---

library_name: transcribe.cpp

pipeline_tag: automatic-speech-recognition

license: other

license_name: swissdial-cc-by-nc-4.0-no-reidentification

license_link: https://form.jotform.com/223344961502048

language:

- gsw

- de

tags:

- gguf

- transcribe.cpp

- whisper

- q4_0

- swiss-german

- automatic-speech-recognition

base_model:

- openai/whisper-large-v3

- Flurin17/whisper-large-v3-peft-swiss-german

---

Whisper Large-v3 Swiss German – GGUF Q4_0

Q4_0-quantized GGUF 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

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

build/bin/transcribe-cli \
  -m whisper-large-v3-swiss-german-Q4_0.gguf \
  --language de \
  input.wav

For Swiss German transcription, de is the recommended Whisper language token

and transcribe is the intended task.

Format and compatibility

• Format: GGUF Q4_0
• Runtime: handy-computer/transcribe.cpp
• Source quantization: the validated F16 GGUF variant
• Not compatible with oMLX, which expects MLX Safetensors

Use

gcoli/whisper-large-v3-swiss-german-mlx-fp16

for oMLX.

Provenance

• Base revision: 1ecca609f9a5ae2cd97a576a9725bc714c022a93
• Adapter revision: 2ae117cf342bc57d6068066181a9d359e98a2961
• transcribe.cpp converter and quantizer revision:

856d7c10a1a864b900e066b7c9801edf373f5148

• Quantization preset: Q4_0

The adapter was merged into the base model in FP16. The resulting F16 GGUF was

then quantized using transcribe-quantize. Tokenizer, frontend and model

metadata remain embedded in the GGUF. quantization_provenance.json records

the exact F16 source checksum.

Evaluation and limitations

The build performs checksum validation and a real runtime transcription test.

Independent Swiss German WER and controlled performance comparisons between

F16, Q8_0 and Q4_0 have not yet been published. Q4_0 trades a smaller model and

lower memory demand for a potentially larger accuracy loss than Q8_0.

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.

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 is limited to non-commercial purposes under CC BY-NC 4.0, attempts to

re-identify SwissDial speakers are prohibited, and SwissDial must be attributed

and cited in research use.

The Whisper base model remains subject to Apache-2.0. This repository grants no

additional rights to upstream models, training data, or software.