STFT-Domain Neural Speech Enhancement With Very Low Algorithmic Latency

Deep learning based speech enhancement in the short-time Fourier transform (STFT) domain typically uses a large window length such as 32 ms. A larger can lead to higher frequency resolution and potentially better enhancement. This however incurs an algorithmic latency of ms online setup, because overlap-add algorithm used inverse STFT (iSTFT) is also performed using same size. To reduce this inherent latency, we adapt conventional dual-window-size approach, where regular input size for but shorter output overlap-add, STFT-domain deep frame-online Based on STFT-iSTFT configuration, employ complex spectral mapping enhancement, neural network (DNN) trained predict real imaginary (RI) components target from mixture RI components. In addition, use DNN-predicted conduct beamforming, results which are extra features second DNN perform postfiltering. The frequency-domain beamformer be easily integrated with our DNNs designed not incur any latency. Additionally, propose future-frame prediction technique further Evaluation noisy-reverberant shows effectiveness proposed algorithms. Compared Conv-TasNet, system achieve performance comparable amount computation, or less maintaining strong at low 2