parakeet-tdt-0.6b-v3-basque

Parakeet TDT 0.6B v3 — Basque (Euskara)

Basque ASR model obtained by fine-tuning nvidia/parakeet-tdt-0.6b-v3 on asierhv/composite_corpus_eu_v2.1.

What this model is

• Base architecture: FastConformer RNNT-TDT (Parakeet TDT 0.6B v3)
• Language: Basque (eu)
• Target task: Automatic speech recognition (16 kHz speech)
• Framework: NVIDIA NeMo
• Artifact format: .nemo

Training data provenance

Fine-tuning used the composite Basque corpus:

• asierhv/composite_corpus_eu_v2.1

This composite corpus includes Basque speech/transcript data from:

• Mozilla Common Voice (Basque)
• Basque Parliament
• OpenSLR (Basque)

Reference datasets:

• Mozilla Common Voice: https://commonvoice.mozilla.org/
• Basque Parliament corpus (as included in the composite dataset above)
• OpenSLR Basque resources: https://www.openslr.org/

Evaluation summary

WER was computed on held-out test splits created from the same composite source family.

Notes:

• Baseline is the original multilingual/English-oriented base model, which performs poorly on Basque (WER > 100% is expected).
• Fine-tuned metrics above are from the final selected checkpoint/model export.

Quick start

1) Load from local .nemo

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.ASRModel.restore_from("parakeet-tdt-0.6b-v3-basque.nemo")
asr_model = asr_model.cuda().eval()  # optional if GPU is available

pred = asr_model.transcribe(
    audio=["/path/to/audio.wav"],
    use_lhotse=False,
    batch_size=1,
    num_workers=0,
    verbose=False,
)

# pred may be list[str] or list[Hypothesis], depending on settings/version
if len(pred) and hasattr(pred[0], "text"):
    print(pred[0].text)
else:
    print(pred[0])

2) Load from Hugging Face repo

import nemo.collections.asr as nemo_asr

asr_model = nemo_asr.models.ASRModel.from_pretrained(
    model_name="xezpeleta/parakeet-tdt-0.6b-v3-basque"
)

out = asr_model.transcribe(
    audio=["/path/to/audio.wav"],
    use_lhotse=False,
    batch_size=1,
    num_workers=0,
    verbose=False,
)

print(out[0].text if hasattr(out[0], "text") else out[0])

3) Batch transcription snippet

audio_paths = [
    "/path/a.wav",
    "/path/b.wav",
    "/path/c.wav",
]

outs = asr_model.transcribe(
    audio=audio_paths,
    use_lhotse=False,
    batch_size=16,
    num_workers=4,
    verbose=True,
)

texts = [o.text if hasattr(o, "text") else o for o in outs]
for path, txt in zip(audio_paths, texts):
    print(path, "=>", txt)

Recommended runtime settings

• Use 16 kHz mono WAV when possible.
• For NeMo 2.7 environments, use use_lhotse=False in transcribe(...) if you observe inference/runtime issues in constrained container setups.
• Increase batch_size gradually to match available GPU memory.

Fine-tuning recipe (high-level)

• Full-model fine-tuning from nvidia/parakeet-tdt-0.6b-v3
• Optimizer: AdamW
• LR: 1e-4
• Scheduler: CosineAnnealing with warmup
• Effective batch size: 64 (batch_size=8, gradient accumulation=8)
• Precision: BF16 mixed
• Hardware used: NVIDIA L40 (48 GB)

Limitations

• Performance can degrade on strong code-switching, noisy far-field audio, or domains far from training data.
• This model is optimized for Basque ASR; behavior on other languages is not the target.
• Proper text normalization and punctuation/casing post-processing may still be needed for production use.

Citation and acknowledgements

If you use this model, please cite/credit:

1. Base model: nvidia/parakeet-tdt-0.6b-v3
2. Training dataset: asierhv/composite_corpus_eu_v2.1
3. Underlying source collections: Mozilla Common Voice, Basque Parliament corpus, OpenSLR Basque resources

License

This derivative model follows the base model and dataset terms. Keep attribution and license obligations from all upstream assets.