JP2007327980A - Digital filter, periodic noise reduction device and noise reduction device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To solve the problem that, in a conventional noise reduction system, noise is emphasized for noise (periodic noise) including line spectrum components, while noise including continuous spectrum components is reduced very well. <P>SOLUTION: In a periodic noise reduction device of the invention, a filter tap in which a filter factor is actually updated, is provided only for a signal of time and/or around the time which is coincident with periodic timing of the periodic noise, on a time axis of delay time made by a plurality of delay elements of adaptive FIR filters. <P>COPYRIGHT: (C)2008,JPO&INPIT

Description

  The present invention relates to a noise reduction device for reducing periodic noise and a digital filter suitable for the noise reduction device.

Currently, there is a problem that it is difficult to make a telephone contact due to noise generated from a ventilation duct using a jet fan or an exhaust duct of a diesel generator. In view of this, a noise re-synthesis method has been studied in which a noise is suppressed by using a difference in characteristics between speech and noise and a speech is enhanced with respect to the speech signal on which such noise is superimposed. This technique effectively suppresses continuous spectrum noise, and the effect has been confirmed by listening experiments using a telephone (see Non-Patent Documents 1 and 2). In addition, it has been confirmed that the continuous spectrum component is effective for noise actually collected.
A. Kawamura, K. Fujii, Y. Itoh, and Y. Fukui, “A new noise reduction method using estimated noise spectrum,” IEICE Trans. Fundamentals, vol. E85-A, no. 4, pp. 784-789, Apr. 2002. Shin Kawamura, Kensaku Fujii, Yoshio Ito, Hiroshi Sokei, “Noise Suppression Based on Linear Prediction Analysis,” IEICE (A), vol. J85-A, no. 4, pp. 415-423, Apr.2002 .

  However, in practice, noise often includes a line spectrum component (hereinafter referred to as periodic noise) composed of a frequency corresponding to the rotation speed and its harmonics in addition to such a continuous spectrum component. On the other hand, in the noise resynthesis method, the reduction effect cannot be obtained with the latter periodic noise, but there is a problem that the periodic noise is emphasized by the amount that the continuous spectrum noise is suppressed. . This is because periodic noise has characteristics similar to vowels and is difficult to distinguish from speech. Therefore, in order to reduce this, it is necessary to devise a way to make the best use of slight differences in characteristics from speech.

  The characteristic difference is that the frequency is known for periodic noise related to the rotational speed of the fan or engine. A method based on the principle of a line enhancer (such as “Sugihara, Fujii, Kawamura, Ito, Sobui,” a study on a method for reducing sinusoidal noise superimposed on speech to reduce noise with known frequencies. ", IEICE Technical Report, US 2002-98, Jan 2003") has been proposed.

  An object of the present invention is to provide a periodic noise reduction device with improved performance and a digital filter therefor.

  In order to solve the above problems, a digital filter according to the present invention is a digital filter used in a periodic noise reduction device, and the impulse response thereof is substantially equal to the periodic timing of the periodic noise. Only have a coefficient value that is neither 0 nor a value in the vicinity thereof.

  In the digital filter, the sum of coefficient values over the entire impulse response may be approximately 1.

  The digital filter may be an IIR type, provided with a feedback loop, having a delay unit in the feedback loop, and a delay time generated by the delay unit may be substantially the same as the period of the periodic noise.

  The digital filter may show an attenuation pattern in which a coefficient value at a time substantially coincident with the periodic timing of the periodic noise becomes smaller as the subsequent stage.

  In the digital filter, the attenuation pattern may be an exponential attenuation pattern.

  The digital filter is of the FIR type, includes a plurality of delay elements connected in series and a plurality of filter taps, and on the time axis of the delay time generated by the plurality of delay elements, Only a signal having a time substantially matching the periodic timing of the periodic noise may be provided with a filter tap in which a filter coefficient that is neither 0 nor a value in the vicinity thereof is set. .

  The digital filter may show an attenuation pattern in which a coefficient value at a time substantially coincident with the periodic timing of the periodic noise becomes smaller as the subsequent stage.

  In the digital filter, the attenuation pattern may be an exponential attenuation pattern.

  In the digital filter, the attenuation pattern may be an attenuation pattern corresponding to the second half of the Hamming window.

  In the digital filter, a coefficient value at a time substantially matching the periodic timing of the periodic noise may indicate a pattern corresponding to a Hamming window.

  In order to solve the above-mentioned problem, another digital filter according to the present invention is a digital filter used in a periodic noise reduction device, which is an adaptive type and has a plurality of delays connected in series. An element and a plurality of filter taps whose filter coefficients can be updated, the sound signal of the periodic noise is used as a main input signal, and the error signal power between the output signal of the digital filter and the sound signal is minimized. The filter coefficient of the filter tap is updated so that the signal having a time substantially matching the periodic timing of the periodic noise on the time axis of the delay time generated by the plurality of delay elements is obtained. Only a filter tap capable of substantially updating the filter coefficient is provided.

  In the above digital filter, whether or not a filter tap is provided for an output signal of a delay element other than the delay element in which the filter coefficient is multiplied to the output signal by a filter tap capable of substantially updating the filter coefficient. Alternatively, a filter tap for multiplying 0 or a value in the vicinity thereof as a filter coefficient may be provided.

  In the digital filter, the adaptive algorithm executed in the digital filter may be a learning identification method.

  In the digital filter, in the filter tap capable of substantially updating the filter coefficient, the total value of the maximum values of the filter coefficient updateable range may be 1 or a value in the vicinity thereof.

  In the digital filter, the filter coefficient may be limited to 0 or a positive value in the filter tap capable of substantially updating the filter coefficient.

  In the digital filter, the filter tap that can substantially update the filter coefficient may show an attenuation pattern such that the maximum value of the filter coefficient updateable range becomes smaller as the subsequent filter tap. .

  In the digital filter, the attenuation pattern may be an exponential attenuation pattern.

  In the digital filter, the attenuation pattern may be an attenuation pattern corresponding to the second half of the Hamming window.

  In the digital filter, in the filter tap capable of substantially updating the filter coefficient, the maximum value of the filter coefficient updatable range may indicate a pattern corresponding to the Hamming window.

In the digital filter, the periodic noise includes a noise component having a first period and a noise component having a second period, and the second period is larger than the first period and is generated by a plurality of delay elements. On the time axis of the delay time to be tightened, it is substantially only for a signal that substantially matches the periodic timing of the first period and a signal that substantially matches the periodic timing of the second period. A filter tap capable of updating the filter coefficient may be provided.

  In the digital filter, the periodic noise includes a noise component having a first period and a noise component having a second period, and the second period is larger than the first period and is generated by a plurality of delay elements. A time substantially coincident with the periodic timing of the first period at a time before a predetermined time on the time axis of the delayed time, and a time axis of the delay time generated by the plurality of delay elements, Only a signal having a time substantially matching the periodic timing of the second period may be provided with a filter tap that can substantially update the filter coefficient.

  In order to solve the above-mentioned problem, still another digital filter according to the present invention is a digital filter used in a periodic noise reduction device, and is an adaptive type. Steps 3 and 3 are executed. In the first step, the filter coefficients are updated in all the filter taps of the digital filter. In the second step, the filter coefficients converged by executing the first step are obtained. A filter tap larger than a predetermined threshold is extracted as the first filter tap, and in the third step, the filter coefficient is allowed to be updated for the first filter tap extracted by executing the second step. For other filter taps, the filter coefficient is set to 0 or a value in the vicinity thereof, and The update of the filter coefficient is not permitted.

  In the digital filter, the adaptive algorithm executed in the digital filter may be a learning identification method.

  In the digital filter, in the third step, the total value of the maximum values of the filter coefficient updateable range of the first filter tap may be 1 or a value in the vicinity thereof.

  In the digital filter, the filter coefficient of the first filter tap may be limited to 0 or a positive value in the third step.

  In order to solve the above-mentioned problem, still another digital filter according to the present invention is a digital filter used in a periodic noise reduction device, which is an adaptive type, and includes a first adaptation algorithm execution unit and a second adaptation filter. An algorithm execution unit, and the first adaptive algorithm execution unit and the second adaptive algorithm execution unit have the same number of filter taps and unit delay time, and the main input signal input to the first adaptive algorithm execution unit is A filter tap that is also input to the second adaptive algorithm execution unit as a main input signal and has a filter coefficient that is converged by the execution of the adaptive algorithm by the second adaptive algorithm execution unit is greater than a predetermined threshold value. In the first adaptive algorithm execution unit, the same as the extracted second filter tap The filter coefficient is allowed to be updated for the filter tap having the loop number, and for the other filter taps, the filter coefficient is set to 0 or a value in the vicinity thereof, and the filter coefficient is updated. It is not allowed.

  In the digital filter, both the adaptation algorithm executed in the first adaptation algorithm execution unit and the adaptation algorithm executed in the second adaptation algorithm execution unit may be learning identification methods.

  In the above digital filter, in the adaptive algorithm of the first adaptive algorithm execution unit, the total value of the maximum values of the filter coefficient updateable range at the filter tap that allows the filter coefficient to be updated is 1 or a value in the vicinity thereof. May be.

  In the digital filter, in the adaptive algorithm of the first adaptive algorithm execution unit, the filter coefficient at the filter tap that is allowed to update the filter coefficient may be limited to 0 or a positive value.

  In order to solve the above problems, a periodic noise reduction device according to the present invention includes an audio signal input unit, an audio signal output unit, and a filter, and an audio signal including periodic noise is included in the audio signal input unit. An audio signal input and input to the audio signal input unit is input to the filter, and a signal generated by subtracting the output signal of the filter from the audio signal input to the audio signal input unit is the audio signal output. The filter is a digital filter described above.

  In order to solve the above problems, a noise reduction device according to the present invention is a noise reduction device including a first noise reduction device and a second noise reduction device connected in series at a subsequent stage thereof, One of the noise reduction device and the second noise reduction device is the periodic noise reduction device, and the other is a device that reduces noise based on a noise resynthesis method using a linear prediction filter. .

  In order to solve the above-described problem, another periodic noise reduction device according to the present invention includes a first microphone, a second microphone, a speaker, and a filter, and the filter is the digital filter described above. Periodic noise is input to the first microphone, an output signal of the first microphone is input to the filter, an output signal of the filter is transmitted to the speaker, and the periodic noise and sound emitted from the speaker are Is input to the second microphone, and the output signal of the second microphone is input to the filter as an error signal for executing the adaptive algorithm.

  According to the present invention, periodic noise can be effectively reduced.

[1. Evaluation methods]
(Periodic noise reduction system based on line enhancer principle)
FIG. 1 shows a periodic noise reduction system (Line Enhancer System: LES) based on the principle of a line enhancer. However, Z ^−N is an element that gives a delay of N sampling periods to an input signal, and ADF (Adaptive Digital Filter) is an adaptive filter. In this study, the most common learning identification method (for example, “Noda Akihiko, Nagumo Niichi,“ System Learning Identification Method, ”Measurement and Control, vol.7, no, .9, pp.597-605 (1968) ”

Is used. Here, h (n) is a coefficient vector set in ADF at time n, μ is a step size, y (n) is a prediction error, and x (n) is a reference signal vector set in ADF.

  Further, in the subsequent simulation, a signal obtained by superimposing the periodic noise shown in FIG. 3 on the male voice shown in FIG. 2 with an SN ratio of 0.0 dB is used as an input signal. FIG. 4 shows the input signal waveform. However, the periodic noise is a set of line spectra whose fundamental frequency is f0 = 160Hz, which simulates the noise of an actual jet fan.

As given. Here, Ts = 1/8000 is a sampling period, and r (k) is a uniform random number from 0 to 1.

  FIG. 5 shows an output signal of LES when the signal of FIG. 4 is given to the system of FIG. However, the delay number N = 66, the tap number M = 200, and the step size μ = 0.05.

  In addition, the system performance is evaluated for the sound quality and noise suppression effect of the output signal. As an evaluation formula of sound quality

As an evaluation formula of the noise suppression effect

Is introduced. Here, K is the total number of samples, and K = 50,000. For the evaluation, three same systems are prepared. One of the signals shown in FIG. 4 is input, and the obtained ADF coefficient is copied to the remaining two. Furthermore, y _s (n) is the output obtained by inputting only the speech x _s (n) to the two systems with the copied coefficients, and y is the output obtained by inputting only the noise x _n (n). _{Let n} (n). In the above evaluation formula, the larger the value, the higher the performance.

[2. Embodiment A]
When the signal shown in FIG. 5 was evaluated using the evaluation method described above, VE = 5.19 dB and SNRout = 19.59 dB. When I tried listening, some distortion was felt in the sound, and a little noise remained in the sound part.

(Examination of update tap limitation: ADF tap coefficient)
FIG. 6 shows the tap coefficient values of the ADF when the periodic noise shown in FIG. 3 is given as an input signal to the system shown in FIG. From FIG. 6, when periodic noise can be reduced, it can be confirmed that the tap coefficient has a peak for each pitch period, and other values have values close to 0.0. Therefore, it is expected that ADF can be specialized in periodic noise reduction and performance can be improved by limiting the taps for coefficient updating to only taps having peak values and setting other coefficient values to 0.0. Is done.

(Limited effect of update tap)
When the predicted noise is completely periodic, the noise is completely canceled if the tap coefficient corresponding to the first period including the number of delays is 1.0 and the others are 0.0. However, FIG. 6 is not so. The reason is explored by simulation.

  First, according to FIG. 6, since it is not 0.0 completely other than the peak, the tap to be updated includes before and after the peak. Further, if the number of taps used for coefficient update becomes very small, it becomes easy to react to voice, so μ = 0.00025. Furthermore, since the noise cycle is 50 samples and the delay number N = 66, the taps that become the peaks are the 34th, 84th, 134th, and 184th taps out of the 200 taps. A simulation is performed by giving a signal shown in FIG. 4 in which voice is added to periodic noise at an SN ratio of 0.0 dB to this system. FIG. 7 shows the number of peaks to be updated and the evaluation values VE and SNRout.

  As shown in FIG. 7, as the number of updated peaks increases, both VE and SNRout are improved.

  The reason for this improvement is confirmed by the waveform. 8 and 9 show tap coefficient values and output signals when the number of update peaks is 1, and FIGS. 10 and 11 show tap coefficient values and output signals when the number of update peaks is 4. FIG. When FIG. 9 and FIG. 11 are compared, when the update peak value is 1, it can be confirmed that the noise increases after the voice is input, and slowly decreases after the voice becomes silent. On the other hand, when the number of update peaks is 4, although the noise similarly increases after the input of the voice, the decrease after becoming no voice is quick. Compared to FIG. 8, the coefficient of each tap is smaller when the number of updated peaks in FIG. 10 is larger, and becomes a coefficient value for reducing periodic noise after the coefficient value is disturbed by speech (hereinafter referred to as coefficient). This is thought to be due to the rapid convergence of Further, when the signal shown in FIG. 9 and the signal shown in FIG. 11 are auditioned, the signal shown in FIG. This is probably because the tap coefficient is smaller when the number of update peaks is larger, and the amount of damage to the voice is also smaller.

  From the above results, as the number of taps having peaks increases, the influence of the tap coefficient from the sound is small, and at the same time, the influence on the sound is also small. In other words, it can be said that a larger number of taps is more effective in reducing periodic noise.

(Effect of increasing the number of taps)
Here, the output signal when the signal of FIG. 4 is given is evaluated by Equation (3) and Equation (4), and compared with LES that updates all tap coefficients.

  According to FIG. 7, the evaluation value of the output signal is improved as the number of taps having a peak increases. However, in LES, an audio echo is output when the number of taps increases. On the other hand, in the method of limiting the update taps, many tap coefficients are 0.0, and the influence from the voice is small. In addition, when the coefficient is updated only for the tap having the peak value and one tap before and after the tap, the number of taps to be updated is 3/50 compared to LES, and the amount of calculation is also slightly increased. Therefore, a simulation is performed with the number of ADF taps M = 1,000 as a test. In this case, the number of update taps is 60, which is smaller than the LES that updates all tap coefficients. However, if the number of taps is reduced, the convergence becomes faster, so it is necessary to reduce the step size in order to suppress the influence from the voice. Therefore, here, μ = 0.00025. Further, as in the above simulation, the delay number N = 66. FIG. 12 shows an output signal when the signal of FIG. 4 is given under the above conditions.

  Referring to FIG. 12, the noise is reduced from the result of FIG. However, noise remains after the voice. As a result of the trial listening, a little noise remains in the voice part, and even though the number of taps is as large as M = 1,000, no echo of the voice is felt.

  The evaluation values were VE = 9.47 dB and SNRout = 21.43 dB. Compared to the evaluation values VE = 5.19 dB and SNRout = 19.59 dB of the LES output signal for updating all tap coefficients, VE is improved by 4.28 dB and SNRout is 1.84 dB. It can be seen that the method of limiting the update tap is effective in reducing periodic noise.

(Application to MRI noise)
MRI (Magnetic Resonance Imaging) noise is periodic and its frequency depends on the application. However, since the frequency is known, a method of limiting the update tap can be applied. Here, a method for limiting update taps is applied to MRI noise, and its effect is confirmed.

  FIG. 13 shows the MRI noise used in this simulation. This noise is a signal having two periods of 400 samples and 3,200 samples. In the subsequent simulation, the signal shown in FIG. 14 in which the male voice shown in FIG. 2 is added to the noise at an SN ratio of 0.0 dB is used as an input signal. FIG. 15 shows an output signal when this signal is given to the LES for updating all tap coefficients, and FIG. 16 shows the value of the tap coefficient at that time. However, the delay number N = 66, the tap number M = 4,000, and the step size μ = 0.05.

  The evaluation values of this output signal are VE = 4.66 dB and SNRout = 5.45 dB. Moreover, it can confirm that the peak value is taken for every 400 taps which are pitch periods from FIG. When this output signal is auditioned, it is confirmed that although the noise is reduced, a large echo of the voice remains.

  Next, the signal shown in FIG. 14 is given to a method for limiting the update taps. Since the period of periodic noise is equivalent to 400 taps, if only taps having tap numbers that are integer multiples of 400 are extracted, they are taps corresponding to the time corresponding to the periodic timing of periodic noise. Become. Therefore, the delay number is N = 66, and the tap to be updated is an integer multiple of 400 plus or minus 1 tap. Only in the tap to be updated, a coefficient value that is neither 0 nor a value in the vicinity thereof is set. The step size is μ = 0.00025. FIG. 17 shows the result of this simulation.

  The evaluation values of this output signal are VE = 6.16 dB and SNRout = 6.36 dB. Both VE and SNRout have improved VE by 1.5 dB and SNRout by 0.91 dB compared to the result of LES that updates all tap coefficients. However, the same level of echo remains.

  This is probably because the number of taps is 4,000, and there is only one peak that is an integral multiple of the 3,200 sample period. Therefore, the number of taps is set to M = 20,000, and the number of peaks that is an integral multiple of 3,200 samples is increased. However, since an echo occurs when the number of taps increases in LES, the peak of an integer multiple of 400 samples to be updated is limited to 3,200 taps. Here, the 400 sample period is the first period, and the 3,200 sample period is the second period. The second period is larger than the first period. Only for signals that substantially match the periodic timing of the first period in the time before 3,200 taps, and signals that substantially match the periodic timing of the second period in the entire tap range. , Update tap is set. In consideration of the increase in the number of taps, the step size is set to μ = 0.0005 here.

  In addition, increasing the number of taps to be updated before and after one peak is considered to make it easier to cope with fluctuations in the cycle. Therefore, 25 taps before and after the tap that takes each peak as a test are updated.

  At this time, FIG. 18 shows an output signal when the signal of FIG. 4 is inputted, and FIG. 19 shows a tap coefficient.

  The evaluation values of this output signal are 9.05 dB and 7.03 dB, and VE is improved by 4.39 dB and SNRout is improved by 1.58 dB compared with the result of LES for updating all the tap coefficients. Compared with the method of limiting the coefficient update with the tap number of 4,000, improvements in VE are 2.89 dB and SNRout is 0.67 dB. When you actually listen to it, you can feel some echoes of the background noise. From this, it was confirmed that the method of limiting the update tap is effective in reducing MRI noise.

  As described above, with regard to the periodic noise reduction method, when the noise frequency is known, the tap coefficient to be updated is limited, so that the influence of the tap coefficient from the voice is reduced. Therefore, it has been confirmed that the noise reduction effect is increased and the influence of the output signal on the sound is reduced. At this time, the step size of LES needs to be smaller than when all the tap coefficients are updated. By increasing the number of taps to be updated before and after one peak, the influence of the tap coefficient from the voice is further reduced. It was confirmed that

[3. Embodiment B]
(Refresh taps only)
In the following, we will examine techniques for further improving performance from techniques for limiting update taps. Hereinafter, the ADF that limits the tap to be updated is referred to as a limited ADF.

  FIG. 20 shows values of tap coefficients obtained after the periodic noise is reduced using the limited ADF and converged. However, the delay number N = 49, the tap number M = 1,000, and the step size μ = 0.00025. An LES output signal obtained under these conditions is shown in FIG. When this signal was evaluated, VE (evaluation value) = 11.25 dB and SNRout = 22.24 dB.

  Here, referring to FIG. 20, the sum of the tap coefficient values is approximately 1. If Mp is the number of periods of noise included in the limited ADF, the average tap coefficient value is approximately 1 / Mp. Yes. From this, even if 1 / Mp is given as a coefficient value to a tap that is a constant multiple of the noise period so that the sum of tap coefficients becomes 1, it is expected that the same periodic noise reduction effect can be obtained.

  In FIG. 22, the tap coefficient value is not given adaptively, 1 / Mp is given to taps that are a constant multiple of the noise period, and 0 is given to other taps, and the number Mp of non-zero taps is changed. This is the value of VE and SNRout when Here, since the frequency of periodic noise is 160 Hz and the sampling period is 1 / 8,000, the period of noise is 50. Therefore, when the number of delays of the delay elements placed in front of the limited ADF is N = 49, the taps having non-zero values are 1,51,101,151,.

  In FIG. 22, the value of SNRout is −1 because the noise has been completely reduced. When the denominator of Equation (4) becomes 0, −1 is forcibly output. This is because. This indicates that when a value is given to a tap that is a constant multiple of the noise period so that the sum of the coefficients is 1, the periodic noise is completely reduced.

  Further, it is confirmed that the value of VE is improved as the number of non-zero taps Mp is increased. This is because 1 / Mp given as the tap coefficient value becomes smaller as Mp becomes larger, thereby reducing the influence of voice on the output signal.

  Also, the reason why the value of VE does not increase linearly is considered to be easily affected by reverberation as the number of taps increases. FIG. 23 shows the output result when the number of taps is 1,000. The VE at this time is 15.18 dB. Compared to FIG. 2, reverberation is confirmed in the silent section. In addition, reverberation is confirmed even when actually listening to the sample. However, since the VE when the number of taps is 1,000 in the limited ADF is 11.25 dB, it can be seen that the value of VE is improved by giving 1 / Mp to the tap coefficient value that is a constant multiple of the noise period. .

  As described above, when the tap coefficient is updated, the limited ADF is affected by the voice. On the other hand, when the coefficient is fixed, the coefficient is not affected by the voice. This means that the step size is reduced in the speech interval, that is, the coefficient update amount is limited (for example, “Tetsu Miyata, Kazuaki Maeda, Tomohiro Amitani, Kensaku Fujii, Yoshio Ito,” 2 microphones were used. This shows that VE is improved by the verification of the noise suppression system using actual devices ", 2005 Haruon Ronshu, 1-6-24, (2005.3).

(Improvement of output sound quality)
In the above description, it was confirmed that when the noise period is known, the periodic noise can be completely reduced, and that the value of VE can be improved by increasing the number of taps. However, when the number of taps increases, there arises a problem that sound reverberation remains in the output signal. Therefore, we will tentatively set the number of taps to 1,000 and consider a method to improve the sound quality of the output sound without changing the number of taps. However, the number of delays is N = 49. FIG. 24 shows the frequency characteristics of the limited ADF at this time.

  Referring to FIG. 24, it can be seen that since the fundamental frequency of noise is 160 Hz, a peak of amplitude 1 stands at a position that is an integral multiple of the fundamental frequency. That is, it can be said that ADF constitutes a narrow bandpass filter. However, since the pass band has a width, it is considered that a large amount of sound passes. Therefore, it is presumed that a decrease in sound quality can be suppressed by further narrowing the pass band. As one of the means, consider applying a hamming window to the tap coefficient to narrow the pass band. However, adjustment is made so that the sum of the tap coefficients becomes 1 even after the window is applied.

  FIG. 25 shows the value of the tap coefficient multiplied by the Hamming window, and FIG. 26 shows the spectrum given by the tap coefficient multiplied by the Hamming window.

  FIG. 27 shows the relationship between the number of taps obtained for the limited ADF having the coefficient multiplied by the Hamming window and VE. Referring to FIG. 27, the VE value is lower when the window processing is performed than when the window processing is not performed. There are two possible reasons for this. One is that the pass bandwidth is doubled with the same number of taps by applying a Hamming window. FIG. 28 shows an output signal when the number of taps is 1,000. Also, according to the trial listening of this signal, the reverberation is more when the signal is applied than when the Hamming window is not applied.

  The reason why the VE was lowered by performing the window processing is that the tap around the center of the filter, that is, the tap output subjected to a relatively large delay, is larger than the case where the window is not applied. This is probably because reverberation increased.

  Therefore, a tap having a smaller delay is given a larger coefficient, and a tap having a larger delay is given a smaller coefficient. This is expected to reduce the reverberation of the voice in the output signal.

  FIG. 29 shows tap coefficients given so that a half of the Hamming window is subjected to a tap coefficient so as to have an attenuation tendency, and FIG. 30 shows frequency characteristics thereof. FIG. 31 shows an output signal at this time. Looking at this frequency characteristic, it is worse in the stop band than when the normal Hamming window is multiplied as a coefficient. FIG. 32 shows the relationship between the number of taps and VE when the coefficient value tends to attenuate and when the coefficient value is constant at 1 / Mp. Compared to FIG. 27, the tendency is almost the same.

  On the other hand, when the signal shown in FIG. 31 is sampled, reverberation can hardly be confirmed. This is presumably because the characteristic deterioration of the stop band was offset by the narrowing of the pass band. From the result of the trial listening, it can be seen that when the tap coefficient value of the limited ADF tends to be attenuated, the influence of the signal having a large delay on the output signal is reduced and the influence of reverberation is reduced. For this reason, it is effective to prevent the deterioration of sound quality by reducing the passband by increasing the number of taps, and by preventing the influence of reverberation caused by increasing the number of taps by providing the tap coefficient with attenuation characteristics. I know that there is.

(Exponential decay type FIR filter)
In the above, it was confirmed that it is effective to give the tap coefficient a damping characteristic. However, even when a tap coefficient multiplied by half of the Hamming window was given, VE was lower than when the coefficient value was constant at 1 / Mp. Therefore, an exponential attenuation type is considered as a method of giving a tap coefficient that has attenuation characteristics and further improves VE. FIG. 33 shows tap coefficients given so as to be an exponential decay type, and FIG. 34 shows frequency characteristics thereof. However, the common ratio is temporarily set to r = 0.9, and the sum of the tap coefficients is adjusted to 1. FIG. 35 shows the output signal at this time and VE was 13.45 dB.

  From this frequency characteristic, it can be seen that the frequency characteristic is similar to that obtained by multiplying half of the Hamming window as a coefficient. FIG. 36 shows the value of VE when r is changed from 0 to 1. As can be seen from FIG. 36, the larger the value of r, the better the VE. Moreover, it is considered that the inclination is small when r> 0.9, because of the effect of reverberation that r receives because r takes a value close to 1.

  At this time, the number of taps having non-zero coefficient values is Mp, and the sum of the tap coefficients of the Mp is adjusted to be 1. Therefore, when the coefficient is given to the exponential attenuation type, when Mp is small, the exponential attenuation is reduced. It is necessary to increase the initial value, and as a result, the sound is damaged. When the signal shown in FIG. 35 was auditioned, no reverberation could be confirmed, but distortion was felt in the portion where the sound was loud. From this, it is presumed that the number of Mp needs to be very large when a coefficient is given to the exponential decay type.

(IIR type periodic noise reduction system)
In the foregoing, it has been found that when the coefficient is given to the exponential decay type, it is necessary to take a very large number of taps. Therefore, a system as shown in FIG. 37 is considered. Here, Z ^−N is an N delay element, and a is a value satisfying 0 <a <1. In this system, by giving [Noise period −1] to N, the same effect as the limited ADF in which the tap coefficient is given to the exponential decay type can be obtained. Is very large.

  First, a simulation for obtaining a value of coefficient a suitable for periodic noise reduction was performed. FIG. 38 shows the values of VE and SNRout when a is changed from 0 to 1.

  From FIG. 38, it was confirmed that the closer the value of a is to 1, the better the VE. Therefore, simulation is performed with a = 0.99. FIG. 39 shows an output signal when the signal of FIG. 4 is input to the IIR type periodic noise reduction system.

  When this signal was evaluated, VE = 25.27 dB and SNRout = 100.2 dB, and the evaluation values of the output signal of the limited ADF were VE = 11.25 dB and SNRout = 22.34 dB. Compared to the limited ADF, It can be confirmed that VE is improved by 14.02dB and SNRout is improved by 77.86dB. When actually auditioning, the reverberation of the voice was not confirmed, and the residual noise was not confirmed. From this, it was confirmed that the IIR type periodic noise reduction system is effective in reducing the periodic noise when the frequency of the periodic noise is known.

  When the IIR type periodic noise reduction system proposed this time is applied to actual noise, fluctuations occur in the period of noise during sampling. For this reason, it is considered necessary to deal with not only an integer multiple of the noise cycle but also several samples before and after that.

  It is also applied when the noise frequency is unknown. As a concrete plan, a filter for detecting the period of noise may be provided in the system so that the number of delays of the system is detected.

[4. Embodiment C]
(FIR filter with periodic noise reduction)
FIG. 40 shows the tap coefficient of ADF when only the noise of FIG. 3 is given as input to LES. However, the number of delays N = 0, the number of taps M = 1,000, and the step size μ = 0.010.

  From FIG. 40, it can be confirmed that the tap coefficient has a peak for each pitch period, and other values have values close to zero. From this, it is possible to specialize ADF for periodic noise reduction by limiting the tap for coefficient update to only taps for which noise having the same period as the input noise is input, and setting other coefficient values to 0. Performance is improved. Also, when the frequency of periodic noise is known, a constant is given to the coefficient value of the tap to which noise having the same period as the input noise is input as an FIR filter without updating the coefficient of the ADF, and the coefficient value of other taps The noise reduction performance is improved by giving 0 to and setting the sum of the tap coefficients to 1.

(System using a filter with coefficients given as constants)
FIG. 41 shows a tap coefficient when a value is given only to a tap to which noise having the same period as the input noise is inputted and 0 is given to other coefficients, and an output signal thereof is shown in FIG. At this time, the given value is 1 / M _p . Where M _p is the number of taps giving the value.

  When the signal in FIG. 42 was evaluated, VE = 15.18 dB, and the noise was completely reduced. When actually listening, no residual noise could be confirmed, but the reverberation of the voice was confirmed.

  A system using an FIR filter that gives constants to specific coefficients and gives 0 to other coefficients is hereinafter referred to as CPPS (Coefficient have Peaks with respect to each Period of noise System).

(Applied to non-integer noise)
The frequency of noise used in the simulation so far is 160 Hz. Since the sampling period is 8,000, the noise period is 50 samples, which is an integer. Therefore, consider application to noise having a frequency at which the period of the noise is not an integer.

  The noise used in subsequent simulations is shown in FIG. As an input signal, a signal in which the periodic noise of FIG. 43 is superimposed on the male voice shown in FIG. 2 with an SN ratio of 0 dB is used. FIG. 44 shows the waveform.

Since the frequency of this noise is 120 Hz, the sampling period is T _s = 1 / 8,000 seconds, so the noise period is 66.66... Sample and not an integer value. Therefore, it is 66.66, 133.33, 200, ... delayed signals that have the same period as the input noise, and since it is not an integer, a value is given to the tap to which noise having the same period as the input noise is input. Can't give. Therefore, in the case of a tap for inputting the noise closest to the noise having the same cycle as the input noise, that is, for noise having a frequency of 120 Hz, 66.66, 133.33, 200,. , 200,... A value is given to a tap to which a delayed signal is input.

  FIG. 45 shows the tap coefficient of the constant coefficient CPPS using the above method, and FIG. 46 shows the output signal. However, in order to compare with the case where the frequency is 160 Hz, the number of taps is adjusted. When the frequency was 160 Hz, the number of taps M = 1,000 and the noise cycle was 50 samples, so the number of taps giving a value was 20. When the frequency is 120 Hz, the noise cycle is 66.66 samples, so the number of taps that gives 20 values is M = 1,333. Thereafter, when the noise frequency of the input signal is 120 Hz, the number of taps is M = 1,333.

  When this output signal was evaluated, VE = 11.95 dB and SNRout = 36.2 dB. Moreover, when actually listening to the sound, although the noise was sufficiently reduced, the reverberation of the voice was confirmed in the same manner as in the case of the constant coefficient CPPS with a noise frequency of 160 Hz.

(Applicable condition)
Based on the above results, when the frequency of periodic noise is known, taps with the same period as the input noise are input, or taps with the period closest to the input noise are input (hereinafter referred to as peak taps). It was confirmed that a system (CPPS) that gives a value only to the other tap coefficient and gives 0 to other tap coefficients is effective in reducing noise. However, this method needs to know the tap at which the noise peaks, so the noise cycle needs to be known.

(ADF with limited update taps)
The aforementioned CPPS cannot be applied unless the noise period is known. Therefore, consider application to noise of unknown frequency. However, the purpose of the simulation here is noise suppression. Therefore, a method is adopted in which a signal of only 50,000 samples of noise is input before FIG. 44, the tap coefficients are converged, and the coefficient is not updated in the section of FIG. 44 in which speech is input.

(System using ADF with limited update taps)
It has been confirmed that the periodic noise reduction performance is improved by limiting the taps of the ADF to be updated with respect to the LES shown in FIG. Hereinafter, the ADF that limits the tap to be updated is referred to as a limited ADF.

FIG. 47 shows tap coefficient values when the signal of FIG. 44 is given as an input signal. However, the delay number N = 0, the tap number M = 1,333, and the step size μ = 1.00 × 10 ⁻⁴ . FIG. 48 shows the output signal of LES obtained under these conditions. When this signal was evaluated, VE = 11.8 dB and SNRout = 35.4 dB. Moreover, when actually listening to the sample, it can be confirmed that the reverberation of the sound is not heard but the noise remains in the sound part.

  Therefore, here, a method for improving the performance in the LES using the limited ADF is considered. Hereinafter, LES using limited ADF is referred to as R-LES (Restricted-LES).

(Improved performance)
First, attention is focused on the fact that CPPS has better performance than R-LES. In CPPS, only the peak tap was given a value, whereas in R-LES, the update tap was the peak tap and several taps before and after. Since the evaluation value of CPPS was better than that of R-LES, it is considered that a method of giving a value only to the peak tap is effective in reducing periodic noise. In view of the influence of the output signal of the filter on the sound, it is not desirable that the filter coefficient value is negative. Therefore, a limitation is added to the limited ADF that only taps that are peaked are updated, and when the coefficient value becomes negative at the time of updating, it is forcibly set to 0. If it becomes negative, it is forced to be 0, which means that the filter coefficient is limited to 0 or a positive value.

  Furthermore, since it has been confirmed that the noise reduction effect is improved by setting the sum of coefficient values to 1 from the CPPS result, the update tap is set so that the sum of coefficients becomes 1 after updating the coefficients of the limited ADF. Is multiplied by a constant.

  The performance of R-LES to which the above method is applied is confirmed. FIG. 49 shows tap coefficients when the signal of FIG. 44 is given as an input. The output signal is shown in FIG. When this signal was evaluated, VE = 11.9 dB and SNRout = 36.5 dB. In addition, when actually listening, no sound reverberation or noise remained.

(Update tap detection)
In conventional systems, the noise frequency must be known. Therefore, in order to cope with noise whose frequency is unknown, a method of automatically detecting a tap to be updated in R-LES is proposed.

  First, ADF that updates all taps is used from the start of signal input until 1,000 samples have elapsed, and the tap coefficients are converged. An average value of tap coefficient values is set as a threshold, a tap having a value equal to or higher than the threshold is set as an update tap, and limited ADF is used after 1,000 samples have elapsed. That is, in the filter, the first step, the second step, and the third step are executed. Here, the first step is a step in which the filter coefficient is updated by all the filter taps. The second step is a step in which filter taps whose filter coefficients converged by executing the first step are larger than a predetermined threshold are extracted. In the third step, the filter coefficient is allowed to be updated for the filter tap extracted by executing the second step, and the filter coefficient is set to 0 for the other filter taps. In addition, it is a step for preventing the update of the filter coefficient from being permitted.

  Further, after that, the tap coefficient is monitored, and for every 1,000 samples, the average value of the updated tap coefficient values is set as a threshold, and a tap having a value equal to or higher than the threshold is newly set as an update tap (update tap setting). A specific configuration for the purpose will be described later). However, a lower limit is provided so that the number of taps to be updated does not become too small, and resetting is not performed when the number of taps to be updated falls below the lower limit. In the above-described CPPS, the number of taps that become a peak is 20, so the lower limit of the taps to be updated is 20 here. The coefficient update of ADF and limited ADF is performed every time one sample passes.

FIG. 51 shows tap coefficient values when the signal of FIG. 43 is input when this method is applied. However, the number of delays N = 0, the number of taps M = 1,333, and the step size μ are μ = 0.01 when ADF is used, and μ = 1.0 × 10 ⁻⁴ when limited ADF is used. 51, the taps being updated are 66.66, 133.33, 200,..., And when the noise frequency is 120 Hz, a signal that is an integral multiple of the noise period is obtained. It is a tap that rounds off the decimals of the input tap. The output signal of the system at this time is shown in FIG.

  When the signal in FIG. 52 was evaluated, VE = 11.9 dB and SNRout = 36.5 dB. In addition, when actually listening, no sound reverberation or noise remained. From this, it was confirmed that even if a method for automatically detecting the tap to be updated was used, the performance was comparable to that of the system in which the coefficient was given as a constant.

(Processing in the voice section)
In the above, the coefficient update of the limited ADF is not performed in the speech section, but here, the coefficient is updated and the performance is compared and examined.

  First, FIG. 53 shows the tap coefficient when the coefficient update of the limited ADF is continued even in the voice section, and FIG. 54 shows the output signal. When this signal was evaluated, VE = 8.01 dB and SNRout = 26.8 dB. In addition, when actually listening, the noise was sufficiently reduced outside the speech section, but residual noise was confirmed in the speech part. Therefore, it is considered that the reason why noise remains in the voice part is that the tap coefficient value of the limited ADF is disturbed when voice is input. Furthermore, the limited ADF has very few taps to update. Since this has the same effect as increasing the step size in the ADF, it is considered that the limited ADF is more easily disturbed in the coefficient value than the ADF. Therefore, in order to suppress the disturbance of the tap coefficient, a method of performing clip processing using an appropriate threshold value and limiting the coefficient update amount of the limited ADF is used.

FIG. 55 shows the update amount of the coefficient of the 200th tap which is the update tap when the output signal of FIG. 54 is obtained. As shown in FIG. 55, the coefficient update amount when the coefficient converges with respect to the noise is within a range of ± 5.0 × 10 ⁻⁵ . However, at the beginning of convergence, the update amount is in the range of ± 1.5 × 10 ⁻⁴ . Therefore, the threshold value is set to 1.0 × 10 ⁻⁴ in order to suppress a large coefficient fluctuation after the coefficient convergence and to sufficiently converge the coefficient in the noise part to be given first.

  FIG. 56 shows an output signal when the clip processing is added to the coefficient update in this way. When the signal in FIG. 56 was evaluated, VE = 10.4 dB and SNRout = 32.4 dB. In addition, when actually listening, no residual noise was confirmed.

(Corresponding to frequency change (new update tap setting method))
The next issue is application to actual noise and response to changes in noise frequency. As a specific method for dealing with a change in frequency, a method of connecting a filter for detecting a noise period in parallel to the system and updating a tap corresponding to the detected period can be considered.

  First, a technique of converging tap coefficients using an ADF that updates all taps until 1,000 samples have elapsed after starting to input a signal, and using a limited ADF as a tap having a value greater than or equal to a threshold, In addition, a technique is also effective in which the tap coefficient is monitored and the average value of the tap coefficient values updated every 1,000 samples is set as a threshold, and a tap having a value equal to or greater than the threshold is newly set as an updated tap. Said. Here, a specific configuration for setting a new update tap for every 1,000 samples will be described.

  FIG. 57 shows an apparatus configuration for newly setting an update tap.

  This apparatus has two adaptive algorithm execution units (a first adaptive algorithm execution unit and a second adaptive algorithm execution unit), and newly sets an update tap as follows. First, the adaptive algorithm is executed by the second adaptive algorithm execution unit. A filter tap having a converged filter coefficient larger than a predetermined threshold is extracted as the second filter tap. Next, in the first adaptive algorithm execution unit, for the filter tap having the same tap number as the previously extracted second filter tap, the filter coefficient is allowed to be updated, and for the other filter taps The filter coefficient is fixed at 0 so that updating is not allowed. Thereafter, this is repeated every 1,000 samples.

  A specific configuration of the apparatus will be described with reference to FIG.

  This apparatus removes noise components from the input signal x (n) and outputs a signal y (n) from which the noise components have been removed.

  In this apparatus, a limited ADF (R-ADF) is provided in the upper stage, and an ADF is provided in the lower stage. A first adaptive algorithm execution unit is configured by the upper limited ADF (R-ADF), and a second adaptive algorithm execution unit is configured by the upper limited ADF (R-ADF) and the lower ADF.

  The upper limited ADF (R-ADF) and the lower ADF have the same number of filter taps and unit delay time.

  The same signal is input to the upper limited ADF (R-ADF) and the lower ADF.

  The output signal v (n) of the upper limited ADF (R-ADF) is added to the output signal of the lower ADF. Then, using the signal e (n) obtained by subtracting the added signal w (n) from the input signal x (n) as an error signal, the adaptive algorithm is executed in the lower ADF.

  In the upper limited ADF (R-ADF), filter coefficient updating is executed only with a limited tap.

  In the lower ADF, in the tap having the same tap number as the tap whose filter coefficient is updated in the upper limited ADF (R-ADF), the filter coefficient remains fixed at 0 and is not updated. The filter coefficient is updated only at the tap.

  Therefore, the error signal e (n) is a pseudo error signal similar to that when the filter coefficients are updated in all taps of the lower ADF.

  The updated tap coefficients of the upper limited ADF (R-ADF) and the lower ADF are monitored, and the average value of the updated tap coefficients is set as a threshold value for each 1,000 samples, and the threshold value is exceeded. Extract taps with values. Then, the upper limited ADF (R-ADF) tap having the same tap number as the extracted tap is newly set as an update tap.

  In this way, an update tap for the upper limited ADF (R-ADF) is newly set for every 1,000 samples.

  The apparatus configuration for setting a new update tap has been described above with reference to FIG.

  Next, another apparatus configuration for newly setting an update tap will be described with reference to FIG.

  FIG. 58 shows an apparatus configuration for newly setting an update tap.

  This apparatus also has two adaptive algorithm execution units (a first adaptive algorithm execution unit and a second adaptive algorithm execution unit).

  This device also has a limited ADF (R-ADF) in the upper stage and an ADF in the lower stage, but the first adaptive algorithm execution unit is configured by the upper limited ADF (R-ADF), and the lower ADF A second adaptive algorithm execution unit is configured.

  The apparatus of FIG. 58 is different from the apparatus of FIG. 57 in the following three points. The first is that the output signal v (n) of the upper limited ADF (R-ADF) is not added to the output signal of the lower ADF. The second point is that the filter coefficient is updated with all taps in the lower ADF. The third point is that the tap number of the tap to be set as the next update tap is extracted only from the tap of the lower ADF. In other respects, the apparatus of FIG. 58 is similar to the apparatus of FIG.

  58 can also set a new update tap.

  The apparatus configuration for setting a new update tap has been described above with reference to FIG.

(Summary)
From the above, it was confirmed that LES (R-LES) using limited ADF is effective in reducing periodic noise, and that a technique for detecting a peak in R-LES is effective.

[5. Embodiment D]
FIG. 59 shows an output signal when the signal of FIG. 4 is given to LES. However, the number of delays was 49, the number of taps was 1,000, and the step size was 0.05. When this signal was evaluated, VE = 4.16 dB and SNRout = 15.88 dB. When this signal was actually auditioned, the sound was distorted and the reverberation of the sound was also confirmed.

(ADF with limited update taps)
FIG. 60 shows LES tap coefficients when only the noise shown in FIG. 3 is input. Referring to FIG. 60, the tap coefficient of LES takes a peak value for each noise period, and takes a small value for the rest. Therefore, it was confirmed that the sound quality evaluation value is improved while the periodic noise reduction performance remains the same by limiting the taps of the ADF to be updated with respect to the LES to only taps having a peak value in the ADF. Hereinafter, the ADF that limits the tap to be updated is referred to as a limited ADF, and the LES that uses the limited ADF is referred to as an R-LES (Restricted-LES). Also, a tap that has a peak in the ADF, that is, a tap having the same signal as the input signal in a signal having perfect periodicity is referred to as a peak tap.

  FIG. 61 shows tap coefficients when the noise shown in FIG. 3 is input to R-LES. At this time, the number of delays was 49, the number of taps was 1,000, and the step size to be updated was very small compared to LES, so it was necessary to reduce the number to 0.0005. FIG. 62 shows an output signal when the signal shown in FIG. 4 is input. When this signal was evaluated, VE = 15.08 dB and SNRout = 31.37 dB. In addition, when actually listening to the sample, the noise was sufficiently reduced, the sound was hardly damaged, and no echo was felt.

(Application to actual noise: Application to MRI noise)
Apply R-LES to actual noise. The MRI noise shown in FIG. 63 is used as the actual noise. It has been confirmed that this noise has a period of 400 samples and 3,200 samples when the sampling period is (1 / 8,000) s. As an input signal used for the simulation, the signal shown in FIG. 65 obtained by synthesizing the signal of FIG. 63 and the sound of FIG.

  As a method for reducing MRI noise, a noise resynthesis method shown in FIG. 64 has already been proposed. In FIG. 64, LPF indicates a linear prediction filter, ADF indicates an adaptive filter, and a learning identification method is used to update the coefficients of both filters. FIG. 66 shows an output signal when the signal of FIG. 65 is input to the noise resynthesis method. At this time, the number of LPF taps was 64, the step size was 0.1, the number of ADF taps was 128, and the step size was 0.01. When this signal was evaluated, VE = 3.99 dB and SNRout = 3.77 dB. In addition, when actually listening, it was confirmed that periodic noise was found and that the sound was damaged. FIG. 67 shows an output signal when only noise is input. When this signal was evaluated, SNRout = 2.44 dB. In FIG. 67, periodic noise can be confirmed.

  As a technique for reducing this periodic noise, an adaptive line enhancer has been proposed. However, when this method is used, there is a problem in that an audio echo is output. Therefore, R-LES is used as a technique for reducing noise without generating echo of this voice. Since R-LES has limited values of taps and is difficult to output voice echoes, it is considered that periodic noise can be reduced while suppressing voice echoes.

  Therefore, a method of inputting the output signal of the noise resynthesis method to the R-LES is considered. That is, the periodic noise reduction device (R-LES) as described above and a device for reducing noise based on a noise resynthesis method using a linear prediction filter are connected in series. It is considered that the sound is damaged in the output signal of the noise resynthesis method because the step size of the adaptive filter is large, so the step size of the adaptive filter of the noise resynthesis method is set to a small 0.001 Suppress the damage. FIG. 68 shows an output signal when only the noise of FIG. 63 is input to the noise resynthesis method at this time.

  Next, a signal obtained by synthesizing the voice of FIG. 2 with an S / N ratio of 0 dB is used as an input signal of the R-LES. At this time, in order to correspond to a period of 3,200 samples of noise, the tap coefficient of the limited ADF is set to 4,000, the step size is set to 0.0005, and the delay number is set to 100. In addition, since the reduction target is real noise, it is considered that fluctuations have occurred in the period of the noise, unlike the noise used previously. Therefore, in order to cope with the fluctuation, in the subsequent simulations, not only the peak taps are updated, but the peak taps and the 10 taps before and after the taps are updated.

  In addition, consider a new evaluation index. VE and SNRout are indices for evaluating R-LES input / output signals. Therefore, as an evaluation formula that expresses the amount of noise reduction in the R-LES output signal from the noise resynthesis method input signal.

Is introduced. Here, K is the total number of samples, K = 50,000, x _n is the input noise to the noise resynthesis method, and y _n is the output obtained by inputting x _n to the noise resynthesis method and is input to R-LES. When the output noise. In the above evaluation formula, the larger the value, the higher the performance.

  The output signal of R-LES at this time is shown in FIG. When this signal was evaluated, VE = 13.93 dB, SNRout = 111.29 dB, and NR = 10.91 dB. When actually listening, it can be confirmed that the sound is clearly heard compared to the noise resynthesis method. As for noise, although the evaluation value is improved as compared with the noise resynthesis method, periodic noise can be confirmed from FIG.

(Corresponding to long-period noise)
In the method in which the output of the noise resynthesis method is input to the R-LES, the period of periodic noise that can be confirmed in the output signal is 3,200 samples. Since the number of taps of R-LES is 4,000, the fact that only one peak corresponding to a long period of 3,200 samples is updated is considered to be a cause of outputting long-period noise. Therefore, consider increasing the number of taps of R-LES. Here, since the MRI noise has a cycle of 400 samples and 3,200 samples, and 10 taps are updated before and after each peak, the number of taps to be updated is 210 when the number of taps is 4,000. Accordingly, when the number of taps is simply increased, the characteristics of R-LES voice reverberation and reduced calculation amount are lost.

  Therefore, in order to make use of the characteristics of R-LES, a method is used in which a peak is obtained every 400 taps up to 4,000 taps and a peak is taken every 3,200 taps after 4,000 taps. Here, the 400 sample period is the first period, and the 3,200 sample period is the second period. The second period is larger than the first period. Only for signals that substantially match the periodic timing of the first period in the time before 4,000 taps, and signals that substantially match the periodic timing of the second period in the entire tap range. , Update tap is set. Thereby, the tap corresponding to the noise of a 3,200 sample period can be updated, and the increase in the number of taps can be suppressed to the minimum. However, similarly to the simulation before increasing the number of taps, the 10 taps before and after each peak are also set as update taps.

  FIG. 70 shows an output signal when this method is used, the number of taps is 20,000, and the output signal of R-LES is used as the input signal of the noise resynthesis method. Further, FIG. 71 shows the tap coefficients at this time. When this signal was evaluated, VE = 15.31 dB, SNRout = 11.92 dB, and NR = 11.54 dB. From the evaluation value, it can be confirmed that the sound quality is slightly improved, but the noise reduction effect is not improved. Furthermore, even if it sees FIG. 70 and it actually listens to the signal of FIG. 70, long-period noise is not reduced. From this, the noise reduction effect could not be improved because the long-period noise could not be reduced, and it is considered that the noise reduction performance is improved by reducing the long-period noise.

  FIG. 71 shows tap coefficients of the limited ADF at this time. In FIG. 71, the coefficient up to about 10,000 taps has a large value. From this, the tap that peaks for noise of 3,200 sample periods up to about 10,000 taps has peaks at about 10 taps before and after an integer multiple of 3,200, but fluctuations occur after that. For this reason, it is considered that taps deviating from 10 taps before and after an integer multiple of 3,200 have peaks. Therefore, in order to cope with long-period noise fluctuations, 100 taps before and after the peak with respect to noise of 3,200 sample periods are set as update taps.

  FIG. 72 shows the output signal at this time, and FIG. 73 shows the tap coefficient. When this signal was evaluated, VE = 8.37 dB, SNRout = 12.84 dB, and NR = 12.39 dB. Although the value of VE is lower than the method of updating 10 taps before and after each peak, it is improved compared to when the number of taps is 4,000. I can't feel it. Furthermore, it can be confirmed that long-period noise can be reduced by looking at FIG. 72 or actually listening. From this, it was confirmed that long-period noise fluctuated greatly, and that the reduction could be dealt with by increasing the number of update taps for the peak.

(Summary)
From the above, it was confirmed that the combination of the periodic noise reduction system (R-LES) using the limited ADF and the noise resynthesis method has an effect on the reduction of MRI noise that is actual noise. . In addition, it was confirmed that it is effective to set a peak for each period for noise having a plurality of periods. Furthermore, it was confirmed that fluctuations in the cycle of noise in actual noise can be handled by setting taps that are updated before and after the peak tap.

[6. Embodiment E]
FIG. 74 shows a periodic noise reduction device 10 according to an embodiment of the present invention. The periodic noise reduction device 10 has the simplest configuration as the periodic noise reduction device according to the present invention.

  The periodic noise reduction device 10 includes an input terminal 11, an output terminal 12, a subtracter 13, and an adaptive FIR filter 20.

  The adaptive FIR filter 20 includes a main signal input unit 21, a signal output unit 22, an error signal input unit 23, delay elements d1 to d15, filter taps t2, t7, t12, an adder 24, and a coefficient update. Part 25.

  An audio signal Sin is input to the input terminal 11. The signal Sin input to the input terminal 11 is sent to the main signal input unit 21 of the adaptive FIR filter 20 and the subtractor 13. The signal Sc output from the signal output unit 22 of the adaptive FIR filter 20 is sent to the subtractor 13. The subtracter 13 subtracts the signal Sc from the signal Sin and outputs the subtraction result as the signal Sout. The signal Sout output from the subtractor 13 is sent to the output terminal 12 and the error signal input unit 23 of the FIR adaptive filter 20. A signal Sout is output from the output terminal 12 as an output signal of the periodic noise reduction device 10.

  The FIR adaptive filter 20 performs the following signal processing. That is, the signal Sin input to the main signal input unit 21 is sent to the coefficient update unit 25 as a main input signal and is sequentially delayed by the delay elements d1 to d15 arranged in series. The delay elements d1 to d15 are continuously connected in series.

Filter coefficients h ₂ , h ₇ , and h ₁₂ are set for the filter taps t 2, t 7, and t 12, respectively.

A signal delayed by the delay elements d1 to d2 is output from the delay element d2. In the filter tap t2, the filter coefficient h ₂ of the signal output from the delay element d2 is multiplied, the multiplication result is sent to the adder 24.

A signal delayed by the delay elements d1 to d7 is output from the delay element d7. At the filter tap t ₇ , the signal output from the delay element d ₇ is multiplied by the filter coefficient h ₇ , and the multiplication result is sent to the adder 24.

From the delay element d12, signals delayed by the delay elements d1 to d12 are output. In the filter taps t12, the filter coefficient h ₁₂ to the signal outputted from the delay element d12 multiplied, the multiplication result is sent to the adder 24.

  The adder 24 adds all the signals sent thereto. The signal Sc as a result of the addition is sent to the signal output unit 22 and is further sent from the signal output unit 22 to the subtracter 13.

  The transfer function H (z) in the system from the main signal input unit 21 to the signal output unit 22 is expressed by the following equation (6).

The coefficient update unit 25 receives the input signal Sin and the signal Sout generated by the subtractor 13 as an error signal via the error signal input unit 23. Coefficient updating unit 25, based on the signal Sin and the signal Sout, to update the filter coefficients h _2, h _7, h ₁₂ of the filter tap t2, t7, t12. The coefficient updating unit 25 updates the filter coefficients h ₂ , h ₇ , and h ₁₂ using an adaptive algorithm so that the power of the signal Sout is minimized. In the coefficient updating unit 25, a learning identification method is used as an adaptive algorithm.

The delay element “Z ^−N ” shown in FIG. 1 is for adjusting the delay time, but is shown in FIG. 1 by the delay element d1 and the delay element d2 shown in FIG. It can also be considered that a member corresponding to the delay element “Z ^−N ” is formed. Further, in the periodic noise reduction device 10 of FIG. 74, the delay element “shown in FIG. 1 is interposed between the branch point 14 where the signal Sin input to the input terminal 11 branches and the main signal input unit 21. A member corresponding to “Z ^−N ” may be provided.

The input signal Sin is an audio signal including periodic noise. If the period of this periodic noise is Tn and the delay time of each of the delay elements d1 to d15 is Td, the relationship of “Tn = 5 × Td” is established. (“Td” coincides with the sampling period “Ts”.)
Various delay times (0 × Td, 1 × Td, 2 × Td,..., 14 × Td, 15 × Td) are generated by the 15 delay elements d1 to d15. The minimum is “0” and the maximum is “15 × Td”. The delay time axis can be conceptualized from various delay times caused by the delay elements d1 to d15.

  On such a delay time axis, a reference delay time can be determined, and a delay time matching the timing for each Tn can be determined. If a delay element that outputs a signal with such a determined delay time is selected, the following is obtained.

  For example, suppose that (Td × 2) is defined as the reference delay time. A delay element that outputs a signal delayed by (Tn × 0) from the reference delay time is a delay element d2. The delay element d7 that outputs a signal delayed by (Tn × 1) from the delay element d2 is the delay element d7. The delay element d12 is a delay element that outputs a signal delayed by (Tn × 2) from the delay element d2.

In the adaptive FIR filter 100 of FIG. 74, filter taps t2, t7, t12 that can update filter coefficients are provided at locations corresponding to the output signals from the delay elements d2, d7, d12. each filter coefficient h _2, h _7, h ₁₂ is multiplied by the output signal from the delay element d2, d7, d12, their multiplication result is sent to the adder 24. On the other hand, no filter tap is provided for other delay elements.

  As described above, the adaptive FIR filter 20 is an adaptive FIR filter that performs limited ADF. As described above, the limited ADF is effective in reducing periodic noise. Therefore, the periodic noise reduction device 10 is effective for reducing periodic noise having a period of Tn.

In the adaptive FIR filter 20 of FIG. 74, no filter tap is provided for delay elements other than the delay elements d2, d7, d12. However, a filter tap is also provided for these delay elements. It can also be done. It is also possible to make the filter tap t0 for multiplying the filter coefficients h ₀ to the signal input to the delay element d1, further provided. In this case, the transfer function of the adaptive FIR filter is as shown in the following equation (7).

However, filter coefficients other than the filter coefficients h ₂ , h ₇ , and h ₁₂ are not updated while their absolute values are set to 0 or close to them. In effect, the adaptive FIR filter then performs limited ADF.

Further, the filter coefficients other than the filter coefficients h ₂ , h ₇ , and h ₁₂ may be updated, but their absolute values may be updated only to a value close to 0. In other words, the filter coefficients substantially remain set to zero or close to their absolute values. In other words, it can be said that only the filter coefficients h ₂ , h ₇ , and h ₁₂ are substantially updatable, and the other filter coefficients are substantially not updatable. Even in this case, the adaptive FIR filter substantially performs limited ADF.

FIG. 75 is a diagram illustrating a range of updatable values of the filter coefficient of the adaptive FIR filter 20. The horizontal axis of FIG. 75 indicates the number of the filter tap, and the vertical axis indicates the updatable range of the filter coefficient. As can be understood from FIG. 75, the updatable range of the filter coefficient h ₂ is “0 or more and 0.5 or less”, and the updatable range of the filter coefficient h ₇ is “0 or more and 0.35 or less”. The updatable range of the coefficient h ₁₂ is “0 or more and 0.25 or less”.

The maximum values that each filter coefficient h ₂ , h ₇ , and h ₁₂ can take are 0.5, 0.35, and 0.25, which are exponentially attenuated. As described above, giving the coefficient to the exponential decay type in this way is effective in reducing the periodic noise.

In addition, the maximum values that each filter coefficient h ₂ , h ₇ , h ₁₂ can take are 0.5, 0.35, and 0.25, and the sum is “1.1”. Thus, the sum of the maximum values of the filter coefficient updatable range of the filter tap that can update the filter coefficient is close to “1”. In this way, it has been experimentally confirmed that it is particularly effective for reducing periodic noise.

The minimum values that can be taken by the filter coefficients h ₂ , h ₇ , h ₁₂ are all zero. This is done so that these filter coefficients are not updated to a negative value having a large absolute value. That is, if these filter coefficients are configured so that the absolute value can be updated to a large negative value, the following problem may occur. In other words, when the value of a certain filter coefficient is updated to a negative value with a large absolute value, other filter coefficients are updated to a positive value with a very large absolute value as if to compensate for it. Because there are things. As described above, it has been experimentally confirmed that, in the filter coefficient of the limited ADF, when a large absolute value appears in both the positive and negative regions, an echo is generated.

FIG. 76 is a diagram showing a range of updatable values of filter coefficients of the adaptive FIR filter of the periodic noise reduction device according to the present invention. This periodic noise reduction device has basically the same configuration as the periodic noise reduction device 10 of FIG. 74, but the adaptive FIR filter includes 199 delay elements d1 to d199 and output signals of these delay elements. filter tap t1~t199 199 for multiplying the filter coefficients are arranged, it is arranged filter tap t0, for multiplying the filter coefficients h ₀ against further input signal to the delay element d1 respect. Each of these delay elements produces a delay time of (1/8000) seconds. That is, the sampling frequency is 8000 Hz. Here, a delay time (= (1/8000) seconds) generated by one delay element is defined as Td.

  The horizontal axis of FIG. 76 shows the number of the filter tap, and the vertical axis shows the updatable range of the filter coefficient. The filter tap number on the horizontal axis also corresponds to the delay time. That is, for example, the signal output from the delay element d80 is input to the filter tap t80. The signal output from the delay element d80 is a signal delayed by (80 × Td), and the signal is input to the filter tap t80. Then, it can be said that the horizontal axis of FIG. 76 is a time axis indicating the delay time normalized by Td.

  FIG. 76 shows the updatable range of the filter coefficient. As shown here, the filter update is performed so as to be effective in reducing the periodic noise whose period Tn is (1/100) second. The possible filter taps are limited. That is, only the filter coefficients of the filter taps t0, t80, and t160 can be updated, and the filter coefficients of the other filter taps are maintained at 0 and cannot be updated.

  Since the delay time Td generated by one delay element is (1/8000) second and the period Tn of periodic noise is (1/100) second, the period Tn of periodic noise is equivalent to 80 taps. To do. For this reason, the filter coefficient capable of updating the filter coefficient is set every 80 taps.

  Further, in order to reduce periodic noise having a period corresponding to 80 taps, it is only necessary to set a filter coefficient updatable for every 80 taps. As shown in FIG. 76, it may be as shown in FIG.

  FIG. 77 shows the updatable range of the filter coefficient. Here, only the filter coefficients of the filter taps t20, t100, and t180 can be updated, and the filter coefficients of the other filter taps are maintained at 0 and updated. I can't do it.

  In FIG. 77, the maximum value that can be taken by the filter coefficient is attenuated toward the subsequent stage. The attenuation pattern is an exponential attenuation pattern.

  However, the update range of the filter coefficient may be determined so that the attenuation pattern is an attenuation pattern corresponding to the second half of the Hamming window. FIG. 78 shows such an updatable range of the filter coefficient.

  Further, the filter coefficient updatable range may be determined so that the maximum value that the filter coefficient can take extends from the first filter tap to the last filter tap and indicates a pattern corresponding to a Hamming window. FIG. 79 shows such an updatable range of the filter coefficient.

  As shown in FIG. 77, only the filter coefficients of the filter taps t20, t100, and t180 can be updated, and the filter coefficients of the other filter taps are maintained at 0 and cannot be updated. However, not only the filter taps t20, t100 and t180 but also the filter taps in the vicinity thereof can be set so that the filter coefficient can be updated.

  FIG. 80 shows the updatable range of the filter coefficient. As shown here, the filter coefficients of the filter taps t20, t100, t180 and the neighboring filter taps, for example, the filter taps before and after the filter tap are shown. It can also be updatable. Even if it does in this way, it is effective in reducing the periodic noise of a period of 100 Hz. Here, the updatable range of the filter coefficient of the filter tap t20 is “0 to 0.5”, the updatable range of the filter coefficient of the filter tap t100 is “0 to 0.35”, and the filter coefficient of the filter tap t180 is The updatable range is “0 to 0.25”, and the updatable range of filter coefficients of the filter taps t19, t21, t99, t101, t179, and t181 is “−0.05 to 0.05”.

  As shown in FIG. 77, setting the filter coefficient updatable range is effective in reducing periodic noise with a period of 100 Hz, but for noise that includes two periodic noises of 100 Hz and 80 Hz. In order to make the reduction effect effective, an updatable range of the filter coefficient can be determined as shown in FIG. In this way, the filter taps t20, t100 and t180 contribute to the reduction of 100 Hz periodic noise, and the filter taps t50 and t150 contribute to the reduction of 80 Hz periodic noise.

  As shown in FIG. 77, when the updatable range of the filter coefficient is determined, it is effective in reducing periodic noise with a period of 100 Hz, but the reduction is reduced with respect to periodic noise whose period is 120 Hz. In order to make the effect effective, as shown in FIG. 82, an updatable range of the filter coefficient may be determined. As understood from FIG. 82, only the filter coefficients of the filter taps t20, t86, t87, t153, and t154 can be updated here. This is due to the following reason. That is, the period Tn of the periodic noise of 120 Hz corresponds to 66.7 taps. When the filter tap t20 is used as a reference, there is no filter tap at a location 66.7 taps away from the filter tap t20, and therefore the filter taps whose filter coefficients can be updated in the neighboring (previous and subsequent) filter taps t86 and t87. It was determined as The filter taps t153 and t154 are determined as filter taps capable of updating the filter coefficient for the same reason.

  75 to 82 described above show the filter coefficient updatable range for the adaptive FIR filter having a relatively small number of taps for easy understanding. However, in the periodic noise reduction device of the present invention, an adaptive FIR filter having a larger number of taps can be used. Examples thereof are shown in FIGS.

  FIG. 83 is a diagram showing a filter coefficient updateable range in an adaptive FIR filter having a total of 1200 taps t0 to t1199 from No. 0 to No. 1199. Only 15 filters out of 1200 are set so that the coefficients can be updated. In these filter taps in which the filter coefficient can be updated, the maximum value that can be taken by the filter coefficient is attenuated toward the subsequent stage, and the attenuation pattern is an exponential attenuation pattern.

  FIG. 84 is also a diagram showing a filter coefficient updateable range in an adaptive FIR filter having a total of 1200 taps t0 to t1199 from 0 to 1199. Again, only 15 filters out of 1200 are set so that the coefficients can be updated. In these filter taps in which the filter coefficient can be updated, the maximum value that can be taken by the filter coefficient is attenuated toward the later stage, and the attenuation pattern is an attenuation pattern corresponding to the second half of the Hamming window.

  In the above, the periodic noise reduction apparatus which is one embodiment of this invention was demonstrated. In the above, the periodic noise reduction apparatus using the adaptive FIR filter in which the learning identification method is executed as an adaptive algorithm has been mainly shown. However, the adaptive algorithm is not limited to the learning identification method, and other algorithms can be applied. For example, an LMS algorithm may be used.

(Application to systems including acoustic systems)
Various filters that can be used in the periodic noise reduction device have been described above. These filters can also be used for systems including acoustic systems.

  FIG. 85 is a diagram showing a configuration when the limited ADF (R-ADF) is applied to a system including an acoustic system.

In this system, a noise source NS that emits periodic noise exists near the left end of the duct. The duct has a branch path, and a speaker S is provided at the lower end of the branch path.
Inside the duct, a microphone M1 is installed near the noise source NS. A microphone M2 is installed in the duct at a position to the right of the branch point. The limited ADF (R-ADF) receives the output signal of the microphone M1 as a main input signal and the output signal of the microphone M2 as an error signal. Further, the output signal of the limited ADF (R-ADF) is sent to the speaker S. The sound radiated from the speaker S interferes with periodic noise generated by the noise source NS.

  When the adaptive algorithm is executed in the limited ADF (R-ADF), a sound wave that cancels the periodic noise from the noise source NS is emitted from the speaker S.

  According to the present invention, the noise reduction performance is improved, which is useful in the technical field of electroacoustics, for example.

It is a schematic block diagram of the periodic noise reduction system by the principle of a line enhancer. It is a figure which shows a male audio | voice signal. It is a figure which shows periodic noise. It is a figure which shows an input signal. It is a figure which shows the output signal of LES. It is a figure which shows the tap coefficient value of ADF in LES. It is a figure which shows the number of taps which take the peak value to update. It is a figure which shows a tap coefficient value. It is a figure which shows an output signal. It is a figure which shows a tap coefficient value. It is a figure which shows an output signal. It is a figure which shows the output signal in the method of limiting an update tap. It is a figure which shows MRI noise. It is a figure which shows an input signal. It is a figure which shows a LES output signal. It is a figure which shows the tap coefficient value of LES. It is a figure which shows an output signal. It is a figure which shows an output signal. It is a figure which shows a tap coefficient value. It is a figure which shows the tap coefficient of limited ADF. It is a figure which shows the output signal of limited ADF. It is a figure which shows the relationship between the number of taps and VE. It is a figure which shows an output signal. It is a spectrum figure of a tap coefficient. It is a figure which shows a tap coefficient (Humming window). It is the spectrum figure of the tap coefficient which applied the Hamming window. It is a figure which shows the relationship between the number of taps and VE. It is a figure which shows an output signal. It is a figure which shows a tap coefficient (Hamming window by 2 times the number of taps). It is the spectrum figure of the tap coefficient which applied the half of the Hamming window. It is a figure which shows an output signal. It is a figure which shows the relationship between the number of taps and VE. It is a figure which shows a tap coefficient (exponential decay type r = 0.9). It is a spectrum figure of the tap coefficient of an exponential decay type (r = 0.9). It is a figure which shows an output signal. It is a figure which shows the relationship between the coefficient r and VE. It is a schematic block diagram of an IIR type periodic noise reduction system. It is a figure which shows the relationship between the coefficient a and VE. It is a figure which shows an output signal (a = 0.99). It is a figure which shows the tap coefficient of ADF in LES. It is a diagram illustrating the tap coefficients of the coefficient values a predetermined type FIR filter (f ₀ = 160Hz). It is a diagram showing coefficient values a certain type of system output signal (f ₀ = 160Hz). A noise signal waveform diagram, (a) shows the cyclic noise f ₀ = 120Hz, (b) is an enlarged view of the noise. It is a figure which shows an input signal. Is a diagram illustrating the tap coefficients of the coefficient values a predetermined type CPPS (f ₀ = 120Hz). Is a diagram showing coefficient values a certain type CPPS output signal (f ₀ = 120Hz). It is a figure which shows the tap coefficient of limited ADF of the conventional method. It is a figure which shows the output signal of the conventional R-LES. It is a figure which shows the tap coefficient of improved limited ADF. It is a figure which shows the output signal of improved R-LES. It is a figure which shows the tap coefficient of peak detection type | mold limited ADF. It is a figure which shows the output signal of a peak detection type | mold R-LES. It is a figure which shows the tap coefficient (update continuation) of limited ADF. It is a figure which shows an output signal (update continuation). It is a figure which shows the coefficient update amount of a tap, (a) shows the coefficient update amount of a 200th tap, (b) shows the enlarged view. It is a figure which shows an output signal (with clip processing). It is a figure which shows the apparatus structure for setting an update tap newly. It is a figure which shows the apparatus structure for setting an update tap newly. It is a figure which shows the output signal of LES. It is a figure which shows the tap coefficient of ADF. It is a figure which shows the tap coefficient of limited ADF. It is a figure which shows the output signal of R-LES. It is a figure which shows a noise waveform, (a) shows MRI noise, (b) is the enlarged view (of MRI noise). It is a block diagram of the system which performs a noise resynthesis method. It is a figure which shows an input signal (MRI noise). It is a figure which shows the output signal of a noise resynthesis method. It is a figure which shows the output signal (only a noise) of a noise resynthesis method. It is a figure which shows the noise output (ADF step size = 0.001) of a noise resynthesis method. It is a figure which shows the output signal of noise resynthesis method + R-LES. It is a figure which shows the output signal of noise resynthesis method + R-LES (the number of taps = 20,000). It is a figure which shows the tap coefficient of noise resynthesis method + R-LES (the number of taps = 20,000). It is a figure which shows the output signal of noise resynthesis method + R-LES (the number of taps = 20,000). It is a figure which shows the tap coefficient of noise resynthesis method + R-LES (the number of taps = 20,000). It is a schematic block diagram of a periodic noise reduction apparatus. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows the updateable range of the filter coefficient of an adaptive FIR filter. It is a figure which shows a structure when limited ADF (R-ADF) is applied to the system containing an acoustic system.

Explanation of symbols

Periodic noise reduction device 10
Input terminal 11
Output terminal 12
Subtractor 13
Adaptive FIR filter 20
Main signal input section 21
Signal output unit 22
Error signal input unit 23
Adder 24
Coefficient update unit 25

Claims (32)

A digital filter used in a periodic noise reduction device,
A digital filter whose impulse response has a coefficient value that is neither 0 nor a value in the vicinity thereof only at a time substantially matching the periodic timing of the periodic noise.   The digital filter according to claim 1, wherein a sum of coefficient values over the impulse response is approximately 1. IIR type,
With a feedback loop,
Having a delay in the feedback loop;
The digital filter according to claim 1 or 2, wherein a delay time generated by the delay unit is substantially the same as a period of the periodic noise.   The digital filter according to claim 3, wherein the digital filter shows an attenuation pattern such that a coefficient value at a time substantially coincident with a periodic timing of the periodic noise becomes smaller as a subsequent stage.   The digital filter of claim 4, wherein the attenuation pattern is an exponential attenuation pattern. FIR type,
A plurality of delay elements connected in series and a plurality of filter taps;
A filter which is neither 0 nor a value in the vicinity thereof only for a signal whose time substantially matches the periodic timing of the periodic noise on the time axis of the delay time generated by a plurality of delay elements. The digital filter according to claim 1, further comprising a filter tap in which a coefficient is set.   The digital filter according to claim 6, wherein the digital filter shows an attenuation pattern such that a coefficient value at a time substantially coincident with a periodic timing of the periodic noise becomes smaller as a subsequent stage.   The digital filter of claim 7, wherein the attenuation pattern is an exponential attenuation pattern.   The digital filter according to claim 7, wherein the attenuation pattern is an attenuation pattern corresponding to the second half of the Hamming window.   The digital filter according to claim 6, wherein a coefficient value at a time substantially coincident with a periodic timing of the periodic noise indicates a pattern corresponding to a Hamming window. A digital filter used in a periodic noise reduction device,
Is adaptive,
A plurality of delay elements connected in series and a plurality of filter taps capable of updating the filter coefficient;
The sound signal of the periodic noise is used as a main input signal, and the filter coefficient of the filter tap is updated so that the power of the error signal between the output signal of the digital filter and the sound signal is minimized,
On the time axis of the delay time generated by a plurality of delay elements, there is a filter tap that can substantially update the filter coefficient only for a signal having a time substantially matching the periodic timing of the periodic noise. Digital filter provided.   No filter tap is provided for an output signal of a delay element other than the delay element in which the filter coefficient is multiplied to the output signal by a filter tap capable of substantially updating the filter coefficient, or 0 or The digital filter according to claim 11, further comprising a filter tap for multiplying a value in the vicinity thereof as a filter coefficient.   The digital filter according to claim 11 or 12, wherein the adaptive algorithm executed in the digital filter is a learning identification method.   The filter tap capable of substantially updating the filter coefficient, wherein the total value of the maximum values of the filter coefficient updatable range is 1 or a value in the vicinity thereof. The digital filter described.   The digital filter according to any one of claims 11 to 14, wherein the filter coefficient is limited to 0 or a positive value in a filter tap capable of substantially updating the filter coefficient.   The filter tap capable of substantially updating the filter coefficient shows an attenuation pattern in which the maximum value of the filter coefficient updatable range becomes smaller as the subsequent filter tap. The digital filter according to item.   The digital filter of claim 16, wherein the attenuation pattern is an exponential attenuation pattern.   The digital filter according to claim 16, wherein the attenuation pattern is an attenuation pattern corresponding to the second half of the Hamming window.   The digital filter according to any one of claims 11 to 15, wherein in the filter tap capable of substantially updating the filter coefficient, the maximum value of the filter coefficient updateable range indicates a pattern corresponding to a Hamming window. filter. The periodic noise includes a noise component having a first period and a noise component having a second period;
The second period is greater than the first period;
On the time axis of the delay time generated by a plurality of delay elements, for a signal that substantially matches the periodic timing of the first period and a signal that substantially matches the periodic timing of the second period The digital filter according to any one of claims 11 to 15, wherein a filter tap capable of substantially updating a filter coefficient is provided. The periodic noise includes a noise component having a first period and a noise component having a second period;
The second period is greater than the first period;
A time substantially matching the periodic timing of the first period and a delay time generated by the plurality of delay elements at a time prior to a predetermined time on the time axis of the delay time generated by the plurality of delay elements. 16. A filter tap capable of substantially updating a filter coefficient is provided only for a signal having a time substantially matching the periodic timing of the second period on the time axis. The digital filter according to any one of the items. A digital filter used in a periodic noise reduction device,
Is adaptive,
In the digital filter, the first step, the second step and the third step are executed,
In the first step, filter coefficients are updated in all filter taps of the digital filter,
In the second step, a filter tap having a filter coefficient that is converged by executing the first step and is larger than a predetermined threshold is extracted as the first filter tap,
In the third step, the filter coefficient is allowed to be updated for the first filter tap extracted by executing the second step, and the filter coefficient is 0 or in the vicinity thereof for the other filter taps. A digital filter that is set to the value of and the filter coefficient is not allowed to be updated.   The digital filter according to claim 22, wherein the adaptive algorithm executed in the digital filter is a learning identification method.   The digital filter according to claim 22 or 23, wherein in the third step, the total value of the maximum values of the filter coefficient updatable range of the first filter tap is 1 or a value in the vicinity thereof.   The digital filter according to any one of claims 22 to 24, wherein in the third step, a filter coefficient of the first filter tap is limited to 0 or a positive value. A digital filter used in a periodic noise reduction device,
Is adaptive,
A first adaptive algorithm execution unit and a second adaptive algorithm execution unit;
The first adaptive algorithm execution unit and the second adaptive algorithm execution unit have the same number of filter taps and unit delay time,
The main input signal input to the first adaptive algorithm execution unit is also input as the main input signal to the second adaptive algorithm execution unit,
A filter tap whose filter coefficient converged by executing the adaptive algorithm in the second adaptive algorithm execution unit is larger than a predetermined threshold is extracted as the second filter tap,
In the first adaptive algorithm executing unit, the filter coefficient is allowed to be updated for a filter tap having the same tap number as the extracted second filter tap, and for the other filter taps, the filter coefficient Is a digital filter in which 0 is set to 0 or a value in the vicinity thereof, and updating of filter coefficients is not allowed.   27. The digital filter according to claim 26, wherein both the adaptation algorithm executed in the first adaptation algorithm execution unit and the adaptation algorithm executed in the second adaptation algorithm execution unit are learning identification methods.   27. In the adaptive algorithm of the first adaptive algorithm execution unit, the total value of the maximum values of the filter coefficient updatable range at the filter tap in which the filter coefficient is allowed to be updated is 1 or a value in the vicinity thereof. 27. The digital filter according to 27.   29. The adaptive algorithm of the first adaptive algorithm execution unit, wherein a filter coefficient at a filter tap that is allowed to update a filter coefficient is limited to 0 or a positive value according to any one of claims 26 to 28. Digital filter. An audio signal input unit, an audio signal output unit, and a filter;
An audio signal including periodic noise is input to the audio signal input unit,
The audio signal input to the audio signal input unit is input to the filter,
A signal generated by subtracting the output signal of the filter from the audio signal input to the audio signal input unit is output from the audio signal output unit,
A periodic noise reduction device, wherein the filter is the digital filter according to any one of claims 1 to 29. A noise reduction device comprising a first noise reduction device and a second noise reduction device connected in series at the subsequent stage,
One of the first noise reduction device and the second noise reduction device is the periodic noise reduction device according to claim 30,
The other is a noise reduction device that is a device that reduces noise based on a noise resynthesis method using a linear prediction filter. A first microphone, a second microphone, a speaker, and a filter;
The digital filter according to any one of claims 11 to 25, wherein the filter is a digital filter.
Periodic noise is input to the first microphone,
The output signal of the first microphone is input to the filter,
The output signal of the filter is sent to the speaker,
The periodic noise and the sound emitted from the speaker are input to the second microphone,
The periodic noise reduction device, wherein an output signal of the second microphone is input to the filter as an error signal for executing an adaptive algorithm.

Images