1. Field of the Invention
The present invention relates to a system for reducing a noise element from a noise superposed voice signal such as environmental noise, etc., and more specifically to a noise reduction apparatus and a noise reducing method for reducing a noise element from a nonvoice environmental noise superposed voice signal input from a microphone in, for example, a mobile telephone system, an IP phone system, etc., improving a signal-to-noise ratio (SNR), and enhancing the speech communication quality.
2. Description of the Related Art
Recently, digital mobile communications systems such as mobile telephones, etc. have become widespread. In such communications, the communications are commonly established with large environmental noise, and it is important to effectively suppress the noise element contained in a voice signal.
In the above-mentioned noise suppression technology, for example, an input signal on a time axis is converted into a signal on a frequency axis (amplitude spectrum and phase spectrum), a suppression gain is obtained from the background noise estimated by a signal of a nonvoice interval, an amplitude spectrum is suppressed, the phase spectrum and the suppressed amplitude spectrum are restored into a signal on a time axis, thereby eliminating the noise (
The problem with the above-mentioned conventional technology is described below by referring to the following four documents.
[Nonpatent Document] S. F. Boll, “Suppression of Acoustic Noise in Speech Using Spectral Subtraction”, IEEE Transaction on Acoustics, Speech, and Signal Processing, ASSP-33, vol. 27, pp. 113-120, (1979)
[Patent Document 1] Japanese Patent Publication No. 3269969 “Background Noise Elimination Apparatus
[Patent Document 2] Japanese Patent Publication No. 3437264 “Noise Suppression Apparatus”
[Patent Document 3] Japanese Patent Application Laid-open No. 2002-73066 “Noise Suppression Apparatus and Noise Suppressing Method”
In Nonpatent Document 1, the technology of spectrum subtraction, obtaining suppressed amplitude spectrum by subtracting the amplitude spectrum of the estimated noise from the input amplitude spectrum, is proposed.
In Patent Document 1, an input signal is converted into a signal on a frequency axis, and a suppression gain is calculated based on the signal-to-noise ratio (SNR) calculated from the input signal and the estimated noise. The method of calculating a suppression gain is to empirically set a relational expression between the SNR and the suppression gain.
In Patent Document 2, when the power in the estimated nonvoice interval is small, the suppression level is lowered to avoid the degradation by suppressed voice interval of small power. When the power in the nonvoice interval is large, the suppression level is enhanced to further suppressing the nonvoice interval, thereby more appropriately suppressing the noise in the nonvoice interval.
In Patent Document 3, the power of a voice signal is obtained from the smoothing spectrum power in a voice-recognized interval, and the power of a no-voice signal is obtained from the smoothing spectrum power in a voice-unrecognized interval, thereby calculating the SNR, strongly suppressing noise on the signal portion having a high SNR, and restricting suppression on the portion distorted by suppression.
However, in the above-mentioned conventional technology, when the estimation of the background noise is incorrect, no appropriate suppression gain can be obtained, and the noise-suppressed voice signal is degraded. For example, when much bubble noise (background noise containing human voice) is contained in the background noise, the interval of bubble noise is not determined as a nonvoice interval, and estimated noise is calculated in an interval of constant noise other than the bubble noise. When the power of the constant noise is smaller than the power of the bubble noise, the estimated noise is underestimated in bubble noise interval, thereby causing insufficient suppression, that is, sufficient suppression cannot be realized.
In Patent Document 2, the power in the estimated voice interval is estimated as the maximum value of the short interval power in a long interval without considering the distribution of voice power. When the distribution of voice power changes depending on the characteristic of human voice and the speaking style is not considered, there is the problem that an appropriate suppression coefficient cannot be necessarily calculated. For example, when the distribution of the voice power is widely performed, there is voice having small power although the maximum value of the voice power is large. Therefore, the voice can be degraded if the suppression is too strong.
Thus, since the pure voice power, which is obtained by subtracting the noise element from an input voice signal, is not detected and its distribution is not estimated in the conventional technology, an appropriate suppression gain cannot be calculated when the background noise is mistakenly estimated.
The present invention has been developed to solve the above-mentioned problems, and aims at providing a noise reduction apparatus and a noise reducing method capable of appropriately suppressing noise when there is various background noise by estimating the information about the pure voice power contained in an input voice signal, and calculating a suppression gain based on the distribution and the range of voice power.
The first noise reduction apparatus according to the present invention having an analysis unit for analyzing the frequency of an input voice signal and converting the signal into a signal of a frequency area, a suppression unit for suppressing the signal of the frequency area, and a synthesis unit for synthesizing and outputting a suppressed signal of a time area using the suppressed signal of the frequency area includes: a voice information estimation device for estimating, using output of the analysis unit, the information for use as basic information in calculating a suppression gain of a signal, which is the information corresponding to at least the pure voice element excluding a noise element in the input voice signal; and a suppression gain calculation device for calculating the suppression gain corresponding to the output of the voice information estimation device and the analysis unit, and providing a calculation result for the suppression unit.
The second noise reduction apparatus according to the present invention having an analysis unit for analyzing the frequency of an input voice signal and converting the signal into a signal of a frequency area, a suppression unit for suppressing the signal of the frequency area, and a synthesis unit for synthesizing and outputting a suppressed signal of a time area using the suppressed signal of the frequency area includes: a noise estimation device for estimating the spectrum of a noise element in the input voice signal; a voice information estimation device for estimating, using output of the analysis unit, the information for use as basic information in calculating a suppression gain of a signal, which is the information corresponding to at least the pure voice element excluding a noise element in the input voice signal; and a suppression gain calculation device for calculating the suppression gain corresponding to the output of the noise estimation device, the voice information estimation device, and the analysis unit, and providing a calculation result for the suppression unit.
The first noise reducing method according to the present invention reduces noise using an analysis unit for analyzing the frequency of an input voice signal and converting the signal into a signal of a frequency area, a suppression unit for suppressing the signal of the frequency area, and a synthesis unit for synthesizing and outputting a suppressed signal of a time area using the suppressed signal of the frequency area, and performs: estimating, using output of the analysis unit, the information for use as basic information in calculating a suppression gain of a signal, which is the information corresponding to at least the pure voice element excluding a noise element in the input voice signal; calculating the suppression gain corresponding to the estimated voice information and the output of the analysis unit, and providing a calculation result for the suppression unit.
The second noise reducing method according to the present invention reduces noise using an analysis unit for analyzing the frequency of an input voice signal and converting the signal into a signal of a frequency area, a suppression unit for suppressing the signal of the frequency area, and a synthesis unit for synthesizing and outputting a suppressed signal of a time area using the suppressed signal of the frequency area, and performs: estimating the spectrum of a noise element in the input voice signal; estimating, using output of the analysis unit, the information for use as basic information in calculating a suppression gain of a signal, which is the information corresponding to at least the pure voice element excluding a noise element in the input voice signal; calculating the suppression gain corresponding to the estimated noise element spectrum, the estimated voice information, and the output of the analysis unit, and providing a calculation result for the suppression unit.
The noise reduction apparatus 1 according to the present invention further comprises at least a voice information estimation device 5, and a suppression gain calculation device 6. The voice information estimation device 5 estimates as voice information, using a signal of a frequency area output by the analysis unit 2, for example, spectrum amplitude, the information which is the basic information for use in calculating a suppression gain of a signal and is the information corresponding to a pure voice element excluding at least a noise element in the input voice signal. The suppression gain calculation device 6 calculates a suppression gain corresponding to the output of the voice information estimation device 5 and the analysis unit 2, and provides the result to the suppression unit 3.
In the embodiment of the present invention, the voice information estimation device 5 can estimate the power of the pure voice element, or can estimate an average value of the power indicating the number of samples totalized from the largest power as a predetermined ratio of the number of samples in the power distribution in each frequency of pure voice for a plurality of previously input voice signal frames.
In this case, the suppression gain calculation device 6 can also calculate the suppression gain for the frame k based on the difference between the power average value PMAXki corresponding to the frequency index i of the frame k currently to be processed and the spectrum power Pki corresponding to the frame k.
Furthermore, according to the embodiment of the present invention, the voice information estimation device 5 can also calculate the power distribution of the noise superposed voice signal as an input voice signal in addition to the estimated value of the power distribution of the pure voice as the information corresponding to the pure voice element, as the information for use in calculating the suppression gain by the voice information estimation device 5 and provide a result for the suppression gain calculation device 6.
In this case, the voice information estimation device 5 can also estimate the probability density function corresponding to the power distribution of the pure voice using two average values of power indicating the number of samples totalized from the largest power in a predetermined ratio of the total number of samples in the power distribution in each frequency of pure voice for a plurality of previously input voice signal frames, and the suppression gain calculation device 6 can divide the power distribution into a plurality of intervals such that the number of samples totalized from the largest power can be a predetermined ratio of the total samples for each of the distribution of the pure voice power and the power distribution of the noise superposed voice signal as the output of the voice information estimation device 5, and can obtain the suppression gain based on the average value of the power in each of the plurality of intervals.
Furthermore, the noise reduction apparatus of the present invention further comprises a noise estimation device for estimating the spectrum of the noise element in the input voice signal in addition to the analysis unit 2, the suppression 5 unit 3, the synthesis unit 4, and the voice information estimation device 5, and the suppression gain calculation device calculates a suppression gain corresponding to the output of the noise estimation device, the voice information estimation device, and the analysis unit 2.
In the noise reduction apparatus, as described above, the voice information estimation device 5 can estimate the power of the pure voice signal, and can also estimate the average value of the power indicating the number of samples totalized from the largest power as a predetermined ratio of the total number or samples in the distribution of the pure voice power for the plurality of voice frames.
In this case, the suppression gain calculation device 6 can also calculate the suppression gain based on the difference between the power average value PMAXki and the spectrum power Pki and the difference between PMAXki and the spectrum noise Nki in response to the input of the power average value PMAXki, the spectrum noise Nki for the current frame as the output of the noise estimation device, and the spectrum power Pki of the current frame.
Otherwise, the suppression gain calculation device 6 can also estimate the lower limit of the pure voice power, calculate the frequency Hki in which inconstant noise has been detected in the plurality of previously input voice frame signals including the current frame using the estimation result, and calculate the suppression gain based on the difference between the power average value PMAXki and the spectrum power Pki, the difference between the power average value PMAXki and the spectrum noise Nki, and the frequency Hki in response to the input of the power average value PMAXki, the spectrum noise Nki, and the spectrum power Pki.
The noise reducing method according to the present invention reduces noise using the above-mentioned analysis unit, the suppression unit, and the synthesis unit, estimates, using the output of the analysis unit, the information for use as basic information in calculating a suppression gain of a signal, which corresponds to the pure voice element excluding the noise in the input voice signal, as voice information, calculates the suppression gain corresponding to the estimation result and the output of the analysis unit, and provides the result for the suppression unit.
The noise reducing method according to the embodiment of the present invention estimates the above-mentioned voice information, estimates the spectrum of the noise element in the input voice signal, calculates the suppression gain corresponding to the estimated voice information, the estimated noise spectrum, and the output of the analysis unit, and provides the result for the suppression unit.
According to the embodiment of the present invention, corresponding to the two methods, a program used to direct a computer to realize the noise reducing method, and a portable storage medium storing the program can also be applied.
According to the present embodiment, the power information about the pure voice can be estimated without estimating noise, and the suppression gain is calculated based on its distribution and range. Therefore, voice suppression can be realized without an influence of the noise estimating capability, thereby obtaining a high quality voice signal. Furthermore, in addition to the power distribution of the pure voice, the power distribution of the noise superposed voice can be used in calculating a suppression gain, and a suppression gain can be calculated with the influence of the noise power superposed on the voice interval. Therefore, the suppression gain can be more correctly obtained as compared with the conventional method of using the noise estimated value estimated in a noise interval even if inconstant noise is superposed.
Furthermore, according to the present invention, in addition to the estimated value of the power information about the pure voice, the noise is further estimated, and the suppression gain is calculated using the result, the suppression gain can be calculated based on the power distribution of the pure voice, the range of the location, and the noise power estimated. Therefore, even if inconstant noise is superposed, the suppression gain can be more correctly obtained as compared with the conventional method using the estimated noise value calculated simply in a noise interval. Furthermore, the suppression gain can also be calculated using the frequency of inconstant noise. Therefore, the noise can be more correctly suppressed, and, for example, the communications quality in a mobile communication can be much improved.
[Nonpatent Document 2] Tsujii, Kamata “Digital Signal Processing Series vol. 1, Digital Signal Processing” 94 to 120 page, published by Shoko Do
[Nonpatent Document 3] Curtis Road, translated by Aoyagi, etc. “Computer Music] pp. 452-457, published by Tokyo Denki University.
The spectrum amplitude as the output of the analysis unit 11 is provided for a voice estimation unit 12, a suppression gain calculation device 14, and a suppression unit 15. The voice estimation unit 12 estimates the information corresponding to the element excluding the noise from the noise superposed input voice signal using the spectrum amplitude of the input signal, that is, corresponding to the pure voice signal, that is, the voice information for use in calculating a suppression gain. In the first embodiment, instead of calculating a suppression gain by estimating noise as explained by referring to
A spectrum power storage unit 13 stores the value of the spectrum power corresponding to, for example, the past 100 frames, and provides it for the voice estimation unit 12 and the suppression gain calculation device 14.
The suppression gain calculation device 14 calculates the suppression gain for adjustment of the spectrum amplitude using the voice information as the output of the voice estimation unit 12 and the spectrum amplitude of the input signal. The suppression unit 15 calculates the suppressed spectrum amplitude using the value of the calculated suppression gain and the spectrum amplitude of the input signal, and provides the result for a synthesis unit 16.
The synthesis unit 16 converts the signal on the frequency axis into a signal on the time axis by an inverse fast Fourier transform IFFT using the suppressed spectrum amplitude and the spectrum phase output by the analysis unit 11, overlaps it on the suppressed voice on the time axis in the previous frame in the overlapping calculation, and outputs the result as the suppressed output voice signal. Described above are the operations of the noise reduction apparatus 10, but the output signal of the synthesis unit 16 is, for example, provided for a voice coding unit 17, and the coding result is transmitted by a transmission unit 18, thereby applying to the voice communications system.
The reason why the synthesis unit 16 overlaps the signal converted on the time axis and the suppressed voice on the time axis in the previous frame in the overlapping addition is that the signal reduced outside the window by the window process in the FFT can be corrected, which is generally executed as the well-known technology.
Then, in step S3, the voice information is estimated. In this example, the voice information as the basic information in calculating a suppression gain is calculated using the spectrum amplitude SAki of an input signal, and the details are described later. The suppression gain Gki is calculated from the voice information calculation result in step S4, and the suppressed amplitude spectrum SA′ki is calculated using the next equation (1) in step S5.
SA′ki=SAki·Gki 0≦i<N (1)
Using the suppressed amplitude spectrum SA′ki and the spectrum phase SPki, the IFFT is performed in step S6, and voice is synthesized by an overlapping addition. In step S7, it is determined whether or not the processes on all input frames have been completed. When it is determined that the processes on all input frames have not been completed, the processes in and after step S1 are repeated. If it is determined that the processes on all frames have been completed, the current process terminates.
wkt=Ht·xkt t=0, . . . , 2N−1 (2)
Then, in step S12, the FFT process is performed on a window signal, and a real part XRki and an imaginary part XIki are obtained as a result. Then, in step S13, the spectrum amplitude SAki is obtained by the following equation (3).
SAki=(XRki2+XIki2)1/2 0≦i<N (3)
Furthermore, in step S14, the spectrum phase SPki is calculated by the next equation (4), thereby terminating the process.
SPki=tan−1(XIki/XRki) 0≦i 21 N (4)
In the equations above, 2N indicates the number of points on the FFT, for example, 128 and 256, and the window function Ht is, for example, a Hamming window.
Pki=SAki2 0≦i<N (5)
Then, in step S17, in an arbitrary period, for example, corresponding to 100 frames in a monitoring period including the current frame, the distribution of the spectrum power is obtained for each frequency (band) index i using the calculated spectrum power. For example, the spectrum power for the higher 10%, that is, the value of 10 spectrum power, is extracted. In step S18, the higher 10%, that is, the average value PMAXki of the spectrum power at a predetermined higher rate, is calculated and output as the voice information to be output by the voice estimation unit 12, thereby terminating the process.
dki=PMAXki−Pki 0≦i<N (6)
Then, in step S21, the suppression gain Gki is calculated using the next equation (7), thereby terminating the process.
Gki=f(dki) 0≦i<N (7)
Another embodiment of the voice information calculating process in step S3 shown in
Then, in step S25, as in
Then, in step S26, the distribution of the pure voice power for each index i of the frequency is assumed to be the Gaussian distribution, and the standard deviation of the Gaussian distribution is calculated by the equation (8).
σki=(PMAX1ki−PMAX2ki)/(a1−a2) 0≦i<N (8)
Then, in step S27, the average m of the Gaussian distribution is calculated by the equation (9).
mki=PMAX1ki−a1·σki 0≦i<N (9)
Thus, based on the standard deviation and the average for the pure voice power, the probability density function of the voice power can be obtained by the following equation (10). In the equation, x indicates the pure voice power.
P1ki(x)={1/(2π)1/2} exp [−(x−mki)2/2 σki2] 0≦i<N (10)
In this example, it is assumed that the power distribution of the pure voice is the Gaussian distribution, but the probability density function can also be obtained by calculating the histogram of the pure voice power.
Then, in step S28 shown in
The practical example of calculating PMAX1ki and PMAX2ki in step S25 is described below further in detail. Assume that the value of the above-mentioned a1 is 3, and the value of a2 is 2, and the PMAX1ki is calculated such that it indicates the power value at a higher 0.3%, and the PMAX2ki is calculated such that it indicates the power value at a higher 4.6%.
That is, in calculating PMAX1ki, for example, the spectrum power of the past 1000 frames is arranged in order from the highest level, and the highest 6 levels are selected. That is, the power at a higher 0.6% is selected, and the average value of the selected spectrum power is obtained. In calculating PMAX2ki, for example, the spectrum power of the past 1000 frames is arranged in order from the highest level, and the highest 92 levels are selected. That is, the power at a higher 9.2% is selected, and the average value of the selected spectrum power is obtained.
First, the noise superposed voice power of the past 100 frames is arranged in order from the highest level, and the average value V2n of the noise superposed voice power of a higher 10 levels is calculated. That is, the average value of the highest 10 noise superposed voice power is assumed to be V21, the second highest 10 noise superposed voice power from the eleventh level is assumed to be V22, . . . , and the average value of ten noise superposed voice power from the 91st level is assumed to be V210. The average value of the pure voice power can also be obtained for the nth interval as V1n.
In step S33 shown in
The suppression gain Gikn obtained in step S33 is a discrete value obtained for each interval, Gikn is interpolated by the following equation (14) in step S34 to calculate the suppression gain as a function of the actual noise superposed voice power signal x, and a suppression gain function is calculated.
Then, in step S35, the value of the suppression gain Gik(x) is calculated using the value of the noise superposed voice power x of the current frame, and the value is output in step S36 and the process terminates.
The second embodiment of the present invention is described below.
If it is determined in step S63 that it is not a noise interval, the process on the frame terminates. If it is a noise interval, then the estimated spectrum noise Nki is updated in step S64.
In this updating process, the spectrum power (noise spectrum power) of the current frame (noise frame) and the calculated past noise spectrum power are multiplied by the respective contribution rates to update the noise spectrum power. Thus, the high frequency element of the power fluctuation for each frame can be eliminated. In this example, the estimated spectrum noise is updated by the following equation (15) where 4 indicates a constant corresponding to the above-mentioned contribution rate.
Nki=ξ·Pki+(1−ξ)N(k−1)i 0≦i<N (15)
When the process starts as shown in
d1ki=PAMXki−Pki 0≦i<N (16)
d2ki=PMAXki−Nki 0≦i<N (17)
Gki=g(d1ki, d2ki) 0≦i<N (18)
In the present embodiment, the suppression gain is determined with the pure voice power information and the noise power information taken into account using two values of d1ki and d2ki. That is, the larger the value of d1ki, the smaller the pure voice power, thereby reducing the suppression gain. In addition the larger the d2ki, the more discrete the distribution of the noise superposed voice power and the distribution of the constant noise power, thereby reducing the contained noise power and increasing the suppression gain. For display, using the equation (19), the function g for providing the suppression gain Gki is set.
g(d1ki, d2ki)=τ−κ·d1ki+μ·d2ki 0≦i<N (19)
PMINki=PMAXki−φki 0≦i<N (20)
In the equation (20), if the input level is constant, it is assumed that the actual width (difference between the largest and smallest power) φki of the pure voice power is assumed to be constant. The value of the actual width can be checked from the distribution of the pure voice power in advance, or can be calculated by assuming the distribution of the pure voice power as the Gaussian distribution, and multiplying the standard deviation σ obtained by observing the power of an input signal by a constant.
Then, in step S76 shown in
Hki=[{H(k−1)i·(k−1)}+1]/kNki+λ≦Pki≦PMINki (21)
Hki={H(k−1)i·(k−1)}/kPki<Nki+λ, PMINki<Pki (22)
That is, Nki+λ indicates the upper limit power of the noise, and frequency Hki of the inconstant noise can be calculated depending on the ratio of the frames having Pki between the upper limit value and the lower limit value PMINki of the distribution of the pure voice power to the total input frames.
Then, in step S77 shown in
Gki=h(d1ki, d2ki, Hki) 0≦i<N (23)
The function h in the equation (23) for calculation of the suppression gain Gki can be determined by, for example, the following equation (24).
h(d1ki, d2ki, Hki)=τ−κ·d1k1+μ·d2ki−ν·Hki 0≦i<N (24).
In
The noise reduction apparatus and noise reducing method according to the present invention have been described above, but the noise reduction apparatus can also be configured as a processor and a common computer system.
In
The storage device 24 can be various types of storage devices such as a hard disk, magnetic disk, etc. These storage devices 24 or ROM 21 store a program, etc. shown in the flowcharts in
The program can also be stored in the storage device 24 from a program provider 28 through a network 29 and the communications interface 23, or can be marketed, stored in a commonly distributed portable storage medium 30, set in the reading device 26, and can be executed by the CPU 20. The portable storage medium 30 can be various types of storage media such as a CD-ROM, a flexible disk, an optical disk, a magneto-optical disk, etc., and the program stored in the storage media is read by the reading device 26 and realizes the suppression of various types of noise including the bubble noise according to the embodiments of the present invention, etc.
Number | Date | Country | Kind |
---|---|---|---|
2003-404595 | Dec 2003 | JP | national |