gcoli/whisper-large-v3-swiss-german-gguf-f16

Whisper Large-v3 Swiss German – GGUF F16

GGUF F16 conversion of openai/whisper-large-v3 with the LoRA adapter Flurin17/whisper-large-v3-peft-swiss-german merged into the weights. This is a standalone checkpoint for `handy-computer/transcribe.cpp`.

Usage

Build transcribe.cpp, convert the input to 16 kHz mono WAV, and run:

ffmpeg -i input.m4a -ar 16000 -ac 1 input.wav

build/bin/transcribe-cli \
  -m whisper-large-v3-swiss-german-F16.gguf \
  --language de \
  input.wav

The model is multilingual. For Swiss German transcription, de is the recommended Whisper language token and transcribe is the intended task.

Compatibility

• —Format: GGUF, mostly F16 with required small tensors stored in F32
• —Runtime: handy-computer/transcribe.cpp
• —Not compatible with oMLX, which expects MLX Safetensors rather than Whisper GGUF. Use `gcoli/whisper-large-v3-swiss-german-mlx-fp16` for oMLX.

Provenance

• —Base revision: 1ecca609f9a5ae2cd97a576a9725bc714c022a93
• —Adapter revision: 2ae117cf342bc57d6068066181a9d359e98a2961
• —transcribe.cpp converter revision: 856d7c10a1a864b900e066b7c9801edf373f5148

Conversion procedure:

1. 1.load the base checkpoint as FP16;
2. 2.merge the PEFT adapter with merge_and_unload(safe_merge=True);
3. 3.save one standalone FP16 Safetensors checkpoint;
4. 4.convert it with scripts/convert-whisper.py from the pinned transcribe.cpp revision;
5. 5.verify the GGUF container, provenance, SHA-256 checksums, and runtime model loading through transcribe-cli.

The GGUF embeds the tokenizer, special tokens, language metadata, Whisper frontend parameters, Mel filterbank, Hann window, and model tensors. No weight quantization below F16 is applied.

Evaluation

The CoreML conversion of the same merged checkpoint was verified against its PyTorch source during conversion. This GGUF build performs a runtime loading and transcription smoke test. Independent Swiss German WER, dialect coverage, and controlled performance benchmarks for this GGUF conversion have not yet been published.

Limitations

Whisper can hallucinate or omit text, particularly with noise, silence, overlapping speakers, uncommon dialects, or specialized vocabulary. Do not use its output as the sole basis for high-impact decisions. Obtain consent before transcribing people.

License and usage conditions

SwissDial CC BY-NC 4.0 with no-reidentification condition

The Swiss German adaptation was trained on the SwissDial dataset. Use of this model is subject to the following inherited conditions:

• —non-commercial use under Creative Commons Attribution-NonCommercial 4.0;
• —no attempt to determine the identity of any SwissDial speaker;
• —attribution of SwissDial and citation of its publication in research use.

This checkpoint is a modified and converted derivative. The Whisper base model remains subject to Apache-2.0. This repository grants no additional rights to upstream models, training data, or software and does not imply endorsement by ETH Zurich, OpenAI, the adapter authors, or the transcribe.cpp authors.