Accounting for deterministic noise components in a MMSE STSA speech enhancement framework

Current approaches to speech enhancement usually consider performance in either the presence of coloured broadband noise or periodic noise, but rarely both. In this paper we present a new speech enhancement technique, derived within the standard minimum mean-square error (MMSE) short time spectral amplitude framework, which jointly compensates for both coloured broadband and periodic noise components. With this approach, termed noise mean subtracted (NMS) MMSE, hidden Markov model based frequency tracking techniques are used to estimate periodic noise components and differentiate them from periodic components associated with the speech signal. They are then removed using complex spectral subtraction. The resulting algorithm is evaluated using perceptual evaluation of speech quality (PESQ) for both male and female speech utterances from the NOIZEUS database. It is shown that when the noise contamination comprises both broadband and periodic components, this NMS MMSE algorithm outperforms the standard MMSE algorithm, derived under the assumption of stochastic noise only. The technique has application in a variety of scenarios, including those involving emergency radio communications.