Blind speech separation in reverberant rooms: effects of microphone positioning

Speech processing is essential in communication systems, where speech signals often suffer from environmental noise and interference from other speakers. Blind speech separation techniques help isolate or enhance specific speech signals in the mixtures. However, in reverberant rooms, this task becomes more complex due to the convolutional nature of the mixtures. Since separation algorithms rely on spatial information, microphone positioning affects performance. This study investigates these effects by comparing separation performance for different microphone positions in reverberant rooms. Two microphones were placed at varying distances from the sources, averaging 210, 140, and 75 centimeters. Another variation analyzed was the distance between microphones: 0, 4, and 57 centimeters. Mixtures of two fixed speech sources were recorded in four rooms with reverberation times of 162, 338, 534, and 1606 milliseconds. The recorded mixtures were separated using different Independent Vector Analysis algorithms. Performance was evaluated using the Scale-Invariant Signal-to-Distortion Ratio and Perceptual Evaluation of Speech Quality. Results showed that performance depended more on spatial information than the algorithms. For each room, increasing the distance between microphones improved performance, though this effect diminished with higher reverberation times. Variations in microphone and source distances did not produce statistically significant effects at any specific reverberation time.