OpenAI Whisper in CoreML format, published by Diction Labs for on-device speech-to-text on Apple hardware. Everything here is our own build, made with whisperkittools (MIT) from OpenAI's MIT-licensed Whisper checkpoints.

Runs on argmax-oss-swift (MIT). Tokenizers are resolved separately from the matching openai/whisper-* repo, so a first run still needs network access even with the weights already on disk.

dictionlabs_whisper-large-v3-turbo-q6q8 is a compressed version of the turbo model above, same weights, reduced precision (encoder 6-bit, decoder 8-bit), 45% of the size. Verified against the full precision build with real transcription tests, not just internal accuracy checks.

Headline (2026-09-04): on 150 clips of real spontaneous speech (VoxPopuli, 113 speakers, en/de/fr) the compressed build scores 11.14% WER vs 10.96% for full precision — a +0.18 pt difference inside sampling noise — and produces the identical output to full precision on 141 of 150 clips. Under added noise down to 0 dB SNR (240 clips) the two are within 0.1 pt. Runs at ~9.6× realtime on an M1 Mac mini via the Neural Engine. Full methodology, noise tables, weight-fidelity figures and caveats: dictionlabs_whisper-large-v3-turbo-q6q8/README.md.

Read-speech check (FLEURS, 5 clips per language, absolute numbers inflated by a simple normalizer):

MIT, same as upstream Whisper.