Blind audio source separation using short+long term AR source models and iterative itakura-saito distance minimization

                    Blind audio source separation (BASS) arises in a       number of applications in speech and music processing such       as speech enhancement, speaker diarization, automated music       transcription etc. Generally, BASS methods consider multichannel       signal capture. The single microphone case is the most       difficult underdetermined case, but it often arises in practice.       In the approach considered here, the main source identifiability       comes from exploiting the presumed quasi-periodic nature of       sources via long-term autoregressive (AR) modeling. Indeed,       musical note signals are quasi-periodic and so is voiced speech,       which constitutes the most energetic part of speech signals. We       furthermore exploit (e.g. speaker or instrument related) prior       information in the spectral envelope of the source signals via       short-term AR modeling. We present an iterative method based       on the minimization of the Itakura-Saito distance for estimating       the sources parameters directly from the mixture using a frame       based analysis.

Document Bibtex

Permalink: http://www.eurecom.fr/publication/3187