Signal processing device, signal processing method, and signal processing program

Abstract

When an observation signal including noise the pitch of which changes with the passage of time has been received, a comb filter selection unit (21) selects, from a plurality of comb filters having different pitches to be removed, a comb filter having a pitch to be removed, which matches the pitch of noise, as a selected comb filter in response to the change of the pitch of the noise with the passage of time. A comb filter application unit (23) applies the selected comb filter to the observation signal to remove noise from the observation signal.

Description

Signal processing device, signal processing method, and signal processing program

[0001] The present disclosure relates to a technique for removing noise contained in an audio signal.

[0002] There is a technology that detects abnormalities in equipment by analyzing audio signals. This technology uses an acoustic sensor such as a microphone to observe the operating sounds of the equipment and determines whether the observed audio signal contains any abnormal sounds.

[0003] However, such technology has the problem of degrading the accuracy of the judgment due to the influence of noise mixed into the observed signal, such as noise from mechanical equipment other than the equipment being inspected (e.g., noise from an air conditioner), environmental noise (e.g., the sound of wind and rain), and electrical noise caused by the measurement.

[0004] The environments where the above inspection technology is required are often machine rooms, production lines, etc. Machine rooms, production lines, etc. are often surrounded by a large amount of machinery. Therefore, in order to properly use the above inspection technology, it is necessary to establish countermeasures against the noise generated by this machinery. Unless such countermeasures can be established, it is difficult to accurately detect abnormalities from the operating sounds of equipment in a real environment.

[0005] In particular, noises generated by mechanical equipment include alarms, chimes, etc. These alarms, chimes, etc. are electronic sounds emitted for the purpose of signals or warnings. Furthermore, these alarms, chimes, etc. often have melodies. Therefore, alarms, chimes, etc. have the characteristic that their pitch frequently changes during playback. For this reason, alarms, chimes, etc. are not stationary. Many noise removal methods that assume the stationarity of noise (e.g., simple spectral subtraction) cannot deal with such non-stationary noise.

[0006] The waveform of such melody noise is periodic within the same interval. Therefore, in the power spectrum domain, the noise waveform appears as several peaks consisting of the fundamental frequency component and its harmonic components. Therefore, if we consider only a single interval, we can remove the noise using a notch filter or comb filter that blocks these peak components.

[0007] Japanese Patent Application Laid-Open No. 2006-331567

[0008] Patent Document 1 discloses an example of such a noise removal technique. Patent Document 1 discloses a technique for removing noise with a periodic waveform using a comb filter. However, with the configuration disclosed in Patent Document 1, in order to accommodate changes in the pitch of the noise, i.e., changes in the period of the noise waveform, it is necessary to separately provide the period by some means. Therefore, in order to accommodate noise that forms a melody, it is necessary to provide the melody pitch at each time by some means. As a result, the technique disclosed in Patent Document 1 has the problem that the configuration for removing noise is complicated.

[0009] The main object of the present disclosure is to solve the above-mentioned problems. Specifically, the main object of the present disclosure is to enable, with a simple configuration, the removal of noise whose pitch changes over time.

[0010] A signal processing device according to the present disclosure includes: a comb filter selection unit that, when an audio signal containing noise whose pitch changes over time is received, selects, as a selected comb filter, a comb filter whose target pitch for removal matches the pitch of the noise from a plurality of comb filters each having a different target pitch for removal in accordance with the change in pitch of the noise over time; and a comb filter application unit that applies the selected comb filter to the audio signal to remove the noise from the audio signal.

[0011] According to the present disclosure, it is possible to remove noise whose pitch changes over time with a simple configuration.

[0012] 1 is a diagram showing an example of a functional configuration of an alarm removal device according to embodiment 1. 2 is a diagram showing an example of a hardware configuration of the alarm removal device according to embodiment 1. 3 is a flowchart showing an example of an operation of the alarm removal device according to embodiment 1. 4 is a flowchart showing an example of an operation of a comb filter selector when a comb filter is applied in the frequency domain by fast Fourier transform according to embodiment 1. 5 is a flowchart showing an example of an operation of a comb filter selector when a comb filter is applied in the frequency domain by fast Fourier transform according to embodiment 1.

[0013] Hereinafter, embodiments will be described with reference to the drawings. In the following description of the embodiments and the drawings, the same reference numerals denote the same or corresponding parts.

[0014] Embodiment 1. *** Overview *** In this embodiment, a configuration will be described for removing noise that changes in pitch over time and is mixed into an observation signal, which is an audio signal obtained from an acoustic sensor. The noise to be removed in this embodiment is composed of multiple discrete pitches, such as alarms and chimes, and different pitches are reproduced over time. The noise to be removed may also be multiple pitches reproduced simultaneously to form a chord.

[0015] In this embodiment, a comb filter is used to remove such noise. A comb filter is a filter that selectively removes noise of a certain pitch. A comb filter has the characteristic of blocking only the fundamental frequency of the pitch to be removed and its harmonic components. A simple comb filter blocks only a single pitch, but it is also possible to combine multiple simple comb filters to obtain a comb filter that blocks multiple pitches. Because such a comb filter blocks multiple pitches simultaneously, it is also possible to remove chords. To remove noise that is part of a melody, an appropriate comb filter can be selected and applied at each time depending on the pitch of the melody.

[0016] In this embodiment, the alarm elimination device 2, which will be described later, stores multiple comb filters with different cutoff characteristics. The alarm elimination device 2 then selects an optimal comb filter from the multiple comb filters based on the observed signal. The alarm elimination device 2 then applies the selected comb filter to the observed signal to output a signal from which noise has been removed. By selecting an optimal comb filter based on the pitch of the noise from the observed signal, it is possible to remove noise whose pitch changes frequently with a simple configuration.

[0017] ***Description of Configuration*** Fig. 1 shows an example of the functional configuration of the alarm removal device 2 according to this embodiment. Fig. 2 shows an example of the hardware configuration of the alarm removal device 2 according to this embodiment. The alarm removal device 2 corresponds to a signal processing device. The operating procedure of the alarm removal device 2 corresponds to a signal processing method. Furthermore, a program that realizes the operation of the alarm removal device 2 corresponds to a signal processing program.

[0018] First, an example of the hardware configuration of the alarm removal device 2 will be outlined with reference to FIG.

[0019] The alarm removal device 2 according to this embodiment is a computer. The alarm removal device 2 includes, as hardware components, a processor 901, a main memory device 902, an auxiliary memory device 903, and an input / output device 904. As shown in FIG. 1 , the alarm removal device 2 also includes, as functional components, a comb filter selector 21, a comb filter storage unit 22, and a comb filter applicator 23. The functions of the comb filter selector 21 and the comb filter applicator 23 are implemented, for example, by programs. The auxiliary memory device 903 stores programs implementing the functions of the comb filter selector 21 and the comb filter applicator 23. These programs are loaded from the auxiliary memory device 903 to the main memory device 902. The processor 901 then executes these programs to perform the operations of the comb filter selector 21 and the comb filter applicator 23, which will be described later. FIG. 2 schematically illustrates a state in which the processor 901 is executing programs implementing the functions of the comb filter selector 21 and the comb filter applicator 23. The comb filter storage unit 22 is implemented, for example, by the auxiliary memory device 903. The input / output device 904 is used to input an observation signal from an external device, the signal input device 1. The input / output device 904 is also used to output a noise-removed signal, which is a signal from which noise has been removed.

[0020] Next, an example of the functional configuration of the alarm removal device 2 will be described with reference to FIG.

[0021] The alarm removal device 2 receives an observation signal from the signal input device 1. As described above, the observation signal is an audio signal containing noise whose pitch changes over time. The alarm removal device 2 removes noise from the observation signal and outputs a noise-removed signal.

[0022] The signal input device 1 is configured with, for example, a microphone and an A / D (Analog / Digital) converter. In the signal input device 1, the A / D converter converts an analog signal observed by the microphone into a digital signal and outputs it to the alarm removal device 2. In the signal input device 1, other than the microphone, for example, an acceleration pickup, an ultrasonic sensor, a laser Doppler accelerometer, or the like may be used as observation means.

[0023] The comb filter selector 21 receives an observed signal from the signal input device 1. The comb filter selector 21 then selects a comb filter suitable for noise removal per unit time in accordance with changes in the pitch of the noise over time. More specifically, the comb filter selector 21 selects a comb filter whose target pitch for removal matches the pitch of the noise per unit time. The unit time is significantly shorter than the change in the pitch of the noise. For example, the unit time is 1 millisecond or less. The unit time can be changed as appropriate depending on the degree of change in the pitch of the noise and the processing power of the processor or processing circuit operating the alarm removal device 2. For example, if the pitch of the noise changes gradually, extending the unit time reduces the number of processes (frequency) required to select a comb filter. This reduces the processing load of the processor or processing circuit operating the alarm removal device 2. The comb filter selector 21 references target pitch information stored in the comb filter storage unit 22, which indicates the target pitch for removal of each of the multiple comb filters. The information on the pitch to be removed is stored in, for example, the auxiliary storage device 903. The information on the pitch to be removed is loaded into the main storage device 902, and the comb filter selector 21 can refer to the information on the pitch to be removed. The comb filter selector 21 identifies the pitch of the noise contained in the observed signal for each unit time and selects a comb filter whose pitch to be removed indicated in the information on the pitch to be removed matches the pitch of the noise. The comb filter selector 21 notifies the comb filter storage unit 22 of the selection result. In other words, the comb filter selector 21 notifies the comb filter storage unit 22 of the selected comb filter (hereinafter referred to as the selected comb filter). The processing performed by the comb filter selector 21 corresponds to a comb filter selection process.

[0024] The comb filter storage unit 22 stores a plurality of comb filters, each of which has a different target pitch for elimination. The comb filter storage unit 22 outputs the selected comb filter notified by the comb filter selector 21 as the selection result to the comb filter application unit 23.

[0025] The comb filter application unit 23 applies the selected comb filter to the observed signal to remove noise from the observed signal. The comb filter application unit 23 then outputs the noise-removed signal, which is the observed signal from which noise has been removed, to a predetermined output destination. The processing performed by the comb filter application unit 23 corresponds to comb filter application processing.

[0026] In this way, the comb filter selector 21 selects a comb filter suitable for noise removal for each unit time in accordance with changes in the pitch of the noise over time. This configuration allows an appropriate comb filter to be applied for each unit time to noise whose pitch changes over time. This makes it possible to effectively remove non-stationary noise such as alarms and chimes. Furthermore, because this configuration selects an appropriate comb filter based solely on the observed signal, there is no need to provide information about the pitch of the noise from an external source. This avoids the need for complex equipment.

[0027] ***Explanation of Operation*** Fig. 3 shows an example of the operation of the alarm removal device 2 according to this embodiment. The flow shown in Fig. 3 is repeated every unit time.

[0028] In step S1 , the comb filter selector 21 receives an observation signal from the signal input device 1 .

[0029] Next, in step S2, the comb filter selector 21 selects a comb filter suitable for noise removal. The method of selecting a comb filter by the comb filter selector 21 will be described in detail later.

[0030] Next, in step S3, the comb filter selector 21 notifies the comb filter storage unit 22 of the selected comb filter.

[0031] Next, in step S4 , the comb filter storage unit 22 outputs the selected comb filter to the comb filter application unit 23 .

[0032] Next, in step S5, the comb filter application unit 23 applies the selected comb filter to the observed signal to remove noise from the observed signal.

[0033] Finally, in step S6, the comb filter application unit 23 outputs the noise-removed signal.

[0034] Next, a specific method for selecting a comb filter by comb filter selector 21 will be described.

[0035] An appropriate comb filter is selected when the pitch of the noise to be removed matches the pitch blocked by the selected comb filter. Therefore, the comb filter selector 21 can identify the pitch of the noise using an analysis technique such as autocorrelation or linear prediction, and then select a corresponding comb filter based on the identified pitch of the noise. However, in a situation where a desired signal is superimposed on the observed signal in addition to noise, the desired signal acts as a disturbance to these analysis techniques. For example, if the purpose is to inspect the operating noise of a machine, the desired signal would be the operating noise of the machine being inspected. In such a case, it is difficult to ensure robust operation of analysis techniques such as autocorrelation or linear prediction.

[0036] Therefore, one possible method for selecting comb filters in the comb filter selector 21 is to select comb filters that maximize the total amount of energy to be blocked. That is, the comb filter selector 21 calculates, for each unit time, the amount of energy removed by each comb filter when each comb filter removes signal components of a pitch that matches the target pitch from the observed signal. The comb filter selector 21 then selects the comb filter with the largest calculated amount of energy removed. The frequency components blocked by the selected comb filter match the fundamental frequency and its harmonics of the noise. As a result, noise is effectively removed. Furthermore, the comb filter passes signals across most frequency bands. Therefore, external disturbances have little effect on the calculation of the amount of energy to be blocked. This provides greater robustness and higher noise removal accuracy than the aforementioned analytical methods that estimate pitch using autocorrelation, linear prediction, etc.

[0037] Furthermore, when applying a comb filter in the frequency domain using a fast Fourier transform, the comb filter selector 21 prepares in advance an array a of index numbers of frequency bins corresponding to the frequencies to be cut off. Then, as shown in FIG. 4 , the comb filter selector 21 references only the frequency bins with the index numbers stored in array a and cuts off those frequency components. That is, each comb filter is assigned a frequency bin corresponding to the frequency component of the pitch to be removed. For each unit time, the comb filter selector 21 calculates the amount of energy in the frequency bins assigned to each comb filter as the amount of energy removed by each comb filter. The comb filter selector 21 then selects the comb filter with the largest calculated amount of energy removed. When applying a comb filter in the frequency domain using a fast Fourier transform, the observed signal is processed block by block. Therefore, the unit time is preferably set to a range of approximately 10 to 30 ms in conjunction with the block length. The unit time can be changed as appropriate depending on the analysis window length or block length of the fast Fourier transform.

[0038] 4 shows an example of the operation of comb filter selector 21 when applying a comb filter in the frequency domain by fast Fourier transform according to this embodiment. The flow shown in FIG. 4 is repeated every unit time.

[0039] In step S21, the comb filter selector 21 initializes a loop counter i to one.

[0040] Next, in step S22, the comb filter selector 21 determines whether the loop counter i is equal to or less than the length (size) of the array a. If the loop counter i is equal to or less than the length (size) of the array a (YES in step S22), the process proceeds to step S23. If the loop counter i exceeds the length (size) of the array a (NO in step S22), the process ends.

[0041] Next, in step S23, the comb filter selection unit 21 assigns the i-th index number of the array a to a variable w.

[0042] Next, in step S24, the comb filter selector 21 cuts off the component of the wth frequency bin.

[0043] Next, in step S25, comb filter selector 21 increments loop counter i by 1. Then, comb filter selector 21 returns to the determination process in step S22.

[0044] The comb filter application unit 23 applies a selective comb filter to the frequency components obtained by performing a fast Fourier transform on the observed signal for each unit time. The comb filter selection unit 21 blocks frequency components by replacing the value stored in the corresponding frequency bin with zero. Alternatively, the comb filter selection unit 21 may block frequency components by flooring (multiplying by a small real number such as 0.001). In this case, the total amount of energy blocked by the comb filter can be quantified as the sum of the amplitudes or powers of those frequency bins. This quantification can be performed in the same way as the application of a comb filter. Specifically, as shown in FIG. 5 , the comb filter selection unit 21 quantifies the frequency components by referencing only the frequency bins corresponding to each index number stored in array a and calculating the sum of the absolute values ​​(or squares of the absolute values) of those frequency components.

[0045] 5 shows an example of the operation of comb filter selector 21 when applying a comb filter in the frequency domain by fast Fourier transform according to this embodiment. The flow shown in FIG. 5 is repeated every unit time.

[0046] In step S31, the comb filter selector 21 initializes a loop counter i to 1 and also initializes a sum of energies s to 0.

[0047] Next, in step S32, comb filter selector 21 determines whether loop counter i is equal to or less than the length (size) of array a. If loop counter i is equal to or less than the length of array a (YES in step S32), the process proceeds to step S33. If loop counter i exceeds the length of array a (NO in step S32), the process proceeds to step S36.

[0048] Next, in step S33, comb filter selector 21 assigns the i-th index number of array a to variable w.

[0049] Next, in step S34, the comb filter selector 21 adds the energy of the wth frequency bin to the variable s.

[0050] Next, in step S35, comb filter selector 21 increments loop counter i by 1. Then, comb filter selector 21 returns to the determination process of step S32.

[0051] In step S36, comb filter selector 21 outputs the value of variable s as the sum of the energies, and the flow then ends.

[0052] A comb filter blocks only a small number of frequency bins. Therefore, by implementing the comb filter selection and application process as described above, highly accurate noise removal can be achieved with extremely low computational cost. Note that a discrete Fourier transform may be used for the comb filter instead of a fast Fourier transform.

[0053] The comb filter storage unit 22 stores multiple comb filters for selection by the comb filter selector 21. For example, the comb filter storage unit 22 stores comb filters that cut off all possible pitches of noise. This allows noise of various pitches to be removed. The following methods for generating such comb filters are possible. Method 1: Determine the fundamental frequencies of the possible pitches. Then, design a comb filter to cut off the fundamental frequency and all of its harmonic frequencies. Method 2: Record samples of possible noise individually for each pitch. Then, design a comb filter to cut off the peaks of each pitch in the power spectrum domain. Method 3: Record a series of samples of noise that form a melody. Decompose the spectrogram into bases for each pitch using nonnegative matrix factorization. Then, design a comb filter corresponding to each pitch based on these bases.

[0054] If the melody noise contains chords, a comb filter that removes all the notes that make up the chords is available as a selection, making it possible to handle chords as well.

[0055] As described above, the comb filter application unit 23 applies the selective comb filter to the observed signal from the signal input device 1. The comb filter application unit 23 then outputs a noise-removed signal. In addition to the frequency domain method described above, the comb filter application unit 23 can apply the comb filter to the observed signal by using the following time domain method. For example, the comb filter application unit 23 can configure an infinite impulse response (IIR) filter by connecting BiQuad notch filters in series to block the fundamental frequency and harmonic frequencies. In this case, the energy blocked by the comb filter can be quantified, for example, using the following method: The energy of the signal before and after applying the comb filter is compared. The amount of energy attenuation is then considered to be the amount of energy blocked by the comb filter.

[0056] Alternatively, the comb filter selector 21 may select multiple comb filters, and the comb filter applicator 23 may apply these comb filters simultaneously. For example, consider a case where the noise melody contains a chord with at most three pitches. In this case, the comb filter storage unit 22 stores comb filters corresponding to all single notes included in the melody. The comb filter selector 21 then selects the three most effective comb filters. That is, the comb filter selector 21 selects three comb filters per unit time whose target pitches correspond to the three pitches that make up the noise. The comb filter applicator 23 then applies the three selected comb filters to the observed signal per unit time to remove noise from the observed signal. This configuration eliminates the need to prepare a comb filter for each chord, thereby reducing implementation costs. Furthermore, when the noise contains reverberation, the noise can be removed, including the reverberation. Specifically, when the reverberation of a previously played pitch is superimposed on the next pitch, creating a chord-like effect, the noise can be removed, including the reverberation. Therefore, a higher noise reduction effect can be obtained.

[0057] In the above configuration, it is assumed that the comb filter selector 21 always selects one of the comb filters. Therefore, one of the comb filters will be applied even if the observed signal does not contain noise. Because a comb filter only blocks a small number of frequency components, such a configuration is usually considered to have little effect on the output signal.

[0058] However, if it is desired to reduce distortion in the output signal, it is also possible to introduce a state in which the comb filter selector 21 does not select any comb filter. When the comb filter selector 21 does not select a comb filter, the comb filter applicator 23 outputs the observed signal from the signal input device 1 as is. The following conditions may be considered for the comb filter selector 21 not to select a comb filter. For example, when a comb filter that maximizes the total amount of energy rejected is known, the comb filter selector 21 determines that the observed signal does not contain noise if the total amount of energy rejected by that comb filter is equal to or less than a predetermined threshold. In other words, if the maximum total amount of energy (maximum amount of removed energy) in any unit time is equal to or less than a predetermined threshold, the comb filter selector 21 determines that the observed signal does not contain noise. In this case, the comb filter selector 21 does not select any comb filter. With this configuration, distortion does not occur in the signal output from the comb filter applicator 23 when noise is not mixed into the observed signal, and good sound quality can be obtained.

[0059] In the above configuration, no particular constraints are imposed on the operation of switching the applied comb filter over time. Therefore, if a large disturbance occurs in the observed signal, an appropriate comb filter may not be selected for each unit time. In this case, the comb filter may be switched more frequently than necessary, causing distortion in the output signal. To avoid this, a configuration may be considered that prevents the comb filter from switching more than necessary. Specifically, a configuration may be considered in which the comb filter selector 21 is penalized for switching the comb filter. For example, a score function may be introduced that includes the sum of the energy blocked by the selected comb filter for each unit time as a reward term and the number of times the selected comb filter switches as a penalty term. The comb filter selector 21 then selects a comb filter that maximizes this score function for each unit time. The comb filter selector 21 can efficiently determine a comb filter switching method that maximizes this score function, for example, using dynamic programming. This stabilizes the comb filter selection result even when a large disturbance occurs. As a result, better sound quality can be obtained.

[0060] ***Description of Effects of the Embodiment*** As described above, according to this embodiment, it is possible to remove noise whose pitch changes over time with a simple configuration. Furthermore, according to this embodiment, by using the array a of frequency bin index numbers, it is possible to achieve highly accurate noise removal with extremely low calculation cost. Furthermore, according to this embodiment, even when the noise contains chords, it is possible to effectively remove chords from the noise. Furthermore, according to this embodiment, even when the noise contains reverberation, it is possible to remove the noise including the reverberation. Furthermore, according to this embodiment, a comb filter is not selected when the observed signal does not contain noise. As a result, no distortion occurs in the output signal, and good sound quality can be obtained. Furthermore, according to this embodiment, the comb filter selection result is stable by using a score function even when a large disturbance occurs. As a result, better sound quality can be obtained.

[0061] The procedure described in this embodiment is an example. Therefore, only a part of the procedure described in this embodiment may be implemented. Also, at least a part of the procedure described in this embodiment may be implemented in combination with a procedure not described in this embodiment. Also, the configuration and procedure described in this embodiment may be changed as necessary.

[0062] *** Supplementary Explanation of Hardware Configuration *** Finally, a supplementary explanation of the hardware configuration of the alarm removal device 2 will be provided. The processor 901 shown in Fig. 2 is an IC (Integrated Circuit) that performs processing. The processor 901 is a CPU (Central Processing Unit), a DSP (Digital Signal Processor), or the like. The main storage device 902 shown in Fig. 2 is a RAM (Random Access Memory). The auxiliary storage device 903 shown in Fig. 2 is a ROM (Read Only Memory), a flash memory, an HDD (Hard Disk Drive), or the like.

[0063] The auxiliary storage device 903 also stores an operating system (OS). At least a portion of the OS is executed by the processor 901. While executing at least a portion of the OS, the processor 901 executes a program that implements the functions of the comb filter selector 21 and the comb filter application unit 23. The processor 901 executes the OS to perform tasks, memory management, file management, communication control, and the like. At least one of information, data, signal values, and variable values ​​indicating the results of processing by the comb filter selector 21 and the comb filter application unit 23 is stored in at least one of the main storage device 902, the auxiliary storage device 903, and a register and cache memory within the processor 901. The program that implements the functions of the comb filter selector 21 and the comb filter application unit 23 may be stored on a portable recording medium such as a magnetic disk, a flexible disk, an optical disk, a compact disk, a Blu-ray (registered trademark) disk, or a DVD. Portable recording media storing the program that implements the functions of the comb filter selector 21 and the comb filter application unit 23 may be distributed.

[0064] Furthermore, the "part" in at least one of the comb filter selector 21 and the comb filter applicator 23 may be interpreted as a "circuit," a "step," a "procedure," a "process," or a "circuitry." The alarm elimination device 2 may be realized by a processing circuit. The processing circuit may be, for example, a logic integrated circuit (IC), a gate array (GA), an application-specific integrated circuit (ASIC), or a field-programmable gate array (FPGA). In this case, the comb filter selector 21 and the comb filter applicator 23 are each realized as part of the processing circuit. In this specification, the term "processing circuitry" refers to a generic concept that encompasses both a processor and a processing circuit. In other words, the processor and the processing circuit are each specific examples of "processing circuitry."

[0065] 1 signal input device, 2 alarm elimination device, 21 comb filter selection unit, 22 comb filter storage unit, 23 comb filter application unit, 901 processor, 902 main memory device, 903 auxiliary memory device, 904 input / output device.

Claims

1. A signal processing device having: a comb filter selection unit that, when an audio signal containing noise whose pitch changes over time is received, selects, as a selected comb filter, a comb filter whose target pitch for removal matches the pitch of the noise from a plurality of comb filters each having a different target pitch for removal in accordance with the change in pitch of the noise over time; and a comb filter application unit that applies the selected comb filter to the audio signal to remove the noise from the audio signal.

2. The signal processing device described in claim 1, wherein the comb filter selection unit calculates, for each unit time, the amount of energy removed from the audio signal when each comb filter removes signal components of a pitch that matches the pitch to be removed from the audio signal as the amount of energy removed by each comb filter, and selects the selected comb filter based on the calculated amount of energy removed by each comb filter.

3. The signal processing device according to claim 2, wherein the comb filter selection unit selects, as the selected comb filter, a comb filter for which the maximum amount of removed energy is calculated.

4. A signal processing device as described in claim 2, wherein each comb filter is assigned a frequency bin corresponding to the frequency component of the pitch to be removed, the comb filter selection unit calculates the amount of energy of the frequency bin assigned to each comb filter for each unit time as the amount of energy removed by each comb filter, and the comb filter application unit applies the selected comb filter to the frequency component obtained by performing a Fourier transform on the audio signal for each unit time.

5. The signal processing device of claim 1, wherein the noise is composed of two or more pitches, the comb filter selection unit selects, for each unit time, two or more comb filters as the selected comb filters, each of which has a pitch to be removed corresponding to one of the two or more pitches that make up the noise, and the comb filter application unit applies, for each unit time, the two or more selected comb filters to the audio signal to remove the noise from the audio signal.

6. The signal processing device described in claim 2, wherein the comb filter selection unit does not select any comb filter as the selected comb filter when the maximum amount of removed energy in any unit time is equal to or less than a predetermined threshold, and the comb filter application unit outputs the audio signal to a predetermined output destination without processing the audio signal when the selected comb filter is not selected.

7. The signal processing device according to claim 1, wherein the comb filter selection unit selects, as the selected comb filter, a comb filter that maximizes a score function that includes the amount of energy removed by each comb filter as a reward term and the number of times the comb filter selected as the selected comb filter is switched as a penalty term.

8. A signal processing method in which, when an audio signal containing noise whose pitch changes over time is received, a computer selects, from a plurality of comb filters each having a different target pitch for removal in accordance with the change in pitch of the noise over time, a comb filter whose target pitch for removal matches the pitch of the noise as a selected comb filter, and the computer applies the selected comb filter to the audio signal to remove the noise from the audio signal.

9. A signal processing program that causes a computer to execute, when an audio signal containing noise whose pitch changes over time is received, a comb filter selection process that selects, from a plurality of comb filters each having a different target pitch for removal in accordance with the change in pitch of the noise over time, a comb filter whose target pitch for removal matches the pitch of the noise as a selected comb filter, and a comb filter application process that applies the selected comb filter to the audio signal to remove the noise from the audio signal.