A Casa Front-end Using the Localisation Cue for Segregation and Then Cocktail-party Speech Recognition

We propose and test a cocktail-party recognition technique based on segregation applied before recognition. This CASA front-end uses the TDOA (Time Delay Of Arrival) evaluated within subbands in order to determine the Relative Level (RL) of two competing speech sources. To perform the evaluation of the model, we have recorded a stereo database ST-NB95 from the mono Numbers95. This is composed of binary mixtures of sentences at 0dB, spatially placed left and right. With variation of frame duration and bandwidth, we quantify the accuracy of reconstruction of the original sources and we evaluate the recognition score on this new database. This work is a part of a triangular comparative study based on the processing of this database, and addressing the robust cocktail-party speech recognition paradigm.