Base model to copy decoder/encoder weights is nvidia/stt_en_fastconformer_transducer_xlarge
Best achieve validation WER is 6%
Hardware
1x H100 SXM
Training
training on huge dataset, running for 11 epochs, and about 210K global steps.
Inference
run single audio transcribe or batch mode inference, see inference.ipynb
Adapters
training adapter on ganji dataset including 10 hours of training with max_steps of 10000 training (about 40 epochs), 1 hour of validation and 1 hour of test.
Hardware to train adapter module, 1 RTX 6000Ada
| Metric | Without Adapter | With Adapter |
|---|---|---|
| WER | 31.20% | 28.22% |
| CER | 8.81% | 6.21% |