omi-health/omi-med-stt-v1

Omi Med STT v1

Omi Med STT v1 is an English medical speech-to-text model for clinical dialogue, built from NVIDIA Parakeet TDT 0.6B v2.

It is designed for local transcription of GP-style consultations, medication review, clinical dictation, and procedure/device/test language. The canonical checkpoint is this NeMo .nemo model. Mac and CPU exports are available in the companion repos below.

Why Use It

• Open-weight, 0.6B-class medical ASR model
• Runs locally through the omi-med-stt CLI
• Stronger medical transcription than base Parakeet v2 in Omi's internal evaluation
• Runtime options for Apple Silicon, NVIDIA CUDA, and Linux/Windows CPU

Quickstart

Recommended CLI:

pip install -U omi-med-stt
omi-med-stt audio.wav

Apple Silicon:

pip install -U "omi-med-stt[mlx]"
omi-med-stt audio.wav

NVIDIA CUDA / canonical NeMo checkpoint:

pip install -U "omi-med-stt[nemo]"
omi-med-stt audio.wav --runtime nemo

Linux or Windows CPU:

pip install -U omi-med-stt
omi-med-stt install-cpp --cpp-backend cpu
omi-med-stt audio.wav --runtime cpp

Direct NeMo use:

from huggingface_hub import hf_hub_download
import nemo.collections.asr as nemo_asr

nemo_path = hf_hub_download("omi-health/omi-med-stt-v1", filename="omimedstt-v1.nemo")
model = nemo_asr.models.ASRModel.restore_from(nemo_path)
print(model.transcribe(["audio_16k_mono.wav"])[0])

Which Repo Should I Use?

Use case	Repo	Notes
Best quality / NVIDIA GPU	omi-health/omi-med-stt-v1	Canonical NeMo checkpoint
Apple Silicon default	omi-health/omi-med-stt-v1-mlx-q8	Smaller MLX q8 export
Apple Silicon full precision	omi-health/omi-med-stt-v1-mlx	Larger MLX export
Linux/Windows CPU	omi-health/omi-med-stt-v1-gguf	GGUF q8_0 via parakeet.cpp

Evaluation

Full evaluation details: omi.health/research/omi-med-stt. Benchmark: 7.18h of real and synthetic clinical speech across dialogue, dictation, medication review, procedures/devices/tests, and general speech. Speed is shown as time to process one hour of audio; lower is faster.

1. Open / Local Models

Lower is better for WER metrics; higher is better for Medical Recall. Local GPU baselines were run on A10 where applicable; VibeVoice-ASR 9B used H100.

Model	WER	M-WER	Drug M-WER	Medical Recall	Speed: time / 1 hour audio (formula-derived x realtime)
VibeVoice-ASR 9B	11.10%	1.78%	1.36%	98.71%	5m 20s (11.2x)
Omi Med STT v1 NeMo	8.30%	2.37%	4.75%	97.95%	25s (146.3x)
Qwen3 ASR 1.7B	10.72%	3.13%	6.11%	97.21%	44s (81.1x)
Whisper Large v3 Turbo (A10)	11.98%	3.93%	5.88%	96.45%	1m 19s (45.8x)
Cohere Transcribe 03-2026	14.88%	5.05%	11.09%	95.16%	25s (146.3x)
Parakeet TDT 0.6B v3	15.26%	8.01%	9.50%	96.34%	23s (157.9x)
Parakeet TDT 0.6B v2 base	16.45%	8.36%	8.60%	96.20%	23s (153.8x)

2. Closed / API Models

Top closed/API comparison plus Omi. Speed note: Omi is local A10 compute with no network round-trip; cloud rows are single-request API timings, so concurrency can improve cloud throughput.

Model	WER	M-WER	Drug M-WER	Medical Recall	Speed: time / 1 hour audio (formula-derived x realtime)
ElevenLabs Scribe v2	6.53%	1.39%	0.23%	99.03%	7m 42s (7.8x)
Gemini 3.1 Pro Preview	7.13%	1.65%	0.23%	98.22%	41m 54s (1.4x)
AssemblyAI Universal-3 Pro Medical	6.94%	1.81%	1.36%	98.57%	28m 6s (2.1x)
Soniox STT Async v4	6.99%	1.95%	3.39%	98.40%	33m 18s (1.8x)
Omi Med STT v1 NeMo	8.30%	2.37%	4.75%	97.95%	25s (146.3x)
Gemini 3.5 Flash	7.99%	2.39%	0.45%	96.66%	19m 18s (3.1x)
Deepgram Nova-3 Medical	7.33%	2.44%	2.26%	97.81%	7m 48s (7.7x)
Reson8 Prerecorded	6.69%	2.58%	6.56%	97.74%	8m 6s (7.4x)
Voxtral Mini Transcribe v2	8.12%	2.79%	5.66%	97.53%	3m 54s (15.4x)
OpenAI GPT-4o Mini Transcribe	10.26%	3.55%	3.39%	96.76%	4m 54s (12.2x)

3. Runtime Artifacts

The canonical NeMo checkpoint is the source-of-truth model. Runtime exports trade a small amount of quality for size or portability.

Artifact	WER	M-WER	Drug M-WER	Medical Recall	Speed: time / 1 hour audio (formula-derived x realtime)
NeMo canonical	8.30%	2.37%	4.75%	97.95%	25s (146.3x)
MLX full precision	8.59%	2.65%	5.20%	97.70%	56s (64.5x)
MLX q8	8.61%	2.75%	5.20%	97.63%	53s (67.4x)
GGUF q8_0	9.12%	3.20%	6.33%	97.53%	2m 53s (20.8x)

Notes

• The model is English-only.
• The CLI converts audio to 16 kHz mono automatically.
• The CPU GGUF path is for portability; use MLX or NeMo when available.

Safety

Omi Med STT v1 is speech-to-text only. It is not a diagnostic, triage, prescribing, or clinical decision model, and it is not clinically validated. Transcripts must be reviewed before any clinical use.

License And Attribution

Weights are released under CC-BY-4.0, matching the base model.

Omi Med STT v1 is a derivative of nvidia/parakeet-tdt-0.6b-v2. It is not an NVIDIA model.

Runtime: Omi-Health/omi-med-stt-runtime under MIT.

Built by Omi Health. Related product: Omi Scribe.