Exploring Tradeoffs in Models for Low-latency Speech Enhancement
This work addresses speech enhancement for real-time applications, but it is incremental as it builds on existing spectrogram-mask-based methods with specific performance trade-offs.
The paper tackled the problem of low-latency speech enhancement by exploring neural network configurations, achieving a 0.4 dB SDR improvement on the CHiME2 task and showing that zero-look-ahead models can perform within 0.03 dB of the best bidirectional model.
We explore a variety of neural networks configurations for one- and two-channel spectrogram-mask-based speech enhancement. Our best model improves on previous state-of-the-art performance on the CHiME2 speech enhancement task by 0.4 decibels in signal-to-distortion ratio (SDR). We examine trade-offs such as non-causal look-ahead, computation, and parameter count versus enhancement performance and find that zero-look-ahead models can achieve, on average, within 0.03 dB SDR of our best bidirectional model. Further, we find that 200 milliseconds of look-ahead is sufficient to achieve equivalent performance to our best bidirectional model.