2011 Signal Separation Evaluation Campaign
Determined convolutive mixtures under dynamic conditions

In this task performance are evaluated through BSS_EVAL and PEASS 2.0. The audio signals in the submission of "F.Nesta" are estimates of the stereo spatial images of the sources ("sim") while in the remaining submissions are generic single channel source signal estimates ("src"). However, some algorithms generate "src" files representing a close estimation of the spatial image of the sources at one channel. Therefore for a more exhaustive comparison we decided to evaluate all the submissions with metrics for both generic source signal estimation (up to an arbitrary linear distortion) and spatial image estimation. Specifically we use the SIR/SDR/SAR metrics defined as in "bss_eval_sources" and SIRi/SDRi/SARi/ISRi/OPSi/TPSi/IPSi/APSi defined as in PEASS 2.0. Performance are evaluated taking the images of two channels as reference (the channels used by each algorithm), time-aligned and rescaled to the estimated source signals. Additionally, as for BSS-EVAL, PEASS 2.0 is evaluated allowing an arbitrary linear time-varying distortion for the target, i.e. the target error distortion is forced to 0. In this case the related performance are indicated with OPS/IPS/APS. Note that similarly to the ISR in "bss_eval_source", the TPS cannot be computed since any distortion is allowed for the target. The filter length for the BSS_EVAL decomposition was set to 1024 (the reverberation time is high).

The evaluation is performed only considering the segments where at most two sources overlap to each other. Since it may happen that there are very small segments, we evaluate the performance only for those >= 1 sec. (with smaller segments BSS_EVAL/PEASS performance may get unstable).
In the "detailed" performance the mean (and standard deviation) values over all the segments are listed. In the "average" performance the mean (and standard deviation) among all the microphone spacings and setup conditions are listed.

For details about each algorithm, click on the corresponding researcher`s name. To listen the submitted audio signals, click on the corresponding "sig" links.

The audio files below are made available under the terms of the Creative Commons Attribution-NonCommercial-ShareAlike 2.0 license. The author of the mixtures is F. Nesta.

Dataset1 ("random source activity of multiple sources in multiple static locations")

Dataset2 ("a continuously moving active source overlapped with a source in a fixed or random location")