Fisher Baseline Experiments
From SpeechWiki
First we need a reasonable traditional phone-based recognizer against which to compare. An implementation of it using GMTK and associated scripts is here.
It uses the phone set, the multi-pronunciation phonetic dictionary and MVA-normalized PLP observations described in Fisher Corpus.
Monophone Model
The initial model is monophone, with the number of states per phone specified in phone.state. The number of gaussians per state will be determined by tuning.
Training on the entire fisher training set (1775831 utterances, specified in Fisher Corpus), takes an exceedingly long time: ~15 hours for a single EM iteration, on a 32 cpu cluster. Possible solutions are:
- Do training on data where word boundaries are observed. (Thanks Chris) For this we need to force-align the entire fisher corpus using some existing recognizer.
- Use a smaller corpus.
- Use teragrid (our allocation of 30000 cpu-hours is only 40 days on our cluster - almost not worth the effort of porting)
- Fool around with better triangulations (not much hope for improvement since the RVs with each frame are densely connected in our DBN)
Decoding
Preliminary Results
Tested on the first 500 utterances of the Dev Set.
Lang Model | Config | WER |
---|---|---|
10k bigram | config 0 | 67.3% |
10k trigram | config 9 | 67.2% |
5k trigram | config 10 | 67.7% |
1k trigram | config 11 | 69.5% |
500 trigram | config 12 | 71.6% |
Uses the 20k vocabulary language model described in Fisher Corpus.