A High SIR Low-overhead Implementation of Single-channel Speech Source Separation

In the field of speech signal processing, speech source mixture separation is a known challenge. It is addressed by finding the closest estimate of the original speech source from the speech mixture. Source separation solutions can be based on multiple channels or single channel model. In multiple channels, multiple speakers and microphones are assumed while in single channel multiple speakers and a single microphone are assumed. One of the most widely used algorithms in the single-channel model is the Ideal Ratio Mask (IRM). Although IRM is efficient, it has a major drawback; the high memory footprint as it stores all frequency components of the Short-time Fourier transform (STFT). This makes it less suitable for embedded applications. We propose a solution based on the optimization of Mel-frequency Cepstrum Coefficient (MFCC) and Non-centroid K-nearest neighbor (Nk-nn) algorithms that minimizes memory utilization and achieves high Signal-to-Interference Ratio (SIR). Our experimental results show that the proposed solution improves SIR while minimizing memory requirements compared to the reference IRM.