Systems 2:
Hidden source states

Weintraub'85

goal: separate two speakers (for recognition)
· decide states of each voice (silent, voiced, u/v)
· derive parameters by state-dependent rules
output: spectrum estimates, masked resynthesis
Decide number of sources first

Moore'86

goal: separate speech from structured noise
· Hidden Markov Models for both sources
· integrate over all pairs that could give observed
output: optimal simultaneous state sequences
No elements -> no grouping