Speech Recognition

·   Major issues:

- isolated word or continuous
- speaker independent, adaptive or individual
- vocabulary size, (grammar complexity)
- signal quality / robustness


·   State of the art

- moderate perplexity, speaker-independent interactive telephone systems (stock quotes)
- transcription of TV broadcasts, conversations at 30-40% word error
- searching alternate Markov model hypotheses is large & slow: ~ real-time on fast CPU


·   Alternatives

- fixed small-vocabulary module
- `cheap & cheerful' trainable templates