twirly-de-turbo-coreml

whisper-large-v3-turbo, CoreML, for WhisperKit

openai/whisper-large-v3-turbo converted to CoreML with whisperkittools, float16, packaged for WhisperKit. No fine-tune — these are the unmodified OpenAI weights.

Used by Twirly on macOS and iOS for Swiss German dictation: the model hears Swiss German and writes Standard German.

openai_whisper-large-v3-turbo/
  MelSpectrogram.mlmodelc
  AudioEncoder.mlmodelc
  TextDecoder.mlmodelc

The files sit in a <variant>/ folder, not at the root: WhisperKit.download(variant:) looks for *<variant>/*, and a repo with the model only at the root makes every attempt fail with modelsUnavailable.

How it was made

whisperkit-generate-model --model-version openai/whisper-large-v3-turbo --output-dir out

ANE dispatch as reported by the conversion: TextDecoder 203/205 ops.

Measured quality (Swiss German → Standard German)

392 segments of the Swiss Parliaments Corpus (CC BY 4.0), 11 270 reference words: 19.9 % WER for these weights at fp16.

For reference, a CC BY-NC Swiss German fine-tune of the same base measured 22.8 % on the same segments — the unadapted base is better, and sellable.

Checked on a real 57-second Swiss German recording through WhisperKit before shipping: the model writes Standard German throughout. What it still misses is the dialect lexicon — poschte comes out as "gepostet", verlämperet as "verlampet". Those are handled downstream in the app.

Licence

MIT, from the base model. Copyright (c) 2022 OpenAI. Changes were made: converted to CoreML by CHO CHI CHO GMBH, Zürich.