加载中…
加载中…
WhisperKit-compatible Core ML conversion of
gcoli/whisper-large-v3-turbo-swiss-german-mit.
Swiss German speech to Standard German text, on device on Apple Silicon. Licence: MIT, as the
source model.
About 3× faster than the large variant, at about 2 WER points less accuracy. On our sample it does not beat untuned large-v3; use it where speed matters more than accuracy.
Use language de and task transcribe:
whisperkit-cli transcribe --model-path <this-folder> --audio-path input.wav --language de
The tokenizer ships with the bundle; point WhisperKit's tokenizer folder at it to run offline.
Text decoder compute units: the Neural Engine (default) and the GPU (cpuAndGPU) gave the same
WER. The Neural Engine is faster; the GPU is numerically closer to PyTorch (see Conversion).
First load: on first use per device, macOS specialises the audio encoder for the Neural Engine. This took 15 to 19 minutes on an M1 Max; later loads take about 1 s.
| Language | Status |
|---|---|
| Swiss German read from a Standard German prompt | measured on this bundle (i4ds test set, below) |
| Spontaneous Swiss German | not verified; likely weaker than the measured value |
| Standard German with a Swiss accent | not measured; expected to work |
| Standard German, English | works |
| French, Italian | presumably works; not checked on this bundle |
282 clips of the FHNW/i4ds All Swiss German Dialects Test Set (MIT), stratified over 17 dialects.
Speakers read Standard German sentences aloud in their dialect, so this is read, prompt-close
Swiss German, not spontaneous speech.
whisperkit-cli 1.1.1 on an M1 Max, greedy decoding, normalisation as in the source model card.
| Model (Core ML) | WER | CER | Time for 40 min audio |
|---|---|---|---|
| this bundle, decoder on Neural Engine | 24.7 % | 12.3 % | 232 s |
| this bundle, decoder on GPU | 24.8 % | 12.2 % | – |
| large variant | 22.8 % | 11.9 % | 754 s |
openai_whisper-large-v3 (no fine-tune) |
23.6 % | 12.0 % | – |
The source checkpoint reaches 23.7 % (dev) / 25.9 % (final) with CTranslate2 and beam 5.
gcoli/whisper-large-v3-turbo-swiss-german-mit @ 540fb2e7a9135048963c5de9822579c96f8962e5argmaxinc/whisperkittools @ 84f77a83c8f530022ae55fbb1a64b3351ef63c7a, coremltools 9.0whisperkit-generate-model.conversion_options.json.Whisper can hallucinate or omit text, especially with noise, silence, rare dialects, overlapping speakers or domain vocabulary. Do not use the output as the sole basis for high-impact decisions. Obtain consent before transcribing people.
MIT, as the source model. Upstream: openai/whisper-large-v3-turbo (MIT), FHNW/i4ds data (MIT),
WhisperKit and whisperkittools (MIT).