Improving the Separation of Concurrent Speech through Residual Echo Suppression

This paper investigates the use of acoustic echo cancellation components in a speech separation system. The basic system uses a classical beamformer architecture, which separates the speech from different speakers based on spatial diversity. In order to get a better suppression of concurrent speech, we add a residual echo suppression stage, which has originally been developed in the area of acoustic echo cancellation. The speech separation performance of the proposed system is evaluated by means of automatic speech recognition experiments. The results show a clear improvement over standard beamforming and postfiltering approaches, with a word error rate of 44.2% compared to 68.1% for a superdirective beamformer (SDB) and 59.8% for an SDB with Zelinksy postfilter.