aarush67/whisper-coreml-models

Whisper.cpp Core ML Models for Apple Silicon

This repository provides prebuilt Core ML models for Whisper.cpp, optimized for Apple Silicon (M1, M2, M3, M4, M5) devices.
These models enable hardware-accelerated speech-to-text using Apple’s Neural Engine via Core ML.

The repository is designed for easy plug-and-play usage with Whisper.cpp using a prebuilt CLI binary.

---

📦 Repository Contents

Each model directory contains everything required to run Whisper.cpp with Core ML acceleration:

• ggml-*.bin – Whisper model weights used by whisper.cpp
• *-encoder.mlmodelc/ – Compiled Core ML encoder bundle

⚠️ Important: The .mlmodelc directories must remain intact. Do not modify, rename, or move their contents.

---

📁 Folder Structure

.
├── tiny/
├── tiny.en/
├── base/
├── base.en/
├── small/
├── small.en/
├── medium/
├── medium.en/
├── large-v1/
├── large-v2/
└── large-v3/

Each folder corresponds to a Whisper model variant and contains the matching .bin and .mlmodelc files.

---

🚀 Model Variants Overview

The table below summarizes the trade-offs between speed, accuracy, and memory usage.

Folder	Model Size	Speed	Accuracy	Notes
tiny	Very Small	⚡ Fastest	⭐ Lowest	Best for real-time, low-resource use
tiny.en	Very Small	⚡ Fastest	⭐ Lowest	English-only
base	Small	⚡ Fast	⭐⭐ Balanced	Good default choice
base.en	Small	⚡ Fast	⭐⭐ Balanced	English-only
small	Medium	⚡ Medium	⭐⭐⭐ Better	Improved transcription quality
small.en	Medium	⚡ Medium	⭐⭐⭐ Better	English-only
medium	Large	🐢 Slower	⭐⭐⭐⭐ High	High accuracy, higher memory usage
medium.en	Large	🐢 Slower	⭐⭐⭐⭐ High	English-only
large-v1	Very Large	🐢 Slow	⭐⭐⭐⭐⭐ Best	Maximum accuracy
large-v2	Very Large	🐢 Slow	⭐⭐⭐⭐⭐ Best	Improved multilingual performance
large-v3	Very Large	🐢 Slow	⭐⭐⭐⭐⭐ Best	Latest and most accurate

---

🛠 Usage Instructions

1. Download the Prebuilt whisper-cli Binary

Download the Core ML–enabled whisper-cli binary directly from GitHub Releases:

https://github.com/aarush67/whisper-cli-for-core-ml/releases/download/v1.0.0/whisper-cli

Recommended directory structure:

.
├── bin/
│   └── whisper-cli
├── medium.en/
│   ├── ggml-medium.en.bin
│   └── medium.en-encoder.mlmodelc/

---

2. Make the Binary Executable

chmod +x bin/whisper-cli

---

3. Run Whisper with a Core ML Model

Place the .bin file and the matching .mlmodelc folder in the same directory.

Example (English-only transcription)

./bin/whisper-cli -m ggml-medium.en.bin -f sample.wav

Whisper.cpp will automatically detect and use the Core ML encoder when available.

---

🧠 Best Practices & Notes

• ✅ Keep .bin and .mlmodelc files together from the same model variant
• ❌ Do not rename, edit, or partially copy .mlmodelc directories
• 🧹 .pt cache files (if generated) are temporary and safe to delete
• 💾 Larger models require significantly more RAM and disk space
• ⚡ Best performance is achieved on Apple Silicon devices with a Neural Engine

---

📜 License

• Repository structure & metadata: MIT License
• Model weights & Core ML artifacts: Governed by the original
  Whisper.cpp / OpenAI Whisper licenses

Please review upstream licenses before commercial or large-scale use.

---

🙏 Credits

• OpenAI Whisper – Original speech recognition models
• whisper.cpp by Georgi Gerganov – High-performance C/C++ implementation
• Apple Core ML – Hardware acceleration on Apple Silicon

---

⭐ Support

If this repository is useful to you:

• ⭐ Star the repository
• 🔗 Support the upstream Whisper.cpp project