Abstract:
We provide a Matlab GUI for the T-ABCD method for blind separation of convolutive mixtures of audio sources. The algorithm works in the time-domain and is based on a complete unconstrained ICA decomposition of the observation space that is spanned by lagged input signals. Almost arbitrary ICA method can be used as the core mechanism of the separation. The method went through several modifications until the current version 2. Some functions have been removed, which simplifies the current GUI (but the original version 1 is available as well). The most important modification compared to version 1 resides in a new scale-invariant reconstruction of signals’ images (responses) on the microphones. In other words, the scaling ambiguity problem is efficiently resolved in version 2, and, therefore, it achieves significantly better signal-to-distortion ratio compared to version 1.
Matlab GUI package in m-code (March 2016): here
Version 1 (June 3, 2013): here.
Corresponding papers:
A detailed description in IEEE TASLP
[1] Z. Koldovský and P. Tichavský, “Time-Domain Blind Separation of Audio Sources on the basis of a Complete ICA Decomposition of an Observation Space”, IEEE Trans. on Speech, Audio and Language Processing, Vol. 19, No. 2, pp. 406-416, ISSN 1558-7916, February 2011. (here)
Pioneering papers
[2] Z. Koldovský and P. Tichavský, “Time-domain Blind Audio Source Separation Using Advanced Component Clustering and Reconstruction”, to be presented on The Joint Workshop on Hands-free Speech Communication and Microphone Arrays (HSCMA 2008), May 6-8, Trento, Italy, 2008. (here)
[3] Z. Koldovský and P. Tichavský, “Time-Domain Blind Audio Source Separation Using Advanced ICA Methods”, Proceedings of 8th Annual Conference of the International Speech Communication Association (Interspeech 2007), pp. 846-849, August 2007. (here)
Particular improvements and extensions
[4] J. Málek, Z. Koldovský, J. Ždánský and J. Nouza, “Enhancement of Noisy Speech Recordings via Blind Source Separation”, Proceedings of the 9th Annual Conference of the International Speech Communication Association (Interspeech 2008), pp. 159-162, ISSN: 1990-9772, September 22-26, Brisbane, Australia, 2008.(here)
[5] Z. Koldovský, P. Tichavský, and J. Málek, “Time-Domain Blind Audio Source Separation Method Producing Separating Filters of Generalized Feedforward Structure,” in Latent Variable Analysis and Signal Separation, Lecture Notes in Computer Science Vol. 6365, pp. 17-24, ISBN: 978-3-642-15994-7, Springer, Heidelberg, Sept. 2010.
[6] Z. Koldovský, P. Tichavský, and J. Málek, “Subband Blind Audio Source Separation Using a Time-Domain Algorithm and Tree-Structured QMF Filter Bank,” in Latent Variable Analysis and Signal Separation, Lecture Notes in Computer Science Vol. 6365, pp. 25-32, ISBN: 978-3-642-15994-7, Springer, Heidelberg, Sept. 2010.
[7] J. Málek, Z. Koldovský, and P. Tichavský, “Adaptive Time-Domain Blind Separation of Speech Signals,” in Latent Variable Analysis and Signal Separation, Lecture Notes in Computer Science Vol. 6365, pp. 9-16, ISBN: 978-3-642-15994-7, Springer, Heidelberg, Sept. 2010.
[8] Z. Koldovský, J. Málek, and P. Tichavský, “Blind Speech Separation in Time-Domain Using Block-Toeplitz Structure of Reconstructed Signal Matrices,” Interspeech 2011, Florence, Italy, Aug. 2011.
A complete SiSEC 2010 dataset for testing and benchmarks (task “Robust blind linear/non-linear separation of short two-sources-two-microphones recordings”): here
Instalation and execution of the T-ABCD package v2.0: Unzip the package into a directory, run Matlab, load microphone recordings into rows of a matrix x, and type (in Matlab)
>> tabcd2(x)
See the readme.txt file.
Example:
Another example:
