Omi Med STT Runtime

Command-line runtime for Omi Med STT v1, an English medical speech-to-text model built from NVIDIA Parakeet TDT 0.6B v2.

The package downloads the right model artifact for your machine and transcribes audio locally.

Install

pip install -U omi-med-stt

Apple Silicon:

pip install -U "omi-med-stt[mlx]"

NVIDIA CUDA / NeMo:

pip install -U "omi-med-stt[nemo]"

The NVIDIA adapter applies the qualified GPU recipe automatically: NeMo 3.0, BF16, local [256,256] attention, greedy-batch TDT decoding with max_symbols=10, timestamps disabled, and duration-sorted batches capped at eight files or 900 audio-seconds. Inputs are normalized through FFmpeg to mono 16 kHz PCM16. A BF16-capable NVIDIA GPU is required; no inference flags are needed beyond --runtime nemo.

Run

omi-med-stt audio.wav

Useful options:

omi-med-stt audio.wav --json
omi-med-stt audio.wav --runtime mlx
omi-med-stt audio.wav --runtime nemo
omi-med-stt audio.wav --runtime cpp
omi-med-stt check

Audio formats. WAV, FLAC, OGG and other libsndfile formats are read directly. Other inputs — .m4a (iPhone Voice Memos / QuickTime), .mp3, .aac, .mp4, .mov, .wma, .opus, .webm, … — are decoded with ffmpeg, which ships with the package, so there's nothing extra to install. If a system ffmpeg is on your PATH it's used instead (e.g. a newer build). Whatever the input, audio is downmixed to mono and resampled to 16 kHz automatically.

Runtime Choices

The canonical model is the NeMo checkpoint. MLX and GGUF are runtime exports.

CPU setup:

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

The CPU path uses a patched parakeet.cpp runtime and downloads the q8_0 GGUF artifact only. It does not download the NeMo or MLX weights.

Runtime Quality

These numbers compare the unchanged runtime artifacts against each other on the same internal benchmark using the runtime recipes shipped in this package. Visit omi.health for the broader model evaluation and product context.

Runtime recipes and checks:

• NVIDIA GPU quality recipe
• Apple MLX q8 quality recipe

The recipe pages include the exact runtime settings, verification commands, and paths to the tests that enforce them.

Model Repositories

• Canonical NeMo: omi-health/omi-med-stt-v1
• Apple Silicon q8: omi-health/omi-med-stt-v1-mlx-q8
• Apple Silicon full precision: omi-health/omi-med-stt-v1-mlx
• Linux/Windows CPU GGUF: omi-health/omi-med-stt-v1-gguf

If the model repositories are private before launch, authenticate first:

huggingface-cli login

CUDA Note

If --runtime nemo fails with a CUDA driver mismatch, install a PyTorch wheel matching your driver before installing the NeMo extra. For example, on CUDA 12.8 hosts:

pip install torch --index-url https://download.pytorch.org/whl/cu128
pip install -U "omi-med-stt[nemo]"

Development

git clone https://github.com/Omi-Health/omi-med-stt-runtime
cd omi-med-stt-runtime
pip install -e ".[dev]"
python scripts/prepublish_check.py --skip-build
python -m pytest -q tests

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

Runtime code is MIT licensed.

Model weights are CC-BY-4.0 and are derived from nvidia/parakeet-tdt-0.6b-v2. Omi Med STT v1 is not an NVIDIA model.

The CPU runtime uses parakeet.cpp.