This test can be used to observe the\u00a0dependence of the algorithms on various factors (the length of data, the number of signals, computational burden, reliability, etc.). An important factor is the\u00a0character\/model\/origin of the signals. Here, we mix artificial and speech signals. The artificial signals\u00a0obey the standard ICA models described in the chapter (nongaussian i.i.d., piecewise Gaussian i.i.d, and Gaussian weak stationary – autoregressive). It means that some of these signals need not be separable by a given method that utilizes only some\u00a0signal diversities (nongaussianity, nonstationarity, non-whiteness).<\/p>\n
Various ICA\/BSS algorithms can be compared\u00a0this way. We recommend trying methods (but not only) from the following links:\u00a0MLSP-Lab<\/a>,\u00a0A.S.A.P.<\/a>,\u00a0Petr Tichavsky<\/a>,\u00a0ICALAB toolbox<\/a>,\u00a0FastICA<\/a>,\u00a0LVA central<\/a><\/p>\n The code of the test is available here<\/a>.<\/p>\n Here, the data of the example that is presented in the chapter are provided. The experimental setup is shown in the picture above.\u00a0Four loudspeakers and microphones were located in a room where the distance of the loudspeakers from microphones was 1.2 m. Each loudspeaker was situated at different angle.\u00a0Two male and two female utterances (dataset A<\/a> and dataset B<\/a>) were each played by the respective loudspeaker at a different angle. The spatial images (dataset A<\/a> and dataset B<\/a>) of the signals were recorded by the microphones. The mixed signals (dataset A<\/a> and dataset B<\/a>) are equal to the sum of the\u00a0spatial images of the individually recorded sources.<\/p>\n Three setups were considered with two (a and b), three (a through c) and four (a through d) sources. The number of used microphones was the same as that of the sources in each setup. The mixtures were separated by three methods: FD-ICA with activity sequence (power ratio) clustering (dataset A<\/a>\u00a0and dataset B<\/a>), FD-ICA with TDOA\u00a0estimations (dataset A<\/a>\u00a0and dataset B<\/a>), and FastIVA (dataset A<\/a>\u00a0and dataset B<\/a>). The results are available for different lengths of the STFT (from 128 through 4096).<\/p>\n This code<\/a>\u00a0simulates\u00a0acoustic mixtures of two, three or four speakers. Room impulse responses between the sources and microphones are generated by\u00a0the image method implemented by Emmanuel Habets; please download the package from\u00a0AudioLabs<\/a>. The sources and the microphones can be placed at any point in the room; default\u00a0locations are defined in the code.<\/p>\n The user can apply any BSS method to\u00a0the artificial\u00a0mixture. As a reference method, the time-domain T-ABCD algorithm is used; please download the T-ABCD package<\/a>. T-ABCD can be used to estimate only short filters (10-80 taps) since its computational complexity rapidly grows with the length of filters. It is robust but it cannot\u00a0handle long tails of the impulse responses (only direct path and early reflections) as the separating filters are short. The reader is encouraged to compare T-ABCD with the FD-ICA approaches described in the chapter.<\/p>\nSeparation of real-world recordings<\/h2>\n
![\"room\"](\"http:\/\/project.inria.fr\/ssse\/files\/2016\/08\/room-300x207.png\")
<\/a><\/p>\nSeparation of simulated convolutive\u00a0audio mixtures<\/h2>\n