- Data: https://github.com/falabrasil/speech-datasets
- Models: https://gitlab.com/fb-resources/kaldi-br
Some notes:
- TDNN-F took about 16h to train for 5 epochs on an RTX 3080.
- i-Vector train in about a day on 8 parallel jobs (2, 2, 2).
- Tests with online decoding + 4-gram carpa rescoring for FalaBrasil model.
- Tests with online decoding using only 1st pass for Vosk model.
- Beam was set to 10 and lattice-beam to 6.0 for fast inference.
- Some values make sense, others do not, especially CORAA's.
- Debugging with tri-sat decoding should be considered before next training.
| dataset | fb v0.1.1 | vosk pt v0.3 | odd |
|---|---|---|---|
| coddef | 2.48 | 9.17 | |
| cetuc | 6.39 | 17.59 | |
| constituicao | 3.18 | 10.74 | |
| coraa | 52.97 | 65.99 | ** |
| cv | 24.03 | 38.81 | |
| lapsbm | 9.84 | 21.25 | |
| lapsstory | 14.06 | 18.60 | |
| mls | 27.52 | 40.65 | * |
| mtedx | 31.04 | 37.65 | * |
| spoltech | 15.83 | 31.21 | |
| vf | 19.32 | 35.34 | |
| westpoint | 6.37 | 16.50 |
Grupo FalaBrasil (2022) - https://ufpafalabrasil.gitlab.io/
Universidade Federal do Pará (UFPA) - https://portal.ufpa.br/
Cassio Batista - https://cassota.gitlab.io/