JP6229576B2 - Sampling frequency estimation device - Google Patents

Description

  The present invention relates to a technique for synchronizing a plurality of signals obtained by separately sampling the same waveform.

  In recent years, recording devices that can easily perform digital recording, such as IC recorders, and devices that can simultaneously record digital recording, such as smartphones, have become popular. Here, digital recording refers to recording a sound signal in the form of a sample sequence obtained by sampling a sound waveform. For example, while recording a live performance video and performance sound away from the player using a smartphone, the performance sound is recorded with an IC recorder placed near the player, and the performance sound recorded by the smartphone is recorded with the IC recorder. In some cases, the recorded performance sound is replaced (or the former is overlaid with the latter). In general, even if the sampling frequencies of all the recording devices are set to be the same in the device settings, fine variations occur in the sampling frequencies of the recording devices. This is due to the fact that the clock generator that determines the sampling frequency does not operate at the exact same clock frequency. Therefore, when the same sound waveform is separately digitally recorded by multiple recording devices, even if the recording start timing is aligned, the sampling frequency varies from recording device to recording device. End up. As a technique for correcting the deviation of the sampling frequency, there are techniques disclosed in the prior art documents of Patent Document 1, Non-Patent Document 1, and Non-Patent Document 2.

  Non-Patent Document 1 discloses a technique for transmitting a reference signal (pilot signal) from a transmitter and detecting and correcting a frequency shift due to a sampling frequency shift from a reference signal included in a signal received on the receiver side. Yes. Patent Document 1 discloses a correction technique in the case where the sampling frequency is different between the side that transmits a measurement signal (TSP signal or the like) and the side that receives it when measuring the transfer characteristic of the sound field. In the technique disclosed in Patent Document 1, a TSP signal is repeatedly transmitted in order to suppress the influence of noise at the time of measurement, and a plurality of measured TSP signals are cut out at regular intervals. By detecting the phase difference, the sampling frequency deviation is estimated and corrected.

  Non-Patent Document 2 discloses a technique for correcting a sampling frequency shift between a plurality of recording devices using statistical signal processing. In the technique disclosed in Non-Patent Document 2, first, a reference signal is determined for each recording signal recorded by a plurality of recording devices. Then, a signal when the sampling frequency is shifted from the reference signal is statistically modeled, and a signal other than the reference signal is applied to the statistical model to estimate the sampling frequency shift.

JP 2002-101500 A

Matsuoka, Yasushi, Nakajima, Yusuke, Yoshimura, Ken, “Acoustic Information Transmission Method for Acquiring Information Using Mobile Terminal Microphone”, NTT DocomoTechnical Journal, vol.14, No.2, 2006 Shigeki Miyabe, Nobutaka Ono, and Shoji Makino, “BLIND COMPENSATION OF INTER-CHANNEL SAMPLING FREQUENCYMISMATMAT WITH MAXIMUM LIKELIHOOD ESTIMATION IN STFT DOMAIN,” proc. ICASSP 2013, pp.674-678

  However, the technique disclosed in Non-Patent Document 1 and the technique disclosed in Patent Document 1 have a problem that there are many restrictions and lacks versatility. For example, the technique disclosed in Non-Patent Document 1 requires a device that generates a reference signal (pilot signal), and has a problem that the recording signal is affected by the reference signal. On the other hand, the technique disclosed in Patent Document 1 has a problem that it cannot be used unless the same signal is repeatedly transmitted at a constant interval. On the other hand, although Non-Patent Document 2 does not have a problem of lack of versatility, its execution requires a large amount of calculation and requires a long calculation time to complete estimation of the sampling frequency deviation. There is a problem I took.

  The present invention has been made in view of the problems described above, and it is possible to realize synchronization of a plurality of signals obtained by independently sampling the same waveform in a shorter calculation time than in the past, and It aims at providing the technology which has high versatility.

  In order to solve the above problem, the present invention uses one of a plurality of signals obtained by sampling the same waveform separately and independently as a reference signal, and one of the remaining signals as a correction target signal. A time shift amount calculation unit that calculates the correlation between both signals for each frame while shifting one of the reference signal and the correction target signal in the time axis direction, and calculates the time shift amount of both signals for each frame according to the calculation result An error calculation unit that calculates, for each frame, a first estimated value that is an estimated value of the sampling frequency error of the correction target signal in each frame from the time shift amount calculated by the time shift amount calculation unit; Statistical processing is performed on the first estimated value calculated for each frame by the error calculating unit to calculate a second estimated value that is an estimated value of the sampling frequency error over the entire correction target signal. Providing sampling frequency estimating apparatus, characterized by having a statistical processing unit for outputting. Since the error in the sampling frequency of the correction target signal is a deviation of the sampling frequency of the correction target signal from the sampling frequency of the reference signal, the sampling frequency of the correction target signal can be obtained from the error and the sampling frequency of the reference signal. it can. Therefore, calculating the estimated value of the error (that is, estimating the error) is equivalent to estimating the sampling frequency of the correction target signal.

  According to such a sampling frequency estimation device, one of a plurality of signals obtained by independently sampling the same waveform is used as a reference signal, and each of the remaining signals is used as a correction target signal. By calculating an estimated value of the sampling frequency error and correcting the error using an existing technique such as time axis companding, it becomes possible to synchronize each correction target signal with the reference signal. The sampling frequency estimation apparatus of the present invention does not require a pilot signal, and each signal does not need to be repeatedly output over a certain period of time. Therefore, the technology disclosed in Non-Patent Document 1 and Patent Document 1 Higher versatility than Although details will be described later, according to the sampling frequency estimation apparatus of the present invention, the error of the sampling frequency of the correction target signal can be calculated in a shorter calculation time than when the technique disclosed in Non-Patent Document 2 is used. The synchronization of a plurality of signals obtained by sampling the same waveform separately and independently can be realized in a shorter calculation time than in the prior art.

  As a specific configuration of the statistical processing unit, it is estimated that there are statistically many errors from the first estimated value calculated for each frame by the error calculating unit (that is, an estimated value of a rough error in each frame). First statistical filter processing for excluding outliers to be performed, and filter processing for smoothing a group of first estimated values from which outliers have been excluded from the first statistical filter processing (for example, processing for calculating an average value) ) And a second statistical filter process consisting of any one of a filter process for selecting a representative value from the group of first estimated values (for example, a process for selecting a median value), A configuration is conceivable in which the processing result of the second statistical filter processing is output as a second estimated value (estimated value of sampling frequency error over the entire correction target signal).

  As a specific example of the first statistical filter processing, each of the first estimated value calculated for each frame by the error calculation unit when the first estimated value is sorted in order of the size, or a predetermined number from the both ends. There is a process of removing those that do not belong to the range determined according to the value as outliers. For example, if the predetermined number is 1/4 of the total number of the first estimated values calculated by the error calculating unit, a value less than the first quartile and a value greater than the third quartile are outliers. Will be excluded. In addition, outliers are excluded by the so-called quartile range method if weights are assigned to the first and third quartiles to define the above range.

  In a more preferred aspect, the time deviation amount calculation unit is characterized in that a frame in which the power of the reference signal and the correction target signal is less than a predetermined threshold is excluded from the time deviation amount calculation target. If the threshold is set to an appropriate value, a frame that does not include the reference signal with sufficient intensity and a frame that does not include the correction target signal with sufficient intensity are excluded from the calculation target of the amount of time deviation. Even if the amount of time shift is calculated with reference to a frame that does not include the reference signal with sufficient intensity or a frame that does not include the correction target signal with sufficient intensity, it includes many errors. The estimated error value calculated based on such a time shift amount is likely to be excluded by the first statistical filter process as an outlier, and the calculation of the time shift amount itself becomes useless in the first place. According to such an aspect, it is possible to avoid a wasteful calculation in the time shift amount calculation unit and further reduce the processing time required for estimating the sampling frequency of the correction target signal, while reducing the sampling frequency error with high accuracy. It becomes possible to calculate by.

  In another preferred embodiment, the time deviation amount calculation unit has a value representing a correlation between the reference signal and the correction target signal (for example, a maximum value among a plurality of cross-correlation functions calculated while shifting the time). Frames that are less than the threshold value are excluded from the calculation target of the time shift amount. If the threshold value is set to an appropriate value, the time shift amount is not calculated for a frame having a low correlation with the frame corresponding to the reference signal among the frames constituting the correction target signal. Even if the time shift amount is calculated for such a frame, it contains a lot of errors, and the error estimated value calculated based on such a time shift amount is regarded as an outlier as the first statistical filter processing. In the first place, the calculation of the amount of time deviation itself becomes useless. Even in such an aspect, it is possible to calculate the sampling frequency with high accuracy while avoiding unnecessary calculation in the time deviation amount calculation unit and further shortening the processing time required to estimate the sampling frequency of the correction target signal. Is possible.

  As another aspect for solving the above problem, there is an aspect that provides a program that causes a general computer such as a CPU (Central Processing Unit) to function as the time shift amount calculation unit, the error calculation unit, and the statistical processing unit. Conceivable. This is because by operating a general computer according to such a program, it becomes possible to cause the computer to function as the sampling frequency estimation device of the present invention. In addition, as a specific provision mode of such a program, a mode in which the program is written and distributed on a computer-readable recording medium such as a CD-ROM (Compact Disk-Read Only Memory), or an electric communication such as the Internet A mode of distribution by downloading via a line is conceivable.

  As another mode for solving the above problem, one of a plurality of signals obtained by independently sampling the same waveform is used as a reference signal, and one of the remaining signals is corrected. As a target signal, the cross-correlation function of both signals is calculated for each frame while shifting one of the reference signal and the correction target signal in the time axis direction, and the time shift amount of both signals is calculated for each frame according to the calculation result. A time shift amount calculation step to be calculated, and an error calculation step of calculating, for each frame, an estimated value of the sampling frequency error of the correction target signal in each frame from the time shift amount calculated in the time shift amount calculation step. The error calculation step performs statistical processing on the error estimation value calculated for each frame to calculate the sampling frequency error estimation value of the correction target signal. Aspect to provide a sampling frequency estimation method characterized by comprising: a statistical processing step of force, also conceivable. In addition, a mode is also conceivable in which a program for causing a general computer such as a CPU to execute the steps of the time deviation calculation step, the error calculation step, and the statistical processing step is conceivable.

It is a block diagram which shows the structural example of the sampling frequency estimation apparatus 10 of one Embodiment of this invention, and the structural example of the signal processing system 1 containing the sampling frequency estimation apparatus 10. FIG. It is a figure which shows the estimation result of the sampling frequency for every flame | frame obtained by the sampling frequency estimation apparatus. It is a figure for demonstrating a quartile and a quartile range. It is a figure for demonstrating the effect of the outlier removal process based on a quartile. It is a figure which shows the experimental result of the evaluation experiment about the technique disclosed by this embodiment and a nonpatent literature 2. FIG. It is a figure which shows the experimental result of the evaluation experiment about the technique disclosed by this embodiment and a nonpatent literature 2. FIG.

Embodiments of the present invention will be described below with reference to the drawings.
(A: Configuration)
FIG. 1 is a block diagram illustrating a configuration example of a sampling frequency estimation apparatus 10 and a configuration example of a signal processing system 1 including the sampling frequency estimation apparatus 10 according to an embodiment of the present invention. In this signal processing system 1, each sound signal (sample) obtained by independently sampling the same sound waveform by each of N (N is a natural number of 2 or more) recording devices (for example, a smartphone or an IC recorder). Column) X _n (t) (n = 1 to N) is input. In addition, what is necessary is just to use an existing technique suitably about the synchronization of the recording start timing in N recording apparatuses. For example, if each recording device can execute communication via an electric communication line such as the Internet, the recording start timing may be adjusted by the communication, and communication by other communication means such as Bluetooth (registered trademark) is possible. For example, the recording start timing may be adjusted by communication by the communication means.

  Although the sampling frequencies of the N recording devices are all set to the same value (for example, 44.1 kHz), the clock generators of the recording devices do not operate at the same clock frequency. There is a slight shift in the sampling frequency of recording equipment. For this reason, even when recording is performed with the recording start timings being adjusted at each recording device and the N sound signals are reproduced with their heads aligned, the sound gradually shifts, and the shift increases as playback proceeds. The signal processing system 1 of the present embodiment is for estimating and correcting an error in sampling frequency between the N sound signals so that these N sound signals can be synchronized.

  As shown in FIG. 1, the signal processing system 1 includes a sampling frequency estimation device 10 and a time axis companding device 20. The N sound signals are given to the sampling frequency estimation apparatus 10. The sampling frequency estimation device 10 uses one of these N sound signals as a reference signal and each of the remaining N−1 sound signals as a correction target signal, thereby correcting each correction target signal with respect to the sampling frequency of the reference signal. The sampling frequency deviation (ie, error) is estimated for each correction target signal, and data indicating the estimation result is given to the time axis companding device 20. The time axis companding device 20 performs time axis companding on each correction target signal so that the error of the sampling frequency estimated for each correction target signal is eliminated. Thereby, synchronization of N sound signals is realized. As the time axis companding algorithm in the time axis companding device 20, an existing technique may be appropriately used. In the present embodiment, by causing the sampling frequency estimation apparatus 10 to execute processing that clearly shows the features of the present embodiment, the estimation of the sampling frequency error of each correction target signal is realized in a shorter calculation time than in the past. It is possible to ensure high versatility. Below, it demonstrates centering on the sampling frequency estimation apparatus 10 which shows the characteristic of this embodiment notably.

  As shown in FIG. 1, the sampling frequency estimation device 10 includes a short-time Fourier transform (indicated as “STFT” in FIG. 1) unit 100, a time shift amount calculation unit 110, an error calculation unit 120, and a statistical processing unit 130. Contains. Each unit illustrated in FIG. 1 may be a hardware module configured by an electronic circuit, or may be a software module realized by operating a CPU (Central Processing Unit) according to a signal processing program.

The STFT unit 100 divides each of the sound signals X _n (t) (n = 1 to N) input to the sampling frequency estimation device 10 into frames of a predetermined number of samples, and performs a short-time Fourier transform for each frame. The signal is converted into a frequency domain signal X _n (f) (f is a variable representing frequency, hereinafter the same), and is supplied to the time shift amount calculation unit 110. Any known conversion algorithm used in the STFT unit 100 may be used as appropriate.

The time shift amount calculation unit 110 selects one of the N sound signals as a reference signal, sequentially selects each of the remaining N−1 sound signals as a correction target signal, and selects the reference signal and the correction target. While shifting one of the signals in the time axis direction, the cross-correlation function of both signals is calculated for each frame, and the amount of time shift between both signals is calculated for each frame. In the following, details of the process performed by the time shift amount calculation unit 110 will be described by taking as an example the case where X _ref (f) is selected as the reference signal and X _k (f) (k ≠ ref) is selected as the correction target signal. Explained.

First, the time shift amount calculation unit 110 changes the value of τ in the cross-correlation function C (τ) for each frame of the reference signal X _ref (f) and the correction target signal X _k (f) (k ≠ ref). While calculating. The reason why the cross-correlation function C (τ) is calculated is that a method for estimating the sample point at which the cross-correlation function is maximized is generally known as a method for correcting the amount of time deviation. In general, the cross-correlation function C (τ) is expressed by the following formula 1 when there are two analog signals x (t) and y (t) in the time domain. In the case of a digital signal, it is expressed by the following formula 2.

The cross-correlation function C (τ) calculated by Equation 1 or Equation 2 corresponds to taking an inner product by shifting either of the two signals by τ in the time axis direction. If the assumption that “the same signal is included somewhere in the two signals” holds, it is considered that the amount of time deviation between the two signals can be estimated by obtaining τ that maximizes the cross-correlation function. . Considering the Fourier transforms X (f) and Y (f) of the two signals, the cross-correlation function C (τ) is calculated by the following equation (3). IFFT () on the right side of Equation 3 is an operator representing inverse Fourier transform, and X ^* (f) represents a complex conjugate of X (f).

The time shift amount calculation unit 110 according to the present embodiment changes the value of τ in the cross-correlation function C (τ) of each frame of the reference signal X _ref (f) and the correction target signal X _k (f) (k ≠ ref). However, the calculation is performed according to Equation 3. Specifically, the time shift amount calculation unit 110 sets the complex conjugate of the signal X _ref (f) for the i-th frame in the reference signal as X ^* (f) on the right side of Equation 3, and sets the i-th in the correction target signal. The signal X _k (f) (k ≠ ref) for the frame is changed to Y (f) on the right side of the equation 3 and the calculation of the equation 3 is performed while τ is changed to specify τ that maximizes the cross-correlation function C (τ). . Then, the time shift amount calculation unit 110 uses the τ specified in this way to estimate the time shift amount N _ki (that is, the cross-correlation function C () for the i-th frame of the correction target signal X _k (f). The error calculation unit 120 is given as τ) that maximizes τ). The same applies to frames with other numbers.

  In the present embodiment, it is advantageous in terms of calculation amount to calculate the cross-correlation function C (τ) by the frequency domain calculation shown in Formula 3 instead of the time domain calculation of Formula 1 or Formula 2. Because there is. In the present embodiment, the STFT unit 100 is provided so that the cross-correlation function C (τ) can be calculated by the calculation shown in Equation 3 in the time deviation calculation unit 110. Therefore, the STFT unit 100 may be omitted if the cross-correlation function C (τ) is calculated by the calculation shown in Equation 1 or 2 in the time deviation calculation unit 110.

Based on the time shift amount N _ki of each frame given from the time shift amount calculation unit 110, the error calculation unit 120 estimates the error (hereinafter referred to as the first value) of the sampling frequency error of the correction target signal with respect to the sampling frequency f _s of the reference signal. (Estimated value of 1) E _ki is calculated for each frame. For example, when the time shift amount for the i-th frame is N _ki and the first sample of the i-th frame in the correction target signal X _k (f) is the S _ki- th sample from the head of the signal. The error calculation unit 120 calculates the first estimated value E _ki for the i-th frame according to the following _Equation 4, and provides the statistical processing unit 130 with it. As described above, the time shift amount N _ki calculated by the time shift amount calculation unit 110 is a shift amount with reference to the head of the reference signal, so that the cross-correlation function of each frame is calculated by the STFT as in this embodiment. When it is obtained, the amount of deviation is based on the frame head. Therefore, the estimated value E _ki of the sampling frequency error of the correction target signal X _k (f) based on the cross-correlation function in the i-th frame is expressed by the following equation (4).

The statistical processing unit 130 performs statistical processing on the first estimated value E _ki calculated for each frame by the error calculating unit 120 to estimate the sampling frequency error over the entire correction target signal (hereinafter, the second estimated value). Value) E is calculated and output to the time axis companding device 20. As shown in FIG. 1, the statistical processing unit 130 includes a first statistical filter processing unit 130a and a second statistical filter processing unit 130b. That is, the statistical processing executed by the statistical processing unit 130 includes processing executed by the first statistical filter processing unit 130a and processing executed by the second statistical filter processing unit 130b. The contents of the processing executed by each statistical filter processing unit are as follows.

The first statistical filter processing unit 130a excludes outliers that are statistically estimated to contain many errors from the first estimated value _Eki calculated for each frame by the error calculating unit 120. Perform filtering. The first estimated value E _ki calculated for each frame by the error calculation unit 120 often includes many errors. The first estimated value E _ki roughly approximates the deviation of the sampling frequency of both signals based only on the i-th frame information of the correction target signal X _k (f) and the reference signal X _ref (f). This is because it is an estimated value. FIG. 2 shows a sampling frequency estimation result for each frame when an experiment is performed by artificially shifting the sampling frequency by 3 Hz. If the error of the sampling frequency of the correction target signal with respect to the sampling frequency of the reference signal is known, the sampling frequency of the correction target signal can be calculated from the sampling frequency of the reference signal and the error. The sampling frequency estimates of are equivalent. As shown in FIG. 2, the sampling frequency estimated for each frame has a large variation because it is difficult to estimate a very small time shift due to the sampling frequency shift with high accuracy.

In the present embodiment, the first statistical filter process is a process of removing, as an outlier, a value that deviates significantly from the first estimated value E _ki calculated by the error calculation unit 120 as compared to other values. It has been adopted. Specifically, in the present embodiment, a process based on a so-called quartile is employed as the first statistical filter process. Here, the quartile means the number of divisions into which the data to be processed is sorted in order of size and then is divided into four equal parts. The first quartile and the second fourth It is called the quantile and the third quartile (see FIG. 3). Further, the difference between the first quartile and the second quartile is called an interquartile range (IQR). The quartile range is an index that represents the degree of variation of the sample.

More specifically, the first statistical filter processing unit 130a first sorts the first estimated values E _ki calculated for each frame by the error calculation unit 120 in the order of the size. Next, the first statistical filter processing unit 130a compares the first estimated value E _ki calculated for each frame by the error calculating unit 120 with a value smaller than the first quartile in the sorting result, or the third A value larger than the quartile is excluded as an outlier, and the remainder (that is, a group of first estimated values E ′ _ki not including an outlier) is delivered to the second statistical filter processing unit 130b. Here, the calculation o () for detecting the outlier is expressed by the following equation 5. Specifically, each of the first estimated values E _ki calculated by the error calculating unit 120 is substituted for e (n) in Equation 5, and if the value of the operation o () is 1, the first estimated value is calculated. For example, the value E _ki is excluded as an outlier. q _L and q _H represent the first and third quartiles, respectively.

In the present embodiment, processing based on the quartile is employed as the first statistical filter processing, but processing using a quartile range may be used in addition to the quartile. Specifically, the operation o () shown in Equation 6 may be performed instead of the operation o () shown in Equation 5 as an operation for identifying whether it is an outlier. The calculation shown in Equation 6 means that the IQR value is added to or subtracted from the first and third quartiles with weights. When α = 0, Equation 6 is identical to Equation 5. A method of calculating with α = 1.5 is widely known. For example, portions corresponding to the upper and lower whiskers of the box whisker chart shown in FIG. 4 are calculated with this.

The second statistical filter processing unit 130b performs second statistical filter processing (specifically, selecting a representative value from the group of first estimated values E ′ _ki from which outliers have been excluded from the first statistical filter processing unit 130a. the performs filtering) for selecting the median value, gives the time scale modification apparatus 20 the processing result as a second estimate E _k. Although the maximum value and the minimum value can be used as the representative value, the median value is considered most preferable. Further, as the second statistical filter processing executed by the second statistical filter processing unit 130b, the group of first estimated values E ′ _ki from which outliers are excluded from the first statistical filter processing unit 130a are smoothed. A filter process (a process of calculating an average value of a group of first estimated values E ′ _ki from which outliers have been excluded from the first statistical filter processing unit 130a) may be employed. According to experiments, a better result was obtained when the filter processing for selecting the median was adopted. For this reason, in this embodiment, the filter process which selects a median is employ | adopted.
The above is the configuration of the sampling frequency estimation apparatus 10.

(B: Effect of the embodiment)
According to the present embodiment, one of the N sound signals is a reference signal, each of the remaining N-1 sound signals is a correction target signal, and the sampling frequency of the correction target signal with respect to the sampling frequency of the reference signal Is estimated by the sampling frequency estimation device 10 for each correction target signal, and the time axis companding is applied to the correction target signal so as to eliminate the error, thereby synchronizing N sound signals. In order to evaluate the effect of the present embodiment, the present applicant uses the technique disclosed in Non-Patent Document 2 as a comparison target, and estimates the sampling frequency error estimation performance and calculation speed (calculation of the sampling frequency error estimation value). An evaluation experiment was conducted from the viewpoint of the time length of calculation time required to complete. The outline of this evaluation experiment is as follows.

  First, a sound signal of 10 commercial songs of 16 bits sampled at a sampling frequency of 44.1 kHz (genre is pop, each song has a duration of 10 seconds) is used as an original signal, and the original signal is used as a reference signal. A signal obtained by artificially performing resampling (± 5 Hz) on the original signal is used as a correction target signal, and an error in the sampling frequency of each correction target signal is disclosed in the sampling frequency estimation device 10 of this embodiment and Non-Patent Document 2. Estimated by technology. In this evaluation experiment, a computer having a 3.4 GHz drive Corei7 3770 as a CPU and a 32 GB RAM is used as the sampling frequency estimation apparatus 10, and MATLAB (registered trademark) is used for mounting each part such as the STFT part 100. Was used. MATLAB (registered trademark) is numerical analysis software of The MathWorks, USA. Similarly, the method disclosed in Non-Patent Document 2 was also implemented on the same computer using C / C ++ and MATLAB (registered trademark). Further, the FFT length is 4096 samples, a Hamming window having a window length of 4096 samples is used as an analysis window, and further, 8192 as a shift size (that is, an update unit of τ) when calculating the cross-correlation function C (τ). , 4096, 2048 and 1024 samples were used, and the data range used was (3/8) × T to (5/8) × T (T is the number of data).

  FIG. 5A is a diagram illustrating an experimental result regarding the estimation performance for the present embodiment, and FIG. 5B is a diagram illustrating an experimental result regarding the estimation performance for the method disclosed in Non-Patent Document 2. As is clear from a comparison between FIG. 5A and FIG. 5B, the technology disclosed in Non-Patent Document 2 exceeds the maximum performance (that is, the estimation error is small). However, for example, recording is performed for 2 hours (7200 seconds), and a practical range in which the time deviation of the corrected correction target signal with respect to the reference signal is suppressed to 5 milliseconds or less (the sampling frequency estimation error is suppressed to within 0.03 Hz). This performance is also achieved in this embodiment. Therefore, no problem in the practical range occurs in this embodiment. Further, from FIG. 5A, it can be seen that the same estimation performance can be realized in this embodiment regardless of the shift size. The shift size affects the calculation amount. That is, the experimental result of FIG. 5A means that according to the present embodiment, a practical range of performance can be sufficiently achieved even if the amount of calculation is reduced.

  FIG. 6A is a diagram illustrating an experimental result regarding the calculation speed for the present embodiment, and FIG. 6B is a diagram illustrating an experimental result regarding the calculation speed for the method disclosed in Non-Patent Document 2. The As is clear from the comparison between FIG. 6A and FIG. 6B, the method of this embodiment is overwhelmingly faster than the method disclosed in Non-Patent Document 2 (estimation of sampling frequency deviation is completed). It is understood that the calculation time required until the calculation is short), and even with the implementation by MATLAB (registered trademark), the calculation speed exceeding the implementation of the method disclosed in Non-Patent Document 2 by C / C ++ is obtained. When the above experimental results are summarized, it can be concluded that according to the present embodiment, a practical range of estimation performance can be achieved in a shorter calculation time than the technique disclosed in Non-Patent Document 2.

  As described above, according to the present embodiment, synchronization of a plurality of sound signals obtained by separately sampling the same sound waveform is realized in a shorter calculation time than the technique disclosed in Non-Patent Document 2. It becomes possible. In addition, in the present embodiment, synchronization is possible only with a sampled sound signal (in other words, a recorded sound signal), and a pilot signal is not required, so that it is compared with the technique disclosed in Non-Patent Document 1. And high versatility. In the present embodiment, each sound signal to be synchronized does not need to be repeatedly transmitted, and has higher versatility than the technique disclosed in Patent Document 1. That is, according to the present embodiment, synchronization of a plurality of signals obtained by independently sampling the same waveform can be realized in a shorter calculation time than in the past, and high versatility can be realized. It becomes possible.

(C: deformation)
Although one embodiment of the present invention has been described above, it goes without saying that the following modifications may be added to this embodiment.
(1) In the above embodiment, a case has been described in which the plurality of signals input to the sampling frequency estimation apparatus 10 are a plurality of sound signals obtained by sampling the same sound waveform separately and independently. However, the plurality of signals input to the sampling frequency estimation apparatus 10 are not limited to sound signals as long as they are obtained by sampling the same waveform separately and independently. Moreover, in the said embodiment, although the process using a quartile was employ | adopted as a 1st statistical filter process, for example, after sorting the estimated value calculated for every flame | frame by the error calculation part 120 in order of the magnitude | size A process may be adopted in which they are divided into three equal parts, and outliers are used that are smaller than the value corresponding to the smaller separation position and larger than the value corresponding to the larger separation position. In short, when the first estimated values calculated for each frame by the error calculation unit 120 are sorted in the order of their size, they belong to a range determined according to a predetermined number of values from both ends or a predetermined number of values from both ends. What is necessary is a process for setting an unexisting value as an outlier.

(2) The statistical processing executed by the statistical processing unit 130 of the above embodiment is processing based on a deterministic approach, and first statistical filter processing that excludes outliers using a quartile method or the like; By selecting a representative value (a median value in the above embodiment) from the processing result of the first statistical filter processing, and using the value as an estimated value of the sampling frequency error over the entire correction target signal, Was composed. However, the first estimated value calculated for each frame is statistically modeled with an exponential function family, and the process for calculating the second estimated value by estimating the model parameter may be adopted as the statistical process. good. Specifically, for example, the above-described modeling is performed using a Laplace distribution, the shape of the distribution is determined by estimating the parameters of the Laplace distribution, a mode value is obtained from the determined distribution, and the mode value is calculated as the second value. The estimated value of the sampling frequency is estimated so that the likelihood of the sampling frequency error is likely.

(3) The time shift amount calculation unit 110 of the above embodiment calculates the time shift amount based only on τ that maximizes the cross correlation function C (τ). It is also possible to leave M pieces of τ as candidates and calculate the amount of time deviation based on these M pieces of τ. For example, the amount of time deviation is calculated from the average value of these M τ. Alternatively, all τs for which the value of the cross-correlation function C (τ) is equal to or greater than a predetermined threshold may be set as candidates for the amount of time deviation. In this case, a normalized cross-correlation function may be used to avoid the influence of power magnitude.

(4) The time shift amount calculation unit 110 may exclude a frame in which the powers of the reference signal and the correction target signal are less than a predetermined threshold from the time shift amount calculation target. If the threshold is set to an appropriate value, a frame that does not include the reference signal with sufficient intensity and a frame that does not include the correction target signal with sufficient intensity are excluded from the calculation target of the time shift amount. Frames that do not contain the reference signal with sufficient intensity and frames that do not contain the signal to be corrected with sufficient intensity have little contribution to the estimation of the sampling frequency deviation in the first place, even if the time deviation amount is calculated for such a frame. It will contain many errors. The first estimated value calculated based on such a time deviation amount is likely to be excluded as an outlier by the first statistical filter processing unit 130a as an outlier, and the calculation of the time deviation amount itself is useless in the first place. Become. According to such an aspect, it is possible to calculate the sampling frequency error with high accuracy while avoiding unnecessary calculation in the time shift amount calculation unit 110.

(5) The time deviation amount calculation unit 110 may exclude a frame in which the maximum value of the cross-correlation function C (τ) is below a predetermined threshold from the calculation target of the time deviation amount. If the threshold value is set to an appropriate value, the time shift amount is not calculated based on the cross-correlation function that is lower than the threshold value. Even if the amount of time deviation is calculated based on the cross-correlation function below the threshold value, it includes a lot of errors, and the first estimated value calculated based on the amount of time deviation is excluded as an outlier. In the first place, the calculation of the amount of time shift itself becomes useless. Also according to such an aspect, it is possible to calculate the sampling frequency with high accuracy while avoiding unnecessary calculation in the time deviation amount calculation unit.

(6) In the above embodiment, the frame size when the reference signal and the correction target signal are divided into frames is fixed. However, in such an aspect, the sample deviation of both signals increases as the frame number increases, and the cross-correlation function It is conceivable that the calculation of C (τ) becomes meaningless (or the cross-correlation function C (τ) cannot be calculated). Therefore, the processing may be stopped when the maximum value of the cross-correlation function C (τ) falls below a predetermined threshold, and some notification may be given to the user of the sampling frequency estimation apparatus 10, and the frame size is increased. Thus, the frames of the reference signal and the correction target signal may be re-divided.

(7) In the above embodiment, the amount of time deviation between both signals in each frame is calculated by calculating the cross-correlation function between the reference signal and the correction target signal for each frame. However, it is of course possible to calculate the cross-correlation between both signals after normalizing both signals and calculate the amount of time deviation based on the calculation result. In addition, a process for calculating a difference signal between the two signals while shifting one of the reference signal and the correction target signal in the time axis direction with respect to the other is executed for each frame, and the maximum value (or power) of the amplitude of the difference signal is set. It may be calculated as a value representing the correlation between both signals, and the amount of time deviation between the frames of both signals may be calculated based on the calculation result, or both ratios may be calculated by calculating the ratio between the sum signal and the difference signal of both signals. A value representing the correlation between the signals may be calculated for each frame, and the time shift amount for each frame of both signals may be calculated according to the calculation result. Further, a value representing the correlation between the reference signal and the correction target signal may be calculated for each frame by pattern matching, and the time shift amount for each frame of both signals may be calculated according to the calculation result. The point is that the correlation between the reference signal and the correction target signal is calculated for each frame, and the time deviation amount for each frame of both signals is calculated according to the calculation result.

DESCRIPTION OF SYMBOLS 1 ... Signal processing system, 10 ... Sampling frequency estimation apparatus, 20 ... Time-axis companding apparatus, 100 ... STFT part, 110 ... Time deviation calculation part, 120 ... Error calculation part, 130 ... Statistical processing part, 130a ... 1st Statistical filter processing unit, 130b... Second statistical filter processing unit.

Claims (5)

One of a plurality of signals obtained by independently sampling the same waveform is used as a reference signal, one of the remaining signals is used as a correction target signal, and one of the reference signal and the correction target signal is set as a time. Calculating a correlation between both signals for each frame while shifting in the axial direction, and calculating a time shift amount for each frame according to the calculation result;
An error calculation unit that calculates, for each frame, a first estimated value that is an estimated value of an error in the sampling frequency of the correction target signal in each frame from the time shift amount calculated by the time shift amount calculation unit;
A statistic that performs statistical processing on the first estimated value calculated for each frame by the error calculating unit to calculate and output a second estimated value that is an estimated value of the sampling frequency error over the entire correction target signal. A processing unit;
A sampling frequency estimation apparatus comprising: The statistical processing executed by the statistical processing unit includes
A first statistical filtering process that excludes outliers that are statistically estimated to contain many errors from the first estimated value calculated for each frame by the error calculating unit;
One of a filter process for smoothing a group of first estimated values from which outliers have been excluded from the first statistical filter process and a filter process for selecting a representative value from the group of first estimated values is a second one. Included as statistical filtering,
The sampling frequency estimation apparatus according to claim 1, wherein the statistical processing unit outputs a processing result of the second statistical filter processing as the second estimated value.   The first statistical filter processing includes a predetermined number of values from both ends or a predetermined number of values from both ends when the first estimated values calculated for each frame by the error calculation unit are sorted in order of size. The sampling frequency estimation apparatus according to claim 2, wherein the sampling frequency estimation apparatus is a process of setting an outlier as a value that does not belong to a range determined according to.   The time shift amount calculation unit calculates the power of each of the reference signal and the correction target signal for each frame, and excludes a frame in which the power of at least one signal is less than a predetermined threshold from the calculation target of the time shift amount The sampling frequency estimation apparatus according to any one of claims 1 to 3, wherein 5. The time shift amount calculation unit excludes a frame whose value representing a correlation between the reference signal and the correction target signal is below a predetermined threshold from the time shift amount calculation target. The sampling frequency estimation apparatus according to claim 1.

Images