A fine-tune of nvidia/parakeet-tdt-0.6b-v3
on the English subset of MultiMedmixed with Common Voice 17 English (train + validation). The mix is the
trick: it pushes the model toward medical vocabulary (TAVI, intervertebral
disc herniation, drug names, dosing instructions) while keeping the everyday
English it already knew.
Outputs cased English text with punctuation. Drop-in for the base Parakeet:
same NeMo API, same long-form support, same timestamps.
🔥 Quick start
pip install nemo_toolkit[asr]
python
1import nemo.collections.asr as nemo_asr
23asr = nemo_asr.models.ASRModel.from_pretrained("yuriyvnv/parakeet-tdt-0.6b-EN-Medical")4print(asr.transcribe(["audio.wav"])[0].text)
📊 Results
One model, one training mix (MultiMed-en train + Common Voice 17-en train +
validation, concatenated and shuffled per epoch — same .nemo for every row
below). The two rows are the same checkpoint evaluated on two different
held-out test sets: one in-domain (medical) and one out-of-domain (general
English). Neither test set was seen during training.
The zero-shot column is the unmodified nvidia/parakeet-tdt-0.6b-v3,
measured on the same test set with the same evaluator. All numbers are normalized (lowercase + strip punctuation), the standard protocol used by the MultiMed paper and the Open ASR Leaderboard, so they are directly comparable to other published results.
Test set (held out)
Samples
Zero-shot WER
Fine-tuned WER¹
Δ WER
Zero-shot CER
Fine-tuned CER¹
🩺 MultiMed English (test)
7,567
19.22
14.31
-4.91 (-25.5%)
14.16
10.37
🗣️ Common Voice 17 EN (test)
16,393
8.58
9.11
+0.53 (+6.2%)
4.01
4.10
For reference, the MultiMed paper's best published result is Whisper-Small multilingual fine-tune at 16.62% WER (arXiv 2409.14074, Table 6). This model beats that by 2.31 absolute / 14% relative, at 0.6B params.
¹ Both fine-tuned numbers are this single model — the one trained on the
MultiMed-en + CV17-en mix.
Two things worth noting:
🎯 Medical: 25% relative WER reduction on MultiMed-en test
(19.22% → 14.31%), beating the published MultiMed paper SOTA at 0.6B params.
🛡️ General English is essentially preserved. On CV17-en test the
same model lands at 9.11% WER vs the base Parakeet's
8.58% — a +0.53 absolute / +6.2% relative
change. That's the total cost of medical specialisation. A previous medical-only version of this
model (trained on MultiMed alone, no CV) scored 13.53% normalized WER on CV17-en — a
real catastrophic-forgetting hit. Including CV17-en in the training mix
is what made the trade-off this cheap.
MultiMed-en eval (~2,816 clips) — drives early stopping
Optimiser
AdamW (β=0.9/0.98, weight decay 1e-3)
Learning rate
5e-5 (cosine annealing, min 1e-6)
Warmup
10% of total steps
Batch size
32
Precision
bf16-mixed
Gradient clipping
1.0
Max clip duration
30 s
Early stopping
val_wer plateau, patience 10
Best epoch
31 (val WER 22.84%)
Hardware
Single NVIDIA H100
The natural CV17:MultiMed ratio (~97.5% : ~2.5%) means each batch of 32 contains
on average ~0.8 medical clips and the rest general English — strong CV gradient
keeps the base distribution intact while medical clips slowly steer the model
into the clinical domain.
🎯 Intended use
Transcription of English medical speech: clinical consultations, surgical
procedures, patient narratives, medical podcasts, healthcare interviews.
A drop-in replacement for nvidia/parakeet-tdt-0.6b-v3 in pipelines
targeting medical English — picks up clinical vocabulary without breaking
general transcription.
📚 Training data
leduckhai/MultiMed
English subset — multilingual medical ASR dataset (~84h English).
fixie-ai/common_voice_17_0
English train + validation splits — Common Voice 17, crowdsourced English speech.
