Improving Low-Latency Mono-Channel Speech Enhancement by Compensation Windows in STFT Analysis

Speech enhancement is a key component in voice communication technology as it serves as an important pre-processing step for systems such as acoustic echo cancellation, speech separation, speech conversions, etc. A low-latency speech enhancement algorithm is desirable since long latency means delaying the entire system’s response. In STFT-based systems, reducing algorithmic latency by using smaller STFT window sizes leads to significant degradation in speech quality. By introducing a simple additional compensation window along with the original short main window in the analysis step of STFT, we preserve signal quality – comparable to that of the original high latency system while reducing the algorithmic latency from 42 ms to 5 ms. Experiments on the full-band VCD dataset and a large full-band Microsoft’s internal dataset show the effectiveness of the proposed method.