Home

"Multichannel Linear Prediction for Blind Audio Source Separation", İ. Bayram, S. Bulek, January, 2017. [arXiv]

Experiments

Below are some results from the manuscript.

Source Separation Experiment (Real Recordings)

The original signal is the mixture of two speakers recorded in a reverberant room with a circular microphone array. Below is the observation recorded at the first microphone.

Eight such observations are available. In order to demonstrate the effects of using an MCLP preprrocessing step, we provide the separation results with and without MCLP for the two sources. The sources are separated by approximately 0.75 π radians. For separation, other than MCLP, a geometric source separation algorithm followed by a simple post-filter is utilized. The details can be found in the manuscript.

Minimum Variance Distortionless Response (MVDR) Processing

The primary claim in the paper is that MCLP boosts the performance of the subsequent separation step. In order to demonstrate that MCLP complements separation methods other than that described in the paper, we also tried separating the sources in the synthetic experiment above using the MVDR beamformer. For comparison, we provide the reconstructions obtained using MVDR, and MCLP followed by MVDR.

Synthetic Source Separation Experiment

In order to create the observations, we used two clean sources and the impulse responses obtained with the sine-sweep method. The impulse responses belong to the same circular array in the previous experiment but obtained in a relatively less reverberant room. The directions of the sources are separated by approximately 0.5 π radians.

The mixture signal observed at the first microphone is right below. Both sources are clearly reverberant.

Eight such observations are available. The same algorithms as in the previous experiment are used, to demonstrate the effect of using an MCLP step.

After applying MCLP, reverberation is supressed significantly. Below is the MCLP output when the first microphone is selected as the reference.

The effect of MCLP may also be observed by viewing the DOA function (see the manuscript for a description). The DOA functions with and without MCLP are shown below. The actual directions of the sources are 0 and 0.5 π radians. Notice that MCLP reveals these directions. Interestingly, MCLP also reveals a small bump between π and 1.5 π radians, possibly a dominant reflection.

Following MCLP, we apply the separation algorithms as outlined in the manuscript. Below are the results. For comparison, we also provide the outputs of the same algorithms when no MCLP is applied.

Synthetic Experiment with Three Sources

Notice that the sources are reverberant. Applying MCLP by taking the first microphone as the reference produces the following signal.

The DOA functions before and after applying MCLP are shown below. Notice that, as in the experiment with two sources, MCLP noticeably suppresses the reflections and maintains the primary source direction.

The reconstructions are obtained similarly as described in the paper. We apply GSS followed by a post-filter. Again, applying MCLP before separation noticeably improves the reconstructions.