A consolidated view of loss functions for supervised deep learning-based speech enhancement

摘要

Deep learning-based speech enhancement for real-time applications recently made large advancements. Due to the lack of a tractable perceptual optimization target, many myths around training losses emerged, whereas the contribution to success of the loss functions in many cases has not been investigated isolated from other factors such as network architecture, features, or training procedures. In this work, we investigate a wide variety of loss spectral functions for a recurrent neural network architecture suitable to operate in online frame-by-frame processing. We relate magnitude-only with phase-aware losses, ratios, correlation metrics, and compressed metrics. Our results reveal that combining magnitude-only with phase-aware objectives always leads to improvements, even when the phase is not enhanced. Furthermore, using compressed spectral values also yields a significant improvement. On the other hand, phase-sensitive improvement is best achieved by linear domain losses such as mean absolute error.