Speech enhancement in modulation domain using discriminative random fields

We explore the utility of Discriminative Random Fields (DRF) in modulation domain for speech enhancement in stationary noise setting. The ideal channel selection framework is used in modulation domain. We extend the 2-D spectro-temporal DRF framework to model 3-D noisy observations in modulation domain. Each Time-(acoustic)Frequency-Modulation(frequency) bin is classified as speech or noise. Thus, a 3-D binary mask is to be inferred. Iterated Conditional Mode is used to estimate the binary mask; the continuity along acoustic and modulation frequencies is captured using Ising model. By capturing the continuity along modulation frequencies, there is no musical noise present in the enhanced speech. The quality of reconstructed speech in different noise settings (additive white Gaussian noise, pink, car and street noise) is measured in terms of average segmental Signal to Noise Ratio, Perceptual Evaluation of Speech Quality (PESQ) and Mean Opinion Score (MOS). We compare and contrast the performance of this modulation domain DRF with that of spectro-temporal DRF and iterative Wiener Filter.