robertteleng/whisper-small-coreml

Whisper small · CoreML encoder for the Apple Neural Engine

OpenAI Whisper small, split for on-device use in an iOS app: the encoder converted to CoreML (.mlpackage) so it runs on the Apple Neural Engine, and the decoder weights kept as safetensors to be run separately.

Converted with robertteleng/coreml-forge:

uv run python scripts/export_whisper.py

What this is, and what it is not

A format conversion, not a new model: same weights as openai/whisper-small, same transcription quality, same languages. Nothing was retrained.

The work is in the conversion: tracing the encoder so CoreML accepts it and the Neural Engine actually runs it, rather than falling back to CPU.

Files

File	What it is
WhisperEncoder_small.mlpackage	Encoder, CoreML, for the Neural Engine
decoder_weights.safetensors	Decoder weights, to run outside CoreML

Limitations

• Encoder only. This is not a drop-in transcriber: you need to run the decoder yourself.
• No on-device benchmarks published here — no latency or memory numbers on real hardware.
• The conversion targets a recent iOS version; older deployment targets may need re-exporting.
• Inherits Whisper's known behaviour, including hallucinated text on silence and uneven quality across languages.

Licence and attribution

Apache-2.0, following openai/whisper-small. The model is OpenAI's; this repository only redistributes a converted format.