omi-medical-edge-1: Open On-Device Medical Speech-to-Text

omi-medical-edge-1 is a 0.6-billion-parameter adaptation of NVIDIA Parakeet TDT 0.6B v2 for local clinical transcription. The weights did not change in this release. The runtime did. CUDA now gives the best WER and throughput, Apple MLX q8 gives the best canonical Medical WER, and CPU remains the portable fallback. Every shipped artifact has its own result on the same sealed 1,513-clip benchmark.

Current runtime results

Audio preparation, attention, decoding, batching and long-file handling are now qualified as three reproducible paths. Among the open models tested on the standing board, CUDA and MLX have the lowest observed WER. MLX q8 has the second-lowest observed M-WER. These are positions in this benchmark draw.

One install, three runtimes

The model ships with an open runtime, omi-med-stt, that auto-selects the right engine and artifact for the machine it's on:

# Apple Silicon → MLX q8
$ pip install -U "omi-med-stt[mlx]"

# NVIDIA CUDA → NeMo full checkpoint
$ pip install -U "omi-med-stt[nemo]"

# Linux / Windows CPU → GGUF q8_0 via parakeet.cpp
$ pip install -U omi-med-stt
$ omi-med-stt install-cpp --cpp-backend cpu

# Then transcribe; the runtime is selected automatically
$ omi-med-stt consultation.wav

Most releases quantize silently. We benchmarked every artifact we ship on the full locked test set, with the same scorer:

Accuracy uses the same frozen board and scorer for every artifact. Speed is hardware-specific. No dictionary, contextual bias or transcript correction is used. CPU recorded the lowest drug-error count in this draw, but the difference is not statistically established. The July 2026 benchmark row at 6.64% WER and 2.16% M-WER came from a separate GPU-served endpoint and remains on the benchmark page for continuity.

MLX q8 remains the Mac default because larger alternatives did not improve the clinical measures enough to justify their memory and speed cost. A 4-bit artifact regressed on drug names, so it was not released. The matrix is tested on macOS, clean Ubuntu and Windows installations, and Azure NVIDIA hardware.

How it was measured

OmiMedSTT-Bench contains 1,513 clips and 7.2 hours of English medical audio across five scenarios. Every system is scored by the same open pipeline. Medical WER measures errors on clinically relevant terms. The test set was locked before final model selection, with zero file-level overlap with training audio. The benchmark page contains the dataset card, scorer and complete 30-system board.

Training data and methodology

omi-medical-edge-1 is fine-tuned from NVIDIA Parakeet TDT 0.6B v2 on English clinical audio. The corpus combines openly licensed datasets, access-controlled clinical sources and Omi synthetic medical speech. Most of the audio is real recordings.

The medical material includes consultation dialogue, dictation, medication review, procedures, devices and tests. General speech remains in the mix to protect ordinary conversational accuracy. Scorer definitions, competitor versions and decode settings are published with the runtime.

The archived PriMock57 benchmark is unsuitable for evaluating this model because part of that public dataset appears in training. The standing benchmark therefore uses a separate sealed test set.

Limitations

• Drug names remain the hardest category on CUDA: 4.75% Drug M-WER, or 21 errors across 442 occurrences.
• Conversation audio is harder than dictation: the historical served evaluation recorded 10.3% WER on consultations and roughly 4% on dictation.
• The current model is English only. Audio is converted to 16 kHz mono automatically.
• Quantization trades some accuracy for size. Each shipped artifact has its own result above.

omi-medical-edge-1 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.

Run and inspect the model

• Canonical model (NeMo): huggingface.co/omi-health/omi-med-stt-v1
• Apple Silicon MLX q8 (Mac default): omi-med-stt-v1-mlx-q8 · MLX full: omi-med-stt-v1-mlx
• CPU GGUF (parakeet.cpp): omi-med-stt-v1-gguf
• Runtime CLI (MIT): github.com/Omi-Health/omi-med-stt-runtime · PyPI
• Pairs with: Omi-Sum 3B for a local transcript-to-SOAP-note pipeline.
• Contact: [email protected]

omi-medical-edge-1 is a derivative of nvidia/parakeet-tdt-0.6b-v2. It is not an NVIDIA model. Built with NVIDIA NeMo; the runtime interoperates with parakeet-mlx and parakeet.cpp.

Cite this model

APA: Omi Health. (2026). omi-medical-edge-1: On-Device Medical Speech-to-Text. https://omi.health/research/omi-med-stt

@misc{omi_med_stt_v1_2026,
  title   = {omi-medical-edge-1: On-Device Medical Speech-to-Text},
  author  = {{Omi Health}},
  year    = {2026},
  url     = {https://omi.health/research/omi-med-stt},
  note    = {0.6B medical ASR, fine-tuned from NVIDIA Parakeet TDT 0.6B v2, CC-BY-4.0}
}

Related research

• Medical Speech-to-Text Benchmark: 30 systems on a sealed clinical test set
• Clinical SOAP Note Safety Evaluation: 6 models and 300 dialogues
• Omi-Sum 3B: open clinical SOAP-note model