On microphone arrangement for multichannel speech enhancement based on nonnegative matrix factorization in time-channel domain

In this paper, we investigate the relationship between the way microphones are arranged and the degree to which speech is enhanced using the transfer-function-gain nonnegative matrix factorization (NMF), which is an amplitude-based speech enhancement method that is suitable for use with an asynchronous distributed microphone array. In an asynchronous distributed microphone array, recording devices can be placed freely and the number of devices can be easily increased. Therefore, it is important that to determine the optimum microphone arrangement and the degree to which the performance is improved by using many microphones. We understood experimental evaluations to show that the performance by supervised NMF can achieve close to the ideal time-frequency masking with a sufficient number of microphones. We also show that the performance is better when more microphones are placed close to each source.