---
license: cc-by-4.0
language:
- en
base_model: omi-health/omi-med-stt-v1
pipeline_tag: automatic-speech-recognition
tags:
- coreml
- apple-silicon
- speech-recognition
- parakeet
- medical
- clinical
---

# Omi Med STT v1 — CoreML

CoreML conversion of [omi-health/omi-med-stt-v1](https://huggingface.co/omi-health/omi-med-stt-v1),
an English medical speech-to-text model for clinical dialogue built from
NVIDIA Parakeet TDT 0.6B v2 (FastConformer + Token-and-Duration Transducer).
Runs fully on-device on Apple Silicon (Neural Engine + CPU) via
[FluidAudio](https://github.com/FluidInference/FluidAudio).

The component layout, I/O contract, tokenizer (1024 SentencePiece tokens,
blank id 1024), and file names are identical to
[FluidInference/parakeet-tdt-0.6b-v2-coreml](https://huggingface.co/FluidInference/parakeet-tdt-0.6b-v2-coreml) —
only the weights differ (medical fine-tune). Any runtime that loads the stock
v2 CoreML build can load this one from a local directory.

## Files

| File | Role | I/O |
|---|---|---|
| `Preprocessor.mlmodelc` | waveform → log-mel | `audio_signal`, `audio_length` → `mel`, `mel_length` |
| `Encoder.mlmodelc` | FastConformer encoder (FP16, 15 s window) | `mel`, `mel_length` → `encoder`, `encoder_length` |
| `Decoder.mlmodelc` | RNNT prediction network (U=1, explicit LSTM state) | `targets`, `target_length`, `h_in`, `c_in` → `decoder`, `h_out`, `c_out` |
| `JointDecision.mlmodelc` | fused single-step joint + decision head | `encoder_step`, `decoder_step` → `token_id`, `token_prob`, `duration` |
| `parakeet_vocab.json` | token id → SentencePiece piece (dict format) | — |

Audio: 16 kHz mono, fixed 15-second window (shorter audio is padded by the
runtime). Precision: FP16 MLProgram, minimum deployment target iOS 17 /
macOS 14. Total size ≈ 1.1 GB.

## Conversion provenance

- Source checkpoint: `omimedstt-v1.nemo` from
  [omi-health/omi-med-stt-v1](https://huggingface.co/omi-health/omi-med-stt-v1)
  (`sha256 eaf0ff7258133b4e597aef024f81a4db1779abeb966e53f00a2250bb54e0fbf4`).
- Exported with the [FluidInference mobius](https://github.com/FluidInference/mobius)
  `parakeet-tdt-v2-0.6b/coreml` pipeline (`convert-parakeet.py --nemo-path …`,
  NeMo 2.3.1, coremltools 9.0b1, PyTorch 2.7.0), then compiled with
  `xcrun coremlcompiler` and renamed to the FluidAudio v2 layout. The embedded
  model `author` field reads "Fluid Inference" because it is set by that
  conversion tooling; the conversion of this checkpoint was performed
  independently.
- Vocabulary exported from the checkpoint's SentencePiece tokenizer in
  FluidAudio's dict format.

## Usage (FluidAudio / Swift)

Place all five artifacts in one directory and construct `AsrModels` directly
(the stock download path would fetch generic v2 weights instead):

```swift
import CoreML
import FluidAudio

let dir = URL(fileURLWithPath: "/path/to/omi-med-stt-v1-coreml")
let ane = AsrModels.defaultConfiguration()          // CPU + Neural Engine
let cpu = MLModelConfiguration(); cpu.computeUnits = .cpuOnly

let models = AsrModels(
    encoder: try await MLModel.load(contentsOf: dir.appending(path: "Encoder.mlmodelc"), configuration: ane),
    preprocessor: try await MLModel.load(contentsOf: dir.appending(path: "Preprocessor.mlmodelc"), configuration: cpu),
    decoder: try await MLModel.load(contentsOf: dir.appending(path: "Decoder.mlmodelc"), configuration: ane),
    joint: try await MLModel.load(contentsOf: dir.appending(path: "JointDecision.mlmodelc"), configuration: ane),
    configuration: ane,
    vocabulary: vocabulary,   // parse parakeet_vocab.json: [Int: String]
    version: .v2
)
let manager = AsrManager(config: .default)
try await manager.loadModels(models)
let result = try await manager.transcribe(audioURL: fileURL)
```

In [MacParakeet](https://github.com/moona3k/macparakeet), select the
"Omi Med STT v1" Parakeet build (Settings, or
`macparakeet-cli models select parakeet-omi-med-v1`) — the app downloads this
repository automatically on first use, or pre-fetch with
`macparakeet-cli models download parakeet-omi-med-v1`.

## Upstream model

Per the upstream model card: trained for clinical dialogue (GP consultations,
medication reviews, procedure discussions); 8.30 % WER and 97.95 % medical
recall on a 7.18-hour clinical benchmark. English only. It formats numerics
clinically (e.g. "500 mg", "130 over 85"). See the
[upstream card](https://huggingface.co/omi-health/omi-med-stt-v1) for details
and limitations.

## Disclaimer — use at your own risk

This model conversion is provided **as-is, without warranty of any kind**,
including no warranty of transcription accuracy. Use it at your own risk.

- **Not a medical device.** Transcripts are informational and require human
  review; do not rely on them for diagnosis, treatment, or any clinical
  decision.
- **Patient privacy / HIPAA.** If you process patient speech or any protected
  health information with this model, *you* are solely responsible for
  compliance with HIPAA and any other applicable privacy law or regulation
  (GDPR, state law, or local equivalents). On-device processing helps
  minimize data exposure but does not by itself make a workflow compliant.
- **Institutional policy.** Check with your institution's compliance,
  privacy, and IT-security policies *before* using this model in any
  clinical, research, or educational setting.
- **No liability.** The publisher of this conversion accepts no
  responsibility or liability for privacy, regulatory, or policy violations
  arising from its use.

## License and attribution

Weights are CC-BY-4.0, following the upstream release by
[Omi Health](https://huggingface.co/omi-health) (`omi-med-stt-v1`), which is
built from NVIDIA's `parakeet-tdt-0.6b-v2` (also CC-BY-4.0). This repository
redistributes a format conversion of those weights with attribution and no
additional restrictions.