Framewise Finite Impulse Response Filtering Based on Time-Frequency Mask for Low-Latency Speech Enhancement

We propose a low-latency speech enhancement method using framewise finite impulse response (FIR) filters based on time-frequency (T-F) mask. In many real-time audio applications, such as hearing aids, low-latency processing is highly required. T-F masking-based speech enhancement algorithms improve speech intelligibility for hearing impaired people in noisy environments, but an algorithmic delay due to frame analysis occurs. To shorten this delay, we replace time-frequency masking with framewise filtering in the time domain. The filters are designed on the basis of the signal and noise spectra in each frame, which are the same information used as that for designing a Wiener-filter-based T-F mask. The latency is shortened by designing a causal FIR filter and predicting the signal and noise spectra only from the information in the past frames. Evaluation experiments showed that causal framewise FIR filtering reduced the delay with little degradation of the performance compared with T-F masking.