Conclusions

·   Still to do:

- integrate nonspeech model
- build iterative explanation framework


·   Summary:

- must exploit high-level constraints to analyze mixtures that include speech
- residual-explanation framework can integrate disparate models .. if they can `project'
need to estimate speech spectra (unnormalize)

·   Unresolved:

- residual is too little information - need masking
- better ASR i.e. track variability, don't ignore it
- bootstrapping: for now, only as good as ASR