfering speech, background speech) and the detection result storage variable VAD_RES ( Substitute 0.0 for K).

  The gain control unit 96 sets the gain VS_GAIN to 1.0 if VAD_RES (K) = 1.0, and sets the gain VS_GAIN to an arbitrary value less than 1.0 if VAD_RES (K) = 0.0. Set to a positive number α.

  Here, the background of detecting the target speech section based on the level of coherence will be briefly described. The concept of coherence can be paraphrased as, for example, the correlation between a signal coming from the right direction in the front direction and a signal coming from the left direction.

  Therefore, the case where the coherence COH is small is a case where the correlation between the signal B1 and the signal B2 is small, and conversely, the case where the coherence COH is large can be paraphrased as a case where the correlation between the signals B1 and B2 is large.

  The input signal when the correlation is small is either when the input arrival direction is greatly deviated to either the right or left direction, or when the signal is clear and has little regularity such as background noise. is there.

  Therefore, it can be said that the section where the coherence COH is small is a disturbing voice section or a background noise section (non-target voice section).

  On the other hand, when the value of the coherence COH is large, it can be said that there is no deviation in the arrival direction, and therefore the input signal comes from the front. Now, since it is assumed that the target speech comes from the front, it can be said that it is the target speech section when the coherence COH is large.

  The VS_GAIN obtained as described above is multiplied by the signal s1 (n) by the voice switch gain multiplier 97 to obtain the output signal y (n).

  However, in the configuration of FIG. 3, in a small amplitude section such as the rising portion of the voice, even if there is a target voice, it is difficult to correlate without a clear pitch characteristic, so the coherence COH is reduced.

  As a result, as illustrated in FIG. 6A, even for the target voice, it is mistakenly determined as a disturbing voice in the small amplitude section of the rising portion and is attenuated by the voice switch processing, so that a loss occurs. There may be a problem that sound that is interrupted is output and sound quality becomes unnatural.

  In order to solve this problem, as illustrated in FIG. 7, there is a voice switch 90 </ b> B having a detection result long-term average unit 98 that performs a long-term averaging process on the target speech segment detection result.

  In the voice switch 90B shown in FIG. 7, the detection result long-term average unit 98 performs the long-term average processing on the detection result storage variable VAD_RES, and the voice switch 90B By controlling the switch, it is possible to suppress omission at the target sound small amplitude portion.

  For example, the detection result long-term average unit 98 obtains the detection result long-term average value VAD_RES_LONG (K) by the arithmetic expression illustrated in Expression (5). Then, the gain control unit 99 compares VAD_RES_LONG (K) with the voice switch operation determination threshold Ψ, and if VAD_RES_LONG (K) <Ψ, the voice switch gain VS_GAIN> α (0.0 ≦ α <1.0) Otherwise, control is performed such that VS_GAIN = 1.0.

As a result, since the voice switch can be operated after the fluctuation of the VAD_RES in the small amplitude portion of the target voice is reduced, the lack of the small amplitude portion of the target voice is suppressed as shown in FIG. can do.

  The long-term average parameter δ is 0.0 <δ <1.0. Here, the meaning of equation (5) is captured. Expression (5) is obtained by determining the determination value VAD_RES (K) for the input speech in the current frame section (Kth frame from the operation start time) and the long-term average value VAD_RES_LONG (K−1) obtained in the previous frame section. ) And the degree of contribution of the instantaneous value VAD_RES (K) to the average value can be adjusted by the magnitude of the value of δ.

  If δ is set to a small value close to 0, the contribution of the instantaneous value to the average value becomes small, so that fluctuations in VAD_RES can be suppressed. Also, if δ is a value close to 1, the contribution of the instantaneous value increases, so the long-term average effect can be weakened.

JP 2006-197552 A Japanese translation of PCT publication 2010-532879

  By the way, since coherence has the meaning of correlation of input signals, the behavior of coherence differs depending on whether it is a consonant or a vowel even within an incoming speech segment.

  For example, when “sa: sa” is spoken, the signal waveform of the consonant part “s” has low regularity, so the coherence is small, and the signal waveform of the vowel part “a” has high regularity, so the coherence is large. .

  In addition, when the speaking speed is changed, the length of the consonant part is not changed, but the length of the vowel part is changed. For example, when the utterance speed is changed when uttering “sa: sa”, when the utterance speed is slow, the consonant part “s” is not lengthened, but the vowel part “a” is lengthened and the utterance speed is fast. Sometimes the vowel part “a” is shortened.

  By the way, if the speech rate is slow, even if a small amplitude part such as a consonant part is misjudged as non-target speech, the proportion of the large amplitude part of the vowel part in the speech interval is high. Since the contribution of misjudgment to the average is reduced, the lack of small amplitude portions is unlikely to occur.

  However, when the utterance speed is high, the ratio of the large amplitude part of the vowel part in the speech section is reduced, and the contribution of the erroneous determination in the small amplitude part to the long-term average becomes large, so the fluctuation of VAD_RES cannot be reduced. In this case, a small amplitude portion is lost.

  Therefore, as described above, the conventional voice switch illustrated in FIG. 7 has a problem that a small amplitude portion of the target voice is lost even if the detection results are averaged over a long period depending on the speech speed.

  Therefore, in order to prevent the loss of the small amplitude part of the target speech by the long-term averaging process, the target sound extraction device and the purpose that can prevent the voice segment from being lost due to the difference in the speech speed and reduce the interruption of the speech There is a need for a sound extraction program.

  In order to solve this problem, the first aspect of the present invention includes (1) frequency analysis means for converting an input signal from the time domain to the frequency domain, and (2) based on the signal obtained by the frequency analysis means, Directivity forming means for forming a plurality of signals having directivity having a blind spot in a predetermined direction; and (3) coherence calculation means for obtaining a coherence value based on the plurality of directivity signals formed by the directivity forming means. And (4) target sound determination means for determining whether or not the target sound is included based on the coherence value obtained by the coherence calculation means, and outputting a detection result value corresponding to the determination result, and (5) purpose The detection result value calculated from the current input frame obtained from the sound determination means and the past detection result value are subjected to weighted averaging processing, and the long-term average value of the detection result value in the current input frame A long-term average processing means to be obtained; (6) an utterance speed detection means for detecting the utterance speed of the target sound included in the input signal based on the coherence value obtained by the coherence calculation means; and (7) an utterance speed detection means. A weighting factor control unit that controls a weighting factor related to the weighted averaging process of the long-term average processing unit according to the detected speech speed; and (8) a long-term average value of detection result values in the current input frame of the long-term average processing unit. And a gain control means for controlling the gain for the input signal, and (9) a multiplication means for multiplying the input signal by the gain controlled by the gain control means. It is an extraction device.

  According to a second aspect of the present invention, a computer is used to (1) frequency analysis means for converting an input signal from a time domain to a frequency domain, and (2) a blind spot in a predetermined direction based on a signal obtained by the frequency analysis means. Directivity forming means for forming a plurality of signals having directivity, (3) coherence calculation means for obtaining a coherence value based on the plurality of directivity signals formed by the directivity formation means, and (4) coherence calculation means. (5) A current input frame obtained from the target sound determination means, which determines whether or not the target sound is included based on the coherence value obtained by the above and outputs a detection result value corresponding to the determination result A long-term average processing means for performing a weighted average process on the detection result value calculated from the above and a past detection result value to obtain a long-term average value in the current input frame; (6) An utterance speed detecting means for detecting the utterance speed of the target sound included in the input signal based on the coherence value obtained by the coherence calculating means; (7) a long-term average according to the utterance speed detected by the utterance speed detecting means; Weight coefficient control means for controlling the weight coefficient related to the weighted average processing of the processing means; (8) controlling the gain for the input signal based on the long-term average value of the detection result values in the current input frame of the long-term average processing means; (9) A target sound extraction program that causes a gain controlled by the gain control means to function as a multiplying means that multiplies an input signal.

  According to the present invention, when a small amplitude portion of a voice section of a target voice is prevented from being lost, by controlling a long-term average parameter in a long-term average process according to the utterance speed, a voice that may be generated due to a difference in the utterance speed It is possible to prevent gaps from being lost by further preventing missing sections.

It is a block diagram which shows the structure of the voice switch of 1st Embodiment. It is a flowchart which shows the conventional voice switch process. It is a block diagram which shows the structure of the voice switch in the case of using coherence for the conventional target voice detection function. It is explanatory drawing explaining the mode of the sound wave arrival input into the microphone m1 and the microphone m2. It is explanatory drawing explaining the directional characteristic by the 1st directivity formation part and the 2nd directivity formation part. It is explanatory drawing explaining misjudgment with a non-target audio | voice in a target audio | voice area | region, and missing a target audio | voice area. It is a block diagram which shows the structure of the voice switch which prevents the omission of a small amplitude part by the long-term average of the conventional objective sound. It is an internal block diagram which shows the detailed internal structure of the long-term average parameter control part of 1st Embodiment. It is explanatory drawing explaining the corresponding | compatible table which matched speech rate v (K) and long-term average parameter (delta) of 1st Embodiment. It is a block diagram which shows the structure of the voice switch of 2nd Embodiment. It is a block diagram which shows the structure of the voice switch of 3rd Embodiment. It is explanatory drawing explaining the directional characteristic by the 3rd directivity formation part of 3rd Embodiment. It is a block diagram which shows the structure of the voice switch of 4th Embodiment. It is a block diagram which shows the structure of the voice switch of 5th Embodiment.

(A) First Embodiment Hereinafter, a first embodiment of a target sound extraction apparatus and a target sound extraction program of the present invention will be described in detail with reference to the drawings.

  In the first embodiment, an embodiment in which the present invention is applied to a voice switch is illustrated.

(A-1) Configuration of First Embodiment (A-1-1) Overall Configuration FIG. 1 is a configuration diagram showing a configuration of a voice switch 100A of the first embodiment. The voice switch 100A of the first embodiment has, for example, a CPU, ROM, RAM, EEPROM, input / output interface, and the like. The function of the voice switch 100A is a processing program stored in the ROM by the CPU. This is realized by executing. Note that the target sound extraction program may be installed through a network, and in that case also constitutes the components shown in FIG.

  In FIG. 1, a voice switch 100A according to the first embodiment includes a microphone m1 and a microphone m2, an FFT unit 101, a first directivity forming unit 102, a second directivity forming unit 103, a coherence calculating unit 104, an utterance speed. It has at least a detection unit 105, a long-term average parameter control unit 106, a target speech section detection unit 107, a detection result long-term average unit 108, a gain control unit 109, and a voice switch gain multiplication unit 110.

  The microphones m <b> 1 and m <b> 2 capture incoming sound waves, convert the captured sound waves into audio signals, and supply the sound signals to the FFT unit 101. Here, although not shown in FIG. 1, an AD conversion unit is provided between the microphones m1 and m2 and the FFT unit 101, and the AD conversion unit converts the audio signals (analog signals) of the microphones m1 and m2 into digital signals. The signal sequence s1 and the signal s2 are given to the FFT unit 101. Note that n indicates the input order of samples.

  The FFT unit 101 receives the input signals s1 and s2 from the microphone m1 and the microphone m2, and performs fast Fourier transform (or discrete Fourier transform) for each frame composed of a predetermined number of samples. Thereby, the input signal sequences s1 and s2 can be expressed in the frequency domain. In addition, the FFT unit 101 converts the frequency domain signal X1 (f, K) obtained from the input signal s1 and the frequency domain signal X2 (f, K) obtained from the input signal s2 into the first directivity forming unit 102 and the first directivity forming unit 102. The second directivity forming unit 103 is provided.

  The first directivity forming unit 102 receives the frequency domain signals X1 (f, K) and X2 (f, K) from the FFT unit 101, and receives a signal B1 (f, K) having strong directivity in a specific direction. The signal B 1 (f, K) is provided to the coherence calculator 14.

  The second directivity forming unit 103 receives the frequency domain signals X1 (f, K) and X2 (f, K) from the FFT unit 101, and has strong directivity in a specific direction different from that of the first directivity forming unit 102. A signal B 2 (f, K) having characteristics is formed, and the signal B 2 (f, K) is given to the coherence calculator 104.

  Here, the first directivity forming unit 102 and the second directivity forming unit 103 form, for example, equations (1) and (2) as a method of forming a signal having directivity having a blind spot in a specific direction. It is possible to apply a method of obtaining by calculation according to). Thereby, the first directivity forming unit 102 performs a calculation according to the equation (1), forms a signal B1 (f, K) having a strong directivity in the right direction, and the second directivity forming unit. 103 performs an operation according to the equation (2) to form a signal B2 (f, K) having strong directivity in the left direction.

  The coherence calculating unit 104 performs coherence COH based on the signal B1 (f, K) acquired from the first directivity forming unit 102 and the signal B2 (f, K) acquired from the second directivity forming unit 103. (K) is required. Further, the coherence calculation unit 104 gives the obtained coherence COH (K) to the utterance voice detection unit 105 and the target voice section detection unit 107.

  Note that various methods can be widely applied as a coherence calculation method by the coherence calculation unit 104. For example, a method in which the coherence calculation unit 104 uses Equation (3) and Equation (4) is used. Can do.

  The target speech detection unit 107 receives the coherence COH (K) from the coherence calculation unit 104, compares the coherence COH (K) with the target speech segment determination threshold value Θ, and the coherence COH (K) is obtained from the target speech segment determination threshold value Θ. If it is larger, it is determined that it is the target speech section, and if it is less than the target speech section determination threshold Θ, it is determined that it is a non-target speech section.

  The target speech detection unit 107 gives a detection result variable VAD_RES (K) indicating the determination result to the detection result long-term average unit 108. Specifically, VAD_RES (K) = 1.0 is set in the case of the target voice section, and VAD_RES (K) = 0.0 is set in the case of the non-target voice section.

  The speech rate detector 105 receives the coherence COH (K) obtained from the current input frame from the coherence calculator 104 and obtains the speech rate based on the coherence COH (K). Further, the speech rate detection unit 105 gives the detected speech rate v (K) to the long-term average parameter control unit 106.

  The long-term average parameter control unit 106 receives the utterance speed v (K) from the utterance speed detection unit 105, obtains the long-term average parameter δ according to the utterance speed v (K), and detects the long-term average parameter δ as a detection result long-term average unit. 108. The details of the long-term average parameter control method by the long-term average parameter control unit 106 will be described later.

  The detection result long-term average unit 108 receives the detection result variable VAD_RES (K) from the target speech segment detection unit 107 and the long-term average parameter δ from the long-term average parameter control unit 106, and performs long-term averaging on the detection result of the target speech segment. Processing is performed to obtain a long-term average value VAD_RES_LONG (K).

  Here, various methods can be applied to the long-term averaging process by the detection result long-term average unit 108 without particular limitation. For example, a method of obtaining using the arithmetic expression of the equation (5) is applied. Can do.

  The gain control unit 109 receives the long-term average value VAD_RES_LONG (K) from the detection result long-term average unit 108 and gives the gain value VS_GAIN corresponding to the long-term average value VAD_RES_LONG (K) to the voice switch gain multiplication unit 110.

  The voice switch gain multiplier 110 receives the gain value VS_GAIN from the gain controller 109, multiplies the input signal s1 (n) by the gain value VS_GAIN, and outputs a signal y (n).

(A-1-2) Detailed Configuration of Long-Term Average Parameter Control Unit FIG. 8 is an internal configuration diagram illustrating a detailed internal configuration of the long-term average parameter control unit 106 according to the first embodiment.

  In FIG. 8, the long-term average parameter control unit 106 according to the first embodiment includes at least an utterance speed input unit 201, a long-term average parameter collation unit 202, a storage unit 203, and a long-term average parameter output unit 204.

  The speech rate input unit 201 inputs the speech rate v (K) from the speech rate detection unit 105 and gives the input speech rate v (K) to the long-term average parameter matching unit 202.

  The storage unit 203 stores a correspondence table in which the speech rate v (K) is associated with the long-term average parameter δ (0.0 <δ <1.0).

  FIG. 9 is an explanatory diagram for explaining a correspondence table in which the speech rate v (K) is associated with the long-term average parameter δ. For example, in FIG. 9, when the utterance speed v (K) detected by the utterance speed detector 105 is x ≦ v (K) <w, the long-term average parameter δ is determined as δ = a. In FIG. 9, the speech speed v (K) has a relationship of... <Z <y <x <w, and the long-term average parameter δ has a relationship of a <b <c <. That is, the long-term average parameter δ decreases as the utterance speed v (K) decreases, and the long-term average parameter δ increases as the utterance speed v (K) increases. As a result, the higher the utterance speed v (K), the lower the contribution rate of VAD_RES in the current target speech segment, and the contribution of misjudgment to the long-term average can be reduced.

  The long-term average parameter matching unit 202 receives the utterance rate v (K) from the utterance rate input unit 201, refers to the correspondence table stored in the storage unit 203, and determines the long-term average parameter corresponding to the utterance rate v (K). δ (0.0 <δ <1.0) is obtained.

  In the first embodiment, the long-term average parameter determining method refers to a case where the long-term average parameter matching unit 202 obtains a long-term average parameter corresponding to the speech rate with reference to the correspondence table illustrated in FIG. However, it is not limited to this method.

  For example, the correspondence table stored in the storage unit 203 is not the correspondence table illustrated in FIG. 9. For example, a reference value of the utterance speed and a reference value of the long-term average parameter at the utterance speed are set and the utterance is set. A correspondence table in which the difference between the reference value of the speed and the input speech speed and the correction value of the long-term average parameter is associated is stored in the storage unit 203, and the long-term average parameter matching unit 202 refers to the correspondence table. Thus, the correction value of the long-term average parameter may be obtained according to the difference from the reference value of the speech rate, and the long-term average parameter δ may be obtained using the correction value and the reference value of the long-term average parameter.

  Further, for example, the storage unit 203 has a correspondence table in which the difference between the reference value of the utterance speed and the input utterance speed is associated with the value of the long-term average parameter, and the long-term average parameter matching unit 202 is associated with the correspondence table. , A long-term average parameter corresponding to the difference from the reference value of the speech rate may be obtained.

  As another method, for example, a long-term average parameter δ is calculated by creating a relational expression in which the value of the long-term average parameter δ decreases as the utterance speed decreases, and the input utterance speed is substituted into the relational expression. It may be. Thereby, the long-term average parameter according to the speech rate can be obtained with high accuracy.

  The long-term average parameter output unit 204 gives the long-term average parameter δ obtained by the long-term average parameter matching unit 202 to the detection result long-term average unit 108.

(A-2) Operation of the First Embodiment Next, the operation of the target sound extraction process in the voice switch 100 of the first embodiment will be described.

  When an audio signal is input to the microphone m1 and the microphone m2, it is converted into a digital signal by an AD converter (not shown), and the input signal series s1 and the signal s2 are given to the FFT unit 101.

  In the FFT unit 101, the signals s1 and s2 are formed in an analysis frame for each predetermined number of samples, fast Fourier transform is performed to convert from the time domain to the frequency domain, and the converted signals X1 (f, K) and X2 are converted. (F, K) is given to the first directivity forming unit 102 and the second directivity forming unit 103.

  When the signal X1 (f, K) and the signal X2 (f, K) are input, the first directivity forming unit 102 receives the input signal according to, for example, the arithmetic expressions of the expressions (1) and (2). Based on X1 (f, K) and the signal X2 (f, K), a signal B1 (f, K) having a specific orientation at the blind spot is formed.

  Similarly, the second directivity forming unit 103 has a directivity direction different from that of the first directivity forming unit 102. For example, according to the arithmetic expressions of the expressions (1) and (2), the signal X1 ( Based on f, K) and the signal X2 (f, K), a signal B2 (f, K) having a blind spot in a specific direction different from the first directivity forming unit 102 is formed.

  Then, when the signal B1 (f, K) and the signal B2 (f, K) each having a blind spot in a specific direction are supplied to the coherence calculation unit 104, the coherence calculation unit 104, for example, The coherence COH (K) is calculated based on the signal B1 (f, K) and the signal B2 (f, K) according to the arithmetic expression of 4).

  The target speech segment detection unit 107 compares the coherence COH (K) obtained by the coherence calculation unit 104 with the target speech segment determination threshold Θ, and when the coherence COH (K) is larger than the target speech segment determination threshold Θ, Assuming that the section is the target voice section, 1.0 is substituted into VAD_RES (K) and provided to the detection result long-term average unit 108. On the other hand, if the coherence COH (K) is equal to or smaller than the target speech segment determination threshold Θ, it is assumed that the segment is a non-target speech segment, and 0.0 is substituted for VAD_RES (K) to be provided to the detection result long-term average unit 108. .

  On the other hand, the coherence COH (K) obtained by the coherence calculation unit 104 is also given to the speech rate detection unit 105. The utterance speed detection unit 105 obtains the utterance speed v (K) according to the coherence COH (K).

  Here, as a method for detecting the speech rate by the speech rate detecting unit 105, various methods can be widely applied as long as the speech rate is obtained based on the coherence COH. For example, the speech rate detection unit 105 can detect the speech rate by the following method.

  For example, since coherence is a cross-correlation between two signals, there is a sound source in front of the microphone m1 and the microphone m2, and the coherence COH is larger than a signal input from the front. On the other hand, there are sound sources in the right direction or left direction of the microphones m1 and m2, and the coherence COH is small for signals input from the right direction or left direction.

  Even if the signal is from the front, the signal of the vowel part (for example, the voice part of “a” when pronounced “sa: sa”) is a highly correlated waveform having a certain degree of periodicity. While the coherence COH is large, the signal of the consonant part has a characteristic that the coherence COH is small because the signal of the consonant part has a weak periodicity and a low correlation.

  Furthermore, when the speech rate changes, the length of the consonant part does not change and the length of the vowel part changes. This is because the length of the consonant part of “sa: sa” does not change when the utterance speed slows down due to the human utterance mechanism, but the length of “a” in the vowel part becomes longer. As the speed increases, the length of the consonant part does not change, and the length of “a” in the vowel part becomes shorter.

  In addition, when the utterance speed is high, the coherence COH at the vowel part decreases rapidly, whereas when the utterance speed is low, the coherence COH at the vowel part decreases slowly. Becomes more prominent in sections where vowels such as double vowels continue.

  Therefore, the speech rate detection unit 105 obtains, for example, a difference between the coherence COH of the current frame section and the coherence COH of the immediately preceding frame section by using the characteristics of the coherence COH described above, and the difference in the coherence is large. In some cases, the utterance speed may be high, and conversely, when the difference in coherence is small, the utterance speed may be determined to be low and the utterance speed may be obtained.

  Specifically, the utterance speed detection unit 105 holds a correspondence table in which the difference in coherence and the utterance speed corresponding thereto are associated in advance, and the utterance speed detection unit 105 refers to the correspondence table to A method for obtaining the speech rate corresponding to the difference between the coherence COH (K) obtained from the previous frame and the coherence COH (K-1) obtained from the immediately preceding frame can be applied. Note that the method of obtaining the speech rate by the speech rate detection unit 105 is not limited to the above detection example.

  Next, the long-term average parameter control unit 106 obtains the long-term average parameter δ according to the utterance speed v (K) obtained by the utterance speed detection unit 105.

  In the long-term average parameter control unit 106, the long-term average parameter matching unit 202 receives the utterance speed v (K) input from the utterance speed input unit 201 and inputs it by referring to the correspondence table stored in the storage unit 203. The long-term average parameter δ corresponding to the utterance speed v (K) is acquired. Then, the long-term average parameter δ is given from the long-term average parameter output unit 204 to the detection result long-term average unit 108.

  In the detection result long-term average unit 108, VAD_RES (K) is given from the target speech section detecting unit 107, and the long-term average parameter control unit 106 is given the long-term average parameter δ. The long-term average value VAD_RES_LONG (K) is obtained according to the following equation.

  Then, the gain control unit 109 compares VAD_RES_LONG (K) with the voice switch operation determination threshold Ψ as in the conventional case, and when VAD_RES_LONG (K) is smaller than the voice switch operation determination threshold Ψ, the voice switch gain VS_GAIN = α ( 0.0 ≦ α <1.0), otherwise VS_GAIN = 1.0.

  Here, the long-term average parameter δ becomes a large value (ie, a value close to 1.0) as the utterance speed v (K) becomes fast, and becomes a small value (ie, 0 as the utterance speed v (K) becomes slow. Value close to .0).

  In Equation (5), when the speech rate is high, the contribution of VAD_RES (K) obtained in the current frame is reduced, and the contribution of VAD_RES_LONG (K−1) in the immediately preceding frame interval is increased. Means that Thereby, when the speech speed is high, it is possible to reduce the contribution to the long-term average value of the misjudgment that occurs in the small amplitude part in the target speech section. Therefore, it is possible to increase the possibility that VAD_RES_LONG (K) becomes larger than the determination threshold value Ψ, and thus it is possible to prevent the target voice from being lost.

  Further, when the speech rate v (K) is slow, the contribution of VAD_RES (K) is increased and the contribution of the long-term average value VAD_RES_LONG (K-1) is decreased compared to the case where the speech rate is fast. . This is because when the utterance speed is slow, the proportion of vowels in the target speech segment is high, so the rate of misjudgment is small. This is a process considering that it is effective for prevention. As described above, since the long-term average parameter δ is appropriately controlled even when the speech rate is low, it is possible to prevent the target speech from being lost.

  Then, the voice switch gain multiplication unit 110 multiplies the input signal s1 (n) by VS_GAIN from the gain control unit 109 to create and output an output signal y (n).

(A-3) Effect of the First Embodiment As described above, according to the first embodiment, even when the speech speed changes, it is possible to prevent the target voice from being lost, so that the sound quality is deteriorated. Can be resolved.

  Thereby, for example, by applying the present invention to a communication apparatus such as a video conference system or a mobile phone, it is possible to expect improvement in call sound quality.

(B) Second Embodiment Next, a second embodiment of the target sound extraction apparatus and target sound extraction program of the present invention will be described with reference to the drawings.

(B-1) Configuration and Operation of Second Embodiment FIG. 10 is a configuration diagram showing the configuration of the voice switch 100B of the second embodiment. In FIG. 10, the voice switch 100B according to the second embodiment includes a microphone m1 and a microphone m2, an FFT unit 101, a first directivity forming unit 102, a second directivity forming unit 103, a coherence calculating unit 104, an utterance speed. Detection unit 105, long-term average parameter control unit 106, target voice section detection unit 107, detection result long-term average unit 108, gain control unit 109, voice switch gain multiplication unit 110, non-target voice section monitoring unit 301, long-term average value initialization It has at least part 302.

  The second embodiment is different from the first embodiment in that in addition to the components of the first embodiment, a non-target voice section monitoring unit 301 and a long-term average value initialization unit 302 are further provided.

  In the first embodiment, the long-term average parameter δ is controlled according to the utterance speed. However, when the contribution rate of the current VAD_RES (K) is reduced, it becomes impossible to accurately respond to the start of the target speech section. When switching from a non-target voice section to a target voice section, etc., although it is originally a target voice section, it is misjudged as a non-target voice section by long-term averaging processing, and the head of the talk is missing at the voice switch Can occur.

  Therefore, in the second embodiment, the non-target speech section monitoring unit 301 and the long-term average value initialization unit 302 are provided in the configuration of the first embodiment, thereby preventing the beginning of the talk from being lost.

  In FIG. 10, the same constituent elements as those in the first embodiment are denoted by the same reference numerals, and the functions and operations of the same constituent elements as those in the first embodiment are the same as those in the first embodiment. Therefore, detailed description here is omitted.

  The non-target voice section monitoring unit 301 monitors the non-target voice section based on the detection result by the target voice section detection unit 107. Specifically, the non-target voice section monitoring unit 301 receives VAD_RES (K) obtained by the target voice section detection unit 107 and monitors the number of frame sections in which VAD_RES is 0.0 continuously.

  The long-term average value initialization unit 302 receives the number of consecutive frame sections of the non-target voice section from the non-target voice section monitoring unit 301, and when the number of continuous frame sections exceeds a threshold, the detection result long-term average unit 108 calculates This initializes the long-term average value and the long-term average parameter used in the above.

  A state in which the number of non-target speech sections continues longer than a threshold can be said to be a state in which a speaker's speech (target speech) is not input. Therefore, in a period in which the target voice is not input, the long-term average value initialization unit 302 initializes the long-term average value and the long-term average parameter, and deletes the contribution of the non-target voice section accumulated in the long-term average value, It is possible to prevent the beginning of the talk from being lost.

  Since the operation after the target voice is input is the same as that of the first embodiment, detailed description thereof will not be given here.

(B-2) Effects of the Second Embodiment As described above, according to the second embodiment, in addition to the effects of the first embodiment, it is possible to prevent the lack of a talk head part, and to further improve the sound quality. Can be improved.

(C) Third Embodiment Next, a third embodiment of the target sound extraction apparatus and target sound extraction program of the present invention will be described in detail with reference to the drawings.

(C-1) Configuration and Operation of the Third Embodiment FIG. 11 is a configuration diagram showing the configuration of the voice switch 100C of the third embodiment. In FIG. 11, the voice switch 100C of the third embodiment includes a microphone m1 and a microphone m2, an FFT unit 101, a first directivity forming unit 102, a second directivity forming unit 103, a coherence calculating unit 104, an utterance speed. The detection unit 105, the long-term average parameter control unit 106, the target speech section detection unit 107, the detection result long-term average unit 108, the gain control unit 109, the voice switch gain multiplication unit 110, and the frequency subtraction unit 40 are included.

  In the third embodiment, a frequency subtracting unit 40 is further added to the components of the first embodiment. As a result, the disturbing voice (speaking voice of a person other than the speaker) and background noise superimposed on the target voice section, which could not be suppressed by the voice switch, can be suppressed, which is higher than the first and second embodiments. Noise suppression performance can be realized.

  The frequency subtraction unit 40 subtracts the non-target audio signal component from the input signal. As shown in FIG. 11, the frequency subtraction unit 40 includes at least a third directivity forming unit 401, a subtraction unit 402, and an IFFT unit 403.

  The third directivity forming unit 401 receives the signal X (f, K) and the signal X2 (f, K) from the FFT unit 101, and as shown in FIG. 12, a directivity signal B3 having a blind spot in the front direction. (F, K) is formed.

  The reason why the third directivity forming unit 401 forms directivity with the front direction as a blind spot is to acquire a noise signal component included in the input signal. Now, since it is assumed that the speaker utters from the front of the microphones m1 and m2, the third directivity forming unit 401 forms a blind spot on the front to obtain non-target speech coming from the side. can do.

  For example, the third directivity forming unit 401 acquires the signal B3 (f, K) according to Expression (6).

B3 (f, K) = X1 (f, K) -X2 (f, K) (6)
The subtraction unit 402 receives the signal B3 (f, K) from the third directivity forming unit 401 and removes the signal B3 (f, K) that is a noise component from the signal X1 (f, K). For example, the subtraction unit 402 acquires the signal D (f, K) after noise removal according to the arithmetic expression of Expression (7).

D (f, K) = X1 (f, K) -B3 (f, K) (7)
The IFFT unit 403 receives the noise removal signal D (f, K) from the subtraction unit 402, converts the frequency domain signal D (f, K) into the time domain, and performs gain multiplication on the converted signal q (n). This is given to the unit 110.

  Note that the long-term average parameter δ is controlled according to the speech rate by the same processing as in the first embodiment, and the gain control unit 109 outputs VS_GAIN to the gain multiplication unit 110.

  The gain multiplication unit 110 multiplies the output signal q (n) obtained from the IFFT unit 403 by VS_GAIN obtained from the gain control unit 109, and outputs an output signal y (n).

(C-2) Effects of the Third Embodiment As described above, according to the third embodiment, in addition to the effects of the first embodiment, the noise component superimposed on the target speech section is removed. Sound quality can be further improved.

(D) Fourth Embodiment Next, a fourth embodiment of the target sound extraction apparatus and target sound extraction program of the present invention will be described with reference to the drawings.

(D-1) Configuration and Operation of the Fourth Embodiment FIG. 13 is a configuration diagram showing the configuration of the voice switch 100D of the fourth embodiment. In FIG. 13, the voice switch 100D of the fourth embodiment includes a microphone m1 and a microphone m2, an FFT unit 101, a first directivity forming unit 102, a second directivity forming unit 103, a coherence calculating unit 104, an utterance speed. It has at least a detection unit 105, a long-term average parameter control unit 106, a target speech section detection unit 107, a detection result long-term average unit 108, a gain control unit 109, a voice switch gain multiplication unit 110, and a coherence filter calculation unit 50.

  In the fourth embodiment, a coherence filter calculation unit 50 is further added to the components of the first embodiment. Thereby, the noise component superimposed on the target voice section, which could not be suppressed by the voice switch, can be suppressed, and higher noise suppression performance than the first and second embodiments can be realized.

  The coherence filter calculation unit 50 receives coef (f, K) obtained by the calculation formula (3) by the coherence calculation unit 104, and converts it into the input signal X1 (f, K) for each coef (f, K) frequency. Multiply. Thereby, it is possible to suppress a signal component having a bias in the arrival direction, a background noise component having a small waveform regularity, and the like.

  The coherence filter calculation unit 50 includes at least a coherence filter coefficient multiplication unit 501 and an IFFT unit 502.

  The coherence filter coefficient multiplication unit 501 receives coef (f, K) from the coherence calculation unit 104, multiplies coef (f, K) by the signal X1 (f, K) according to the equation (8), and outputs a signal after noise suppression. D (f) is generated.

D (f, K) = X1 (f, K) × coef (f, K) (8)
The IFFT unit 502 receives the noise-suppressed signal D (f, K) from the coherence filter coefficient multiplication unit 501, converts the frequency domain signal D (f, K) into the time domain, and the converted signal q (n ) Is provided to the gain multiplication unit 110.

  Note that the long-term average parameter δ is controlled according to the speech rate by the same processing as in the first embodiment, and the gain control unit 109 outputs VS_GAIN to the gain multiplication unit 110.

  Further, the gain multiplication unit 110 multiplies the output signal q (n) from the IFFT unit 502 by VS_GAIN from the gain control unit 109 to obtain the output signal y (n), and obtains this output signal y (n). Output.

(D-2) Effects of the Fourth Embodiment As described above, according to the fourth embodiment, in addition to the effects of the first embodiment, the noise component superimposed on the target speech section is suppressed. Therefore, the sound quality can be further improved.

(E) Fifth Embodiment Next, a fifth embodiment of the target sound extraction apparatus and target sound extraction program of the present invention will be described with reference to the drawings.

(E-1) Configuration and Operation of Fifth Embodiment FIG. 14 is a configuration diagram showing a configuration of a voice switch 100E of the fifth embodiment. In FIG. 14, the voice switch 100E of the fifth embodiment includes a microphone m1 and a microphone m2, an FFT unit 101, a first directivity forming unit 102, a second directivity forming unit 103, a coherence calculating unit 104, an utterance speed. The detection unit 105, the long-term average parameter control unit 106, the target speech section detection unit 107, the detection result long-term average unit 108, the gain control unit 109, the voice switch gain multiplication unit 110, and the Wiener filter calculation unit 60 are included.

  In the fifth embodiment, a Wiener filter calculation unit 60 is further added to the components of the first embodiment. As a result, it is possible to suppress background noise superimposed on the target voice section that could not be suppressed by the voice switch, and higher noise suppression performance than the first and second embodiments can be realized.

  The Wiener filter calculation unit 60 removes a noise component by multiplying a coefficient obtained by estimating a noise characteristic for each frequency from a signal in a noise section. For the processing by the Wiener filter calculation unit 60, an existing technique can be applied. For example, the technique described in Patent Document 2 can be applied, and detailed description thereof is omitted here.

  The Wiener filter calculation unit 60 includes a Wiener filter coefficient calculation unit 601, a Wiener filter coefficient multiplication unit 602, and an IFFT unit 603.

  The Wiener filter coefficient calculation unit 601 determines whether or not it is a non-target speech section based on the detection result VAD_RES detected by the target speech section detection unit 107. The Wiener filter coefficient wf_coef (f, K) is estimated by the calculation of the equation 3 described above, while the Wiener filter coefficient is not estimated in the case of the target speech section.

  The Wiener filter coefficient multiplication unit 602 multiplies the signal X1 (f, K) by the Wiener filter coefficient wf_coef (f, K) obtained by the Wiener filter coefficient calculation unit 601 according to the equation (9) to obtain a signal after noise suppression. D (f, K) is obtained.

D (f, K) = X1 (f, K) × wf_coef (f, K) (9)
The IFFT unit 603 receives the noise-suppressed signal D (f, K) from the Wiener filter coefficient multiplication unit 602, converts the frequency domain signal D (f, K) into the time domain, and the converted signal q (n ) Is provided to the gain multiplication unit 110.

  Note that the long-term average parameter δ is controlled according to the speech rate by the same processing as in the first embodiment, and the gain control unit 109 outputs VS_GAIN to the gain multiplication unit 110.

  Further, the gain multiplication unit 110 multiplies the output signal q (n) from the IFFT unit 603 by VS_GAIN from the gain control unit 109 to obtain the output signal y (n), and obtains this output signal y (n). Output.

(E-2) Effect of Fifth Embodiment As described above, according to the fifth embodiment, in addition to the effect of the first embodiment, the background noise component superimposed on the target speech section is suppressed. Therefore, the sound quality can be further improved.

(F) Other Embodiments (F-1) In the third to fifth embodiments described above, the technique for suppressing noise by the frequency subtraction technique, the coherence filter, and the Wiener filter has been described. Any one of the frequency subtraction technique, the coherence filter, the Wiener filter, any two, or all the techniques described in the embodiments may be combined. Thereby, higher noise suppression performance can be realized.

(F-2) In the first to fifth embodiments described above, the voice switch includes two microphones m1 and m2, and a directivity signal B1 (f) having a blind spot in the right direction and a blind spot in the left direction. And the case of obtaining coherence based on B2 (f).

  However, the present invention is not limited to this, and includes four microphones and four directivity forming units that form four types of directivity signals, up, down, left, and right, and a signal B1 (f) having a blind spot in the right direction, left direction Alternatively, the coherence COH may be obtained based on the signal B2 (f) having a blind spot in the vertical direction, the signal B3 (f) having a blind spot in the upward direction, and the signal B4 (f) having a blind spot in the downward direction.

In this case, the coherence calculation unit may obtain the coherence COH according to the equations (10) and (4).

(10)
(F-3) In the present invention, the method of controlling the long-term average parameter δ according to the speech rate has been described. However, the lack of target speech occurs not only due to the speech rate, but also due to variations in the distance between the microphone and the speaker. This problem can also be improved by applying the present invention. In this case, instead of the speech rate detection unit, a distance detection unit that estimates the distance between the microphone and the speaker by a known method is provided, and the long-term average parameter control unit controls the long-term average parameter according to the distance. In this way, a correspondence table between distances and long-term average parameters may be stored in the storage unit.

100A-100B ... Voice switch,
101 ... FFT unit, 102 ... first directivity forming unit,
103 ... 2nd directivity formation part, 104 ... Coherence calculation part,
105 ... utterance speed detection unit, 106 ... long-term average parameter control unit,
107 ... target voice section detection unit, 108 ... detection result long-term average part,
109: Gain control unit, 110: Gain multiplication unit,
301: Non-target voice section monitoring unit 302: Long-term average value initialization unit,
40 ... frequency subtraction unit, 50 ... coherence filter calculation unit,
60 ... Wiener filter calculation unit,
201 ... utterance speed input unit, 202 ... long-term average parameter matching unit, 203 ... storage unit, 204 ... long-term average parameter output unit

Claims (7)

A frequency analysis means for converting the input signal from the time domain to the frequency domain;
Directivity forming means for forming a plurality of signals having directivity each having a blind spot in a predetermined direction based on the signal obtained by the frequency analysis means;
A coherence calculating means for obtaining a coherence value based on a plurality of directivity signals formed by the directivity forming means;
Determining whether or not the target sound is included based on the coherence value obtained by the coherence calculating means, and outputting a detection result value corresponding to the determination result; and
By performing a weighted averaging process on the detection result value in the input frame obtained from the target sound determination means and the long-term average value of the detection result value obtained in the frame immediately before the input frame, the input frame A long-term average processing means for obtaining a long-term average value of the detection result values in
Utterance speed detection means for detecting the utterance speed of the target sound included in the input signal based on the coherence value obtained by the coherence calculation means;
Weight coefficient control means for controlling a weight coefficient related to the weighted average processing of the long-term average processing means according to the speech speed detected by the speech speed detecting means;
Gain control means for controlling a gain for an input signal based on a long-term average value of detection result values in the input frame of the long-term average processing means;
A target sound extraction apparatus comprising: gain multiplication means for multiplying an input signal by a gain controlled by the gain control means. The weight coefficient control means is
A storage unit for storing a correspondence table in which the utterance speed is associated with the weighting factor in the long-term average processing unit;
A weighting factor determining unit that determines the weighting factor corresponding to the speech rate obtained from the speech rate detecting means with reference to the correspondence table;
The target sound extraction apparatus according to claim 1, further comprising: an output unit that supplies the weighting factor determined by the weighting factor determination unit to the long-term average processing unit. Non-target sound monitoring means for observing the detection result value of the target sound determination means and monitoring the length of the non-target sound period in which the target sound is not included;
Initializing means for initializing parameters related to the weighted average processing of the long-term average processing means when the non-target sound period length exceeds a threshold based on the monitoring result of the non-target sound monitoring means. The target sound extraction device according to claim 1, wherein A non-target sound signal generating means for forming a blind spot in the target sound direction from the signal obtained by the frequency analyzing means and obtaining a non-target sound signal;
Subtracting means for subtracting the non-target sound signal from the input signal obtained by the frequency analyzing means;
The target sound extraction apparatus according to any one of claims 1 to 3, further comprising: a frequency subtraction unit that includes an inverse frequency conversion unit that converts a signal after noise removal obtained by subtraction into a time domain. A coherence filter coefficient multiplication means for multiplying the signal obtained by the frequency analysis means by a coherence coefficient obtained by a coherence calculation means to obtain a signal component having a bias in the arrival direction and a noise-suppressed signal that suppresses background noise; , Inverse frequency transform means for transforming the signal after the coherence filter coefficient multiplication into the time domain

The target sound extraction device according to claim 1, further comprising a coherence filter calculation unit. Based on the detection result value from the target sound determination means, a Wiener filter coefficient calculation unit that updates a Wiener filter coefficient by a predetermined method only in the case of a non-target sound section;
A Wiener filter coefficient multiplier for multiplying the input signal obtained from the frequency analysis means by the Wiener filter coefficient obtained by the Wiener filter coefficient calculator;
6. The Wiener filter operation means further comprising: an inverse frequency conversion section that converts the frequency domain signal obtained by the Wiener filter coefficient multiplication section into a time domain and gives the multiplication section to the multiplication means. The target sound extraction apparatus according to any one of the above. Computer
Frequency analysis means for converting the input signal from the time domain to the frequency domain,
Directivity forming means for forming a plurality of signals having directivity having blind spots in predetermined directions based on the signals obtained by the frequency analysis means,
A coherence calculating means for obtaining a coherence value based on a plurality of directivity signals formed by the directivity forming means;
A target sound determination unit that determines whether or not the target sound is included based on the coherence value obtained by the coherence calculation unit, and outputs a detection result value according to the determination result;
By performing a weighted averaging process on the detection result value in the input frame obtained from the target sound determination means and the long-term average value of the detection result value obtained in the frame immediately before the input frame, the input frame A long-term average processing means for obtaining a long-term average value of the detection result values in
Utterance speed detection means for detecting the utterance speed of the target sound included in the input signal based on the coherence value obtained by the coherence calculation means;
Weight coefficient control means for controlling a weight coefficient related to the weighted average processing of the long-term average processing means according to the speech speed detected by the speech speed detecting means;
Gain control means for controlling a gain for an input signal based on a long-term average value of detection result values in the input frame of the long-term average processing means;
An objective sound extraction program for causing a gain controlled by the gain control means to function as a gain multiplication means for multiplying an input signal.

Images