Informed Single Channel Speech Separation with time-frequency exemplar GMM-HMM model

In recent studies, the problem of Single Channel Speech Separation (SCSS) have been efficiently tackled by introducing additional cues from the original target source in the form of Informed Source Separation (ISS). In this paper, a more realistic situation is considered where an additional user/listener generated Exemplar source is introduced to aid the separation process instead of using the original target source. The Exemplar source consists of patterns that need to be transformed, extracted, regulated and calibrated to generate an utterance dependent (UD) model that could statistically represent the target source. The proposed method uses general speaker independent (SI) features along with the generated UD features are modelled and combined in a joint probability model to achieve reliable separation. Unlike most model-based approaches, the proposed method does not require Speaker Dependent training on individual sources of the mixture, and is therefore much more efficient and less restrictive.