parakeet-tdt-0.6b-librispeech100-ft
Fine-tuned version of nvidia/parakeet-tdt-0.6b-v3 on LibriSpeech train-clean-100, using NVIDIA NeMo. Fine-tuned with a warm start from the base model, a low learning rate (1e-4), and learning-rate warmup.
Results
Evaluated on LibriSpeech test-clean (2620 utterances):
| Metric | Value |
|---|---|
| WER (raw) | 2.03% |
| WER (normalized) | 2.00% |
| Word accuracy | 98.00% |
Raw and normalized WER are nearly identical, indicating the model transcribes accurately including punctuation and casing (no formatting inflation).
Model description
- Base model: nvidia/parakeet-tdt-0.6b-v3 (FastConformer encoder + TDT decoder, ~600M params)
- Architecture: EncDecRNNTBPEModel (Token-and-Duration Transducer)
- Language: English
- Framework: NVIDIA NeMo
How to use
import nemo.collections.asr as nemo_asr
model = nemo_asr.models.ASRModel.from_pretrained("bhaveshsoni0023/parakeet-tdt-0.6b-librispeech100-ft")
print(model.transcribe(["audio.wav"])[0].text) # 16 kHz mono
Training details
- Data: LibriSpeech train-clean-100 (~100 hours), dev-clean for validation
- Base tokenizer kept (no tokenizer change)
- Optimizer: AdamW, lr 1e-4, warmup 2000 steps
- Precision: bf16, batch size 32, 20 epochs
- Audio: 16 kHz mono
License and attribution
Released under CC-BY-4.0, inherited from the base model and LibriSpeech. Base model: nvidia/parakeet-tdt-0.6b-v3 (NVIDIA), CC-BY-4.0. Attribution required. Data: LibriSpeech (Panayotov et al., 2015), CC-BY-4.0. Framework: NVIDIA NeMo — https://github.com/NVIDIA-NeMo/Speech