We discuss an identifification framework for noisy speech mixtures. A 

block-based generative model is formulated that explicitly incorporates 

the time-varying harmonic plus noise (H+N) model for a number of latent 

sources observed through noisy convolutive mixtures. All parameters 

including the pitches of the source signals, the amplitudes and phases of 

the sources, the mixing fifilters and the noise statistics are estimated by 

maximum likelihood, using an EM-algorithm. Exact averaging over the 

hidden sources is obtained using the Kalman smoother. We show that 

pitch estimation and source separation can be performed simultaneously. 

The pitch estimates are compared to laryngograph (EGG) measurements. 

Artifificial and real room mixtures are used to demonstrate the viability 

of the approach. Intelligible speech signals are re-synthesized from the 

estimated H+N models