Patent Grant
7088831
1. Field of the Invention
The present invention is related to processing audio data, and more particularly to audio source separation in anechoic environments.
2. Discussion of Prior Art
There is increased interest in using microphone arrays in a variety of audio source separation and consequently speech processing applications. Small arrays of microphones have improved on single microphone systems in speech separation and directional detection of sources for hands free communication and a variety of other speech enhancement and audio source separation applications. Blind or parametric source separation approaches have been applied to distinguish between input from different microphones but with limited success. Challenges such as reverberation, noise, and acoustical echoes still plague many approaches to blind separation of audio signals.
One method for automatically compensating for attenuation due to differences in the calibration of the microphones has been attempted, which implements a deconvolution stage on the order of about a thousand taps. This is computationally expensive and may be difficult to implement in real-time.
A mixing model has been proposed wherein a decorrelation criterion is determined for integer delays, therefore the approach assumes that the distance between the microphones is less than the distance from the sources. However, such assumptions about sources being far-field may not hold well, and thus the model may not be a good approximation of the environment. Another proposed refinement to the mixing model includes higher order tap coefficients. The overall model corresponds to a constrained physical situation.
Another set of related spatial filtering techniques are antenna array processing techniques. Such techniques assume information about the microphone array layout as a given. For example, a delay and attenuation compensation (DAC) separation approach does not necessarily make this assumption, however weaker information such as the distance or a bound on the distance between sensors can help during a parameter estimation phase.
Still other proposed techniques use robust beamforming. Adaptive beamformers assume a known direction of arrival. Beamforming can be applied to source separation to deconvolve source estimates. Various source separation approaches have attempt to combine independent component analysis (ICA) or blind source separation (BSS) and elements of a beamformer to improve the performance of ICA/BSS techniques. However, no known system or method exists for real-time source separation by delay and attenuation compensation.
Therefore, a need exists for a system and method of real-time source separation by delay and attenuation compensation in a time domain.
According to an embodiment of the present invention, a method is provided for separating at least two audio channels recorded using an array of at least two microphones. The method equalizes variances of a first channel and a second channel on a current data frame, recursively expresses means and variances of mixtures, and normalizes the second channel to a variance level substantially similar to a variance of the first channel.
On a current block of m data samples xj (t), 1≦t≦m 1≦j≦2, and index k, a current block mean {overscore (x)}j can be determined according to:
A running mean {overscore (x)}j(k−1) can be updated by:
{overscore (x)}j(k)=(1−β){overscore (x)}j(k−1)+β{overscore (x)}j
where β is a learning rate.
A current block variance Varj is determined according to:
A running variance vj(k−1) is updated by:
vj(k)=(1−β)vj(k−1)+βVarj
Normalizing the second channel further includes normalizing an average energy to be similar to an average energy of the first channel according to:
The method determines delay parameters by minimizing a cross-covariance between two sources. The cross-covariance between the outputs is expanded as:
Ry
where Rx
The source separated outputs of the delay parameter estimation module are output in real-time.
The calibration module compensates for attenuations at the microphones.
The delay parameter determines relative delays of arrival of wave fronts at each microphone.
According to an embodiment of the present invention, a method is provided for separating at least two audio channels recorded using an array of at least two microphone. The method includes constraining a mixing model of the at least two audio channels in a time domain to direct path signal components, and defining a plurality of delays with respect to a midpoint between microphones, wherein delays depend on the distance between sensors and the speed of sound. The method further includes inverting a mixing matrix, corresponding to the mixing model, in the frequency domain, and compensating for a plurality of true fractional delays and attenuations in the time domain, wherein values of the delays and attenuations are determined from an output decorrelation constraint.
The method includes estimating a complex filter for each microphone, wherein the complex filters define the mixing model.
The mixing matrix corresponding to the mixing model comprises two delay parameters and two parameters corresponding to the speed of sound.
The output decorrelation constraint is a function of two unknown delays and unknown scalar coefficients. An attenuation coefficient has a value substantially equal to one.
The method imposes a minimum variance criterion for a reverberant case over all linear filtering combinations of X1 and X2.
According to an embodiment of the present invention, a program storage device is provided, readable by machine, tangibly embodying a program of instructions executable by the machine to perform method steps for separating at least two audio channels recorded using an array of at least two microphones. The method includes equalizing variances of a first channel and a second channel on a current data frame, recursively expressing means and variances of mixtures, and normalizing the second channel to a variance level substantially similar to a variance of the first channel.
Preferred embodiments of the present invention will be described below in more detail, with reference to the accompanying drawings:
a is a diagram of a system for executing code according to an embodiment of the present invention;
b is a diagram of a system for separating mixed input according to an embodiment of the present invention;
The present invention provides a system and method for separating two or more audio signals recorded using an array of microphones assuming an anechoic mixture model. The complexity and performance factors of an embodiment of the present invention has been measured. One with ordinary skill in the art will appreciate that various other embodiments can be built upon these results.
Although elements of the present invention are derived from blind source separation principles, the system and method implement an anechoic propagation model to reduce the complexity of the mixing model and make it possible to effectively identify and invert a mixing process using second ordered statistics. For sources far away from the microphone array, for example, greater than one meter, the model can be simplified to depend on just a few parameters. According to an embodiment of the present invention, these parameters include relative delays in the arrival of wave fronts and attenuations at the microphones. The method estimates the parameters of a mixture to compensate for the true values according to a delay and attenuation compensation (DAC) method.
It is to be understood that the present invention may be implemented in various forms of hardware, software, firmware, special purpose processors, or a combination thereof. In one embodiment, the present invention may be implemented in software as an application program tangibly embodied on a program storage device. The application program may be uploaded to, and executed by, a machine comprising any suitable architecture. Preferably, the machine is implemented on a computer platform having hardware such as one or more central processing units (CPU), a random access memory (RAM), and input/output (I/O) interface(s). The computer platform also includes an operating system and micro instruction code. The various processes and functions described herein may either be part of the micro instruction code or part of the application program (or a combination thereof) which is executed via the operating system. In addition, various other peripheral devices may be connected to the computer platform such as an additional data storage device and a printing device.
It is to be further understood that, because some of the constituent system components and method steps depicted in the accompanying figures may be implemented in software, the actual connections between the system components (or the process steps) may differ depending upon the manner in which the present invention is programmed. Given the teachings of the present invention provided herein, one of ordinary skill in the related art will be able to contemplate these and similar implementations or configurations of the present invention.
According to an embodiment of the present invention, to achieve a real-time implementation of a source separation model, an estimation of the direction of arrival can be determined based on a cross-covariance. Further, variations in the microphones, such as differences in gain, can be accounted for. The system and method have been evaluated using a segmental signal-to-noise ratio (SNR) measure for a large collection of data collected in both anechoic and echoic environments.
Referring to
A general convolutive model for the mixing of two source signals at two sensors can be written as:
x1(t)=h11s1(t)+h21s2(t)
x2(t)=s1(t)+s2(t) (1)
where hi represents unknown relative transfer functions of the first sensor versus the second sensor, t is time, wherein t is an index in the current frame of data, and 1 represents convolution. s1 and s2 are the source signals.
With a low complexity source separation method, the treatment of the mixing problem can be simplified by considering only direct path signal components, rather than using a general convolutive propagation model. The component from one source arrives at the sensors with a fractional delay between the time of arrival at two closely spaced sensors. The fractional delay is a delay between sensors that is not generally an integer multiple of the sampling period and depends on the position of the source with respect to the array axis and the distance between sensors. The DAC mixing model in the time domain can be written as the follows:
x1(t)=s1(t−δ1)+c1∃s2(t−δ2)
x2(t)=c2∃s1(t+δ1)+s2(t+δ2) (2)
where: c1, c2 are two positive real numbers, accounting for non-calibrated microphones, and for deviations from the far-field assumption. s1 and s2 are two sources, and x1 and x2 are mixtures at the respective microphones. Equation 2 describes a mixing matrix for the mixing model in the time domain, in terms of four parameters, δ1, δ2, c1, and c2.
According to an embodiment of the present invention this mixing matrix is inverted. This can be performed in the frequency domain, and results in the following time domain solution:
y1(t)=h(t,δ1,δ2,c1,c2)1(x1(t+δ2)−c1x2(t−δ2))
y2(t)=h(t,δ1,δ2,c1,c2)1(−c2x1(t+δ1)+x2(t−δ1)) (3)
where the convolutive filter h accounts for the division with the determinant of the mixing matrix. In practice the criteria above can be simplified to a decorrelation between fractionally delayed sensor recordings:
y1(t)=x1(t+d1)−c1x2(t)
y2(t)=c2x1(t+d2)+x2(t) (4)
This is possible due to the freedom to shift signals under the assumption of decorrelation at any lag.
The DAC method performs source separation by compensating for the true fractional delays and attenuations in the time domain with values determined from an output decorrelation constraint:
Ry1y2(τ)=E[y1(t)y2(t+τ)]=0,∀τ (5)
as a function of two unknown delays d1 and d2 and unknown scalar (attenuation) coefficients c1 and c2. E[●] is the time average of the quantity between square brackets. Attenuation coefficients c1 and c2 have values close to one (1) (e.g., c1≈c2≈1) under the far-field source assumption. This is equivalent to the following criterion:
A generalization of the solution in the reverberant case (e.g., Equation 1) can be obtained by imposing a minimum variance criterion, for example, argminGi1;Gi2Var(Yi−Si) over all linear filtering combinations of X1 and X2:
Yi=Gi1X1+Gi2X2 (7)
The implementation includes the estimation of complex filters H1 and H2 defining the mixing model in Equation 1:
Complexity and performance characteristics of the simple method, particularly on real environment data can influence decisions for more complex approaches to deal with reverberant conditions.
According to an embodiment of the present invention, the method can simplify the delay estimation by dealing with attenuations in a calibration phase and evaluating output decorrelation based on the covariance of the mixtures. Calibration can be performed online. The calibration accounts for dissimilarities in microphones, e.g., neither identical nor calibrated off-line.
Ideally, c1=c2=1 under the far-field assumption, and microphones have identical gain characteristics. In practice however, it can be difficult to impose the latter condition. Referring to
The running mean {overscore (x)}j(k−1) can be updated by, for example:
{overscore (x)}j(k)=(1−β){overscore (x)}j(k−1)+β{overscore (x)}j
where β is a learning rate, for example, β=0.1. The current block variance Varj can be determined according to, for example:
The running variance vj(k−1) can be updated by, for example:
vj(k)=(1−β)vj(k−1)+βVarj
The second channel can be normalized so that its average energy to be similar to that of the first channel:
The recursive formulas above have a direct online implementation. Furthermore, the attenuation parameters in Equation 4 can be dropped, simplifying the estimation of delays.
The cross-covariance between y1 and y2, the outputs, can be expanded as follows:
Ry
where Rx
Delay parameters can be estimated by minimizing this expression 904. Note that to determine sub-unit-delayed versions of cross-correlations, the delay parameters can be determined for a number of lags L.
A real-time application can be implemented as a multi-threaded Windows task on a Pentium III PC. The inputs can come from the auxiliary input of the standard PC sound card, while outputs are continuously streamed to, for example, headphones. At least one thread performs the I/O of audio data in real time. At least another thread is responsible for the analysis, calibration, delay estimation and synthesis of the demixed signals.
Calibrated data are fed into the delay parameter estimation module, which can use, for example, the Amoeba optimization method as taught by W. H. Press et al. Numerical Recipes in C. Cambridge University Press, 1988, to find a local solution. Delay values are constrained based on d, thus, the solution is global. Optimization uses the cost function (Equation 10), wherein an initial simplex can be selected including, for example, three pairs of delays. The initial simplex is centered at the delays of last data block (d1+0:05; d2+0:05), (d1−0:05; d2−0:05), and (d1+0:05; d2−0:05) (in samples). Solutions (d1*;d2*) of the optimization can be smoothed using a learning rate α, the equation can be written as follows:
dj=djk=(1−α)·djk−1+α·dj*,j=1,2 (11)
Delays can be sorted to insure stability to the permutation problem. The correspondence between delays and sources is unique when sources are not symmetrical with respect to the receiver axis. Thus, the sorted delays can be used to directly generate separated outputs.
According to an embodiment of the present invention, an important characteristic of the DAC approach is the artifact-free nature of the outputs.
A method implementing the present invention was evaluated on real data recorded in an anechoic room and in a strongly echoic environment. As shown in
The real-time method was successful in separating voices from anechoic mixtures, even when sources had similar spectral power characteristics. The method generally separates at least one voice in echoic voice mixtures, while achieving about three to four-dB segmental SNR improvement on average. A frame size of 512 samples was chosen.
For anechoic data sets,
Results for echoic data sets are shown in
The delay estimation method converges close to the true delay values provided voice is present after processing only about 150–200 milliseconds of anechoic data or about 2500 samples at 16 kHz sampling frequency.
The present invention has been tested on more than one thousand combinations of voices recorded in real anechoic and echoic environments. The performance of the system is good on anechoic data. Although the method is designed for anechoic environments, its complexity and performance on real data represent a basis for designing more complex approaches to deal with reverberant environments.
Having described embodiments for a method of audio source separation by delay and attenuation compensation, it is noted that modifications and variations can be made by persons skilled in the art in light of the above teachings. It is therefore to be understood that changes may be made in the particular embodiments of the invention disclosed which are within the scope and spirit of the invention as defined by the appended claims. Having thus described the invention with the details and particularity required by the patent laws, what is claimed and desired protected by Letters Patent is set forth in the appended claims.
| Number | Name | Date | Kind |
|---|---|---|---|
| 5208786 | Weinstein et al. | May 1993 | A |
| 5515445 | Baumhauer et al. | May 1996 | A |
| 5694474 | Ngo et al. | Dec 1997 | A |
| 6317703 | Linsker | Nov 2001 | B1 |
| 20010031053 | Feng et al. | Oct 2001 | A1 |
| Number | Date | Country | |
|---|---|---|---|
| 20030112983 A1 | Jun 2003 | US |
