Adaptive noise smoothing for speech enhancement filters

Noise estimation is one of the main keys for successful speech enhancement processing. Traditional noise estimators based on inter-frame smoothing of the estimated power spectrum (PSD) of the background noise leads to artificial artifacts on the residual noise well known as `musical noise´. In this paper we show the relationship of the estimated noise PSD and the denoised speech quality. This study leads to the development of a new technique to smooth the estimated noise power spectrum over frequency bins of the same frame. The smoothing processing is governed adaptively by the PSD of the noisy speech and the estimated noise. Compared to traditional inter-frame smoothing, experimental results show that this intraframe smoothing has a good impact on the denoised speech and the problem of musical noise is well reduced.