# Whisper Large V3 Turbo · Core ML (Option A1)

Optimized Core ML export of `openai/whisper-large-v3-turbo` targeting Apple Silicon. This repository distributes the production bundle used to reach ~0.024 real-time factor (RTF) described in the Option A1 implementation plan.

## Contents

- `whisper-large-v3-turbo-coreml-fp16/` – unpacked model bundle with encoder/decoder mlpackages, tokenizer assets, and metadata.
- `whisper-large-v3-turbo-coreml-fp16.tar.gz` – identical bundle packaged as a tarball for one-shot download.
- `whisper-large-v3-turbo-coreml-fp16.sha256` – checksum for the tarball.

Inside the directory you will find:

- `Encoder.mlpackage`
- `DecoderWithCache.mlpackage` (primary, tensor-cache decoder)
- `DecoderFull.mlpackage` (fallback)
- `DecoderStateful.mlpackage` (experimental; see note below)
- Tokenizer + mel filter assets and metadata JSON files
- `README.md` and `MODEL_CARD.md` with usage and performance details

## Usage

1. Download the snapshot via `huggingface-cli` or `snapshot_download`:

   ```python
   from huggingface_hub import snapshot_download
   model_path = snapshot_download("DRTR-J/whisper-large-v3-turbo-coreml-fp16")
   ```

2. Point your Core ML integration (Swift, Rust, Tauri, etc.) at `DecoderWithCache.mlpackage` to achieve the advertised RTF ≈ 0.024 on Apple Silicon.

3. The experimental `DecoderStateful.mlpackage` is included for completeness but still fails ANE planning (Core ML error -14); keep it disabled in production.

For full background and performance tables, consult `whisper-large-v3-turbo-coreml-fp16/MODEL_CARD.md`.

## License

- Base weights: OpenAI Whisper license
- Packaging and auxiliary assets: MIT