As the performance of single-channel speech separation systems has improved, there has been a shift in the research community towards tackling more challenging conditions that are more representative of many real-world applications, including the addition of noise and reverberation. The need for ground truth in training state-of-the-art separation systems leads to a requirement of training on artificial mixtures, where single-speaker recordings are summed digitally. However, this leads to two separate approaches for creating noisy mixtures: one in which noise has been artificially added, maintaining perfect ground truth information, and one in which the noise is already present in the single-speaker recordings, allowing for in-domain training. In this work, we document a severe negative impact in both training and evaluation of models in the latter paradigm. We provide an explanation for this – the implicit task of separating noise – and propose an improved training objective that allows errors resulting from failing to separate noise to be minimized.
ASJC Scopus subject areas