Custom Local Whisper Models

Why use a custom local model?

• Domain accuracy: Fine‑tuned models (e.g., medical, legal, meetings, accents) can outperform general models on your data.
• Latency and cost: Local inference avoids network latency and API costs, though it can be slightly slower than Core ML optimized models if you're only using the GGML whisper.cpp compatible format.

Requirements

• A local Whisper model that you are trying to add should be whisper.cpp compatible format with the .bin extension (e.g., ggml-large-v3-turbo.bin).
• Optional: For Core ML support, the Core ML model should have exactly the same name as the GGML model, with the format [model-name]-encoder.mlmodelc (e.g., ggml-large-v3-turbo-encoder.mlmodelc).

Available Fine-Tuned Models

Language-Specific Models (Ready to Use):

• Finnish/Swedish (Large-v3)
• Nepali (Large-v3)
• German (Large-v3-Turbo)
• Hebrew (Large-v3)
• Korean (Medium)
• Swedish (Large)
• Hindi (Base)

Models Requiring Conversion (PyTorch format):

• Medical (Large-v3)
• Clinical (Large-v3)
• Russian (Large-v3-Turbo)
• Chinese (Large-v3-Turbo)

Conversion Resources

• GGML whisper.cpp compatible conversion: whisper.cpp GGML format
• Core ML conversion: whisper.cpp Core ML support

For more background, see VoiceInk docs: Custom Local Whisper Models.

How to import in VoiceInk

1. Open VoiceInk → AI Models.
2. Go to the Local tab.
3. Scroll to the bottom and click Import Local Model….
4. Select your .bin file. You will be able to use it now with VoiceInk.
5. Click Set as Default on the imported model to use it for transcription.

Deleting and managing models

• From the model card menu, choose Delete Model to remove the file from VoiceInk. The card disappears immediately.
• Show in Finder reveals the actual .bin file in the models directory.