Analyzing sound mixtures

·   Real sound = many sources (speech, ...)
    How to analyze / decompose?

·   The CASA approach

- bottom-up cues
- assemble into larger structures


·   ASR approaches

- train on mixed (average static noise)
- parallel models (decomposition)
    but:
- calculating joint probabilities
- relative levels (cepstral domain)
- suitability of state model for nonspeech?

·   What do people do?

- ?: hypotheses pruned by gen-purp bottom-up
- i.e. a combination...