omi-med-stt 0.1.18

CLI for Omi Med STT v1 medical speech-to-text

Omi Med STT Runtime

Runtime CLI for Omi Med STT v1, an English medical speech-to-text model.

This repository contains runtime code only. It does not contain model weights, private benchmark data, or training data.

Safety And Intended Use

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.

Runtimes

omi-med-stt supports three runtime paths:

Runtime	Best for	Artifact
mlx	Apple Silicon Macs	omi-health/omi-med-stt-v1-mlx-q8
cpp	Linux and Windows CPU fallback	omi-health/omi-med-stt-v1-gguf / omi-med-stt-v1-q8_0.gguf
nemo	NVIDIA CUDA servers and canonical NeMo checkpoint use	omi-health/omi-med-stt-v1

The source-of-truth model is the NeMo checkpoint. MLX and GGUF are runtime exports.

Default quantization:

• Apple Silicon MLX defaults to the q8 MLX export.
• Linux/Windows CPU defaults to the q8_0 GGUF export through parakeet.cpp.
• NVIDIA CUDA/NeMo defaults to the canonical full NeMo checkpoint. A NeMo q8 checkpoint is not provided because the current NeMo runtime path does not have a validated drop-in q8 .nemo inference artifact.

Install

From PyPI:

pip install -U omi-med-stt

For Apple Silicon / MLX:

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

For CUDA/Linux NeMo:

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

From this repository, for runtime development:

pip install -e ".[dev]"

Models

Platform	Default artifact	Eval (locked private benchmark)
Apple Silicon	omi-med-stt-v1-mlx-q8 (0.94 GB)	WER 8.61% · M-WER 2.75%
NVIDIA CUDA	omi-med-stt-v1 (.nemo, 2.5 GB)	WER 8.30% · M-WER 2.37%
Linux/Windows CPU	omi-med-stt-v1-gguf (q8_0, 0.93 GB)	WER 9.12% · M-WER 3.20%

Full evaluation and methodology: https://github.com/Omi-Health/omi-med-stt-runtime/blob/main/docs/evaluation-summary.md · aggregate summary in docs/evaluation-summary.md.

CUDA Driver / Torch Compatibility

PyPI's default torch wheels currently target CUDA 13 and require NVIDIA driver >= 580. On hosts with an older driver (for example cloud GPU images with driver 570.x / CUDA 12.8), --runtime nemo fails with "The NVIDIA driver on your system is too old". Install a torch build that matches your driver before installing the nemo extra:

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

Check your driver's supported CUDA version with nvidia-smi.

Basic Usage

Simple path:

omi-med-stt audio.wav

Explicit MLX:

omi-med-stt audio.wav --runtime mlx

Explicit NeMo:

omi-med-stt audio.wav --runtime nemo

Explicit parakeet.cpp / GGUF:

omi-med-stt audio.wav --runtime cpp

JSON output:

omi-med-stt audio.wav --json

Dependency/runtime check:

omi-med-stt check

parakeet.cpp / GGUF Runtime

The cpp runtime is powered by parakeet.cpp, a C++/ggml inference engine for NVIDIA Parakeet ASR models.

Omi Med STT v1 includes a post-Conformer medical adapter. Until this adapter extension is upstreamed, omi-med-stt builds parakeet.cpp with the adapter patch included in this repository and caches the resulting libparakeet shared library plus parakeet-cli.

The default cpp path uses the parakeet.cpp C API directly. For long audio it keeps the model loaded once and sends in-memory 16 kHz PCM chunks to libparakeet, avoiding repeated model setup and temporary chunk WAV reads. parakeet-cli remains available as a fallback and for developer debugging.

Normal use:

omi-med-stt audio.wav --runtime cpp

Windows CPU is a first-class target. Install once, then transcribe:

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

On Windows CPU the runtime defaults to a capped thread count min(os.cpu_count(), 8) instead of blindly using the upstream engine default. This is more stable on small 4-vCPU machines and can be overridden:

omi-med-stt audio.wav --runtime cpp --cpp-backend cpu --cpp-threads 4
$Env:OMI_MED_STT_CPP_THREADS = "4"
omi-med-stt audio.wav --runtime cpp --cpp-backend cpu

Pre-install the C++ runtime explicitly:

omi-med-stt install-cpp

Choose a backend:

omi-med-stt install-cpp --cpp-backend cpu

Manual override remains available for developers:

omi-med-stt audio.wav --runtime cpp --parakeet-cli /path/to/parakeet-cli

Runtime toggles:

# force the old subprocess path
OMI_MED_STT_CPP_DISABLE_CAPI=1 omi-med-stt audio.wav --runtime cpp

# require C API and fail instead of falling back
OMI_MED_STT_CPP_REQUIRE_CAPI=1 omi-med-stt audio.wav --runtime cpp

# portable build instead of native CPU flags
OMI_MED_STT_GGML_NATIVE=OFF omi-med-stt install-cpp --force --cpp-backend cpu

For Linux and Windows CPU, install-cpp first downloads a pinned prebuilt runtime bundle from the Omi Med STT runtime GitHub release. If no bundle is available for the host, it falls back to building from parakeet.cpp with parakeet-cpp-omi-adapter.patch. The source-build fallback requires git and cmake. On Windows, normal use requires only the shared parakeet.dll runtime; it does not require a separate parakeet-cli.exe.

The cpp runtime downloads only the selected GGUF file. It does not download the NeMo .nemo checkpoint or the MLX model.safetensors.

By default, the cpp runtime is pinned to a specific Hugging Face repository revision and verifies the SHA256 checksum for the official f16/q8_0 GGUF files. You can override this for experiments:

omi-med-stt audio.wav --runtime cpp --revision main
omi-med-stt audio.wav --runtime cpp --no-verify-checksum

Long Audio

Omi Med STT v1 is based on Parakeet and can handle long-form audio. Start with the simple path:

omi-med-stt consult.wav

An explicit chunked path is still available for constrained environments:

omi-med-stt transcribe-long consult.wav --chunk-seconds 25 --overlap 3

Model Access

The runtime defaults to these Hugging Face model repositories:

• omi-health/omi-med-stt-v1
• omi-health/omi-med-stt-v1-mlx-q8
• omi-health/omi-med-stt-v1-mlx for the full MLX export if explicitly selected
• omi-health/omi-med-stt-v1-gguf

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

huggingface-cli login

Pre-Publish Checks

Use these before pushing runtime changes to GitHub or publishing a new PyPI version. The checked-in unit tests do not download models and do not require audio files:

pip install -e ".[dev]"
python scripts/prepublish_check.py --skip-build

For real model smoke tests, put a small permitted audio file under local_smoke/ or pass any private local path. local_smoke/ and audio/model artifacts are git-ignored.

Apple Silicon Mac:

python scripts/prepublish_check.py --audio local_smoke/sample.wav --runtime mlx --skip-build

Linux with NVIDIA GPU:

python scripts/prepublish_check.py --audio local_smoke/sample.wav --runtime nemo --skip-build
python scripts/prepublish_check.py --audio local_smoke/sample.wav --runtime cpp:cuda --skip-build

Linux CPU:

python scripts/prepublish_check.py --audio local_smoke/sample.wav --runtime cpp:cpu --skip-build

Windows CPU, from PowerShell:

py -m scripts.prepublish_check --audio local_smoke\sample.wav --runtime cpp:cpu --skip-build
py -m scripts.prepublish_check --audio local_smoke\consult.wav --runtime cpp:cpu --long --skip-build

Long-audio smoke, using the runtime chunk path:

python scripts/prepublish_check.py --audio local_smoke/consult.wav --runtime cpp:cpu --long --skip-build

Windows NPU is currently an explicit expected skip. There is no NPU backend for Omi Med STT yet; use cpp CPU on Windows, MLX on Apple Silicon, or NeMo on a CUDA Linux machine.

The script writes the latest local report to local_smoke/prepublish_last.json.

Documentation

Additional public-safe maintainer documentation lives under docs/:

• docs/evaluation-summary.md — aggregate benchmark and runtime results.
• docs/release-log.md — runtime release history and launch readiness.
• docs/runtime-qa.md — test matrix and smoke checks.
• docs/repo-hygiene.md — what stays public vs private.

Agent/maintainer context is in AGENTS.md.

Attribution

This runtime uses or interoperates with:

• NVIDIA NeMo / Parakeet, for the base ASR architecture.
• parakeet-mlx, for Apple Silicon MLX inference.
• parakeet.cpp, for GGUF / C++ / ggml inference.

See NOTICE.md.

License

Runtime code in this repository is MIT licensed.

Model weights are governed separately by the model repositories. Omi Med STT v1 is a derivative of nvidia/parakeet-tdt-0.6b-v2, whose model weights are licensed under CC-BY-4.0.

Omi Med STT v1 — links

Full evaluation & methodology	https://github.com/Omi-Health/omi-med-stt-runtime/blob/main/docs/evaluation-summary.md
Canonical model (NeMo, source of truth)	https://huggingface.co/omi-health/omi-med-stt-v1
PyPI	https://pypi.org/project/omi-med-stt/
Related: SOAP summarization (sum-small)	https://huggingface.co/omi-health/sum-small

Built by Omi Health · Scribe product page · Contact: [e-mail]