LibriSpeech Conformer BPE-CTC
Final checkpoint for dla-asr-hw.
16 Conformer blocks, hidden size 512, 8 attention heads, RoPE, RMSNorm, 80 normalized log-mel features, 16 kHz input. The CTC vocabulary has 1025 entries: 1024 SentencePiece tokens shifted by one, plus blank at index zero.
This is a PyTorch state_dict for the repository's ConformerModel. It is not a Transformers from_pretrained model. Download the checkpoint and the matching tokenizer, then follow the repository inference instructions.
Training
The acoustic model was trained through character CTC on LibriSpeech clean-100 and other-500, BPE-CTC on the same 600 hours initialized from that encoder, and a final 30-epoch continuation on clean-100 + clean-360 + other-500. The published checkpoint is epoch 29 of the final stage.
The BPE tokenizer was trained only on train-clean-100 and train-other-500 transcriptions. Development and test text were not used for tokenizer training. There is no optimizer state in the exported checkpoint.
Evaluation
Corpus metrics, percentages. Final decoding uses the external LibriSpeech 4-gram for beam search followed by SpeechBrain Transformer LM rescoring. The language models are not included in this repository.
| Split | Greedy WER | Rescored WER | Rescored CER |
|---|---|---|---|
| dev-clean | 7.45 | 3.98 | 1.68 |
| dev-other | 17.12 | 10.15 | 5.84 |
| test-clean | 7.45 | 4.17 | 1.71 |
| test-other | 17.26 | 10.48 | 5.81 |
The decoding parameters were selected on dev-other with the preceding BPE checkpoint and fixed for this evaluation. See metrics.json for exact results. This model is intended for English read speech; results on other domains have not been established.