Real-time Phase-Isolation Algorithm for Speech Separation

Beamforming techniques are applied to microphone arrays with the aim of separating sources and improving intelligibility, by means of spatial filtering. The non-stationary nature of speech implies the use of adaptive beamformers and several solutions have been implemented. Furthermore, interfering signals coming from the same direction as the target signal, cannot be filtered by the beamformer. The method presented in this paper is an alternative to adaptive beamforming, combining a simple delay-and-sum beamformer with a time-frequency masking method based on phase information. The beamformer is steered to the desired source and a function related to the phase differences between the steered signals at the microphones is evaluated to reject any interference that passed through the beamformer. Thus, the algorithm does not need to constantly adapt the filter coefficients and takes advantage of both beamforming properties and time-frequency separation techniques. The separation performance of the method has been evaluated in a noisy and reverberant environment using different arrays, talkers and scenarios. Real data are used to show the performance of the real-time algorithm when isolating one of the sources in the mixture.