Downloads · 30 days
133
14% of all-time downloads
jayr23/whisper-large-v3-turbo-swiss-german-ct2
whisper-large-v3-turbo-swiss-german-ct2 is a automatic speech recognition model from jayr23. Use it when you need speech turned into text. It is set up for ctranslate2. The card lists the license as apache-2.0.
This model is a fine-tuned version of openai/whisper-large-v3-turbo for Swiss German automatic speech recognition. It is intended to transcribe Swiss German speech into Standard German text.
Downloads · 30 days
133
14% of all-time downloads
All-time downloads
931
Public
Repo size
1.6 GB
Likes
1
Public
Click a slice to open those files.
.bin1.6 GB · 100%
From the Hugging Face model README
This model is a fine-tuned version of openai/whisper-large-v3-turbo for Swiss German automatic speech recognition. It is intended to transcribe Swiss German speech into Standard German text.
The current upload corresponds to the best checkpoint from the March 20, 2026 retraining run: checkpoint-750.
openai/whisper-large-v3-turbocheckpoint-750301 hours of curated private Swiss German training audiosafetensors24 hours across 8 major Swiss German dialects with Swiss German and High German transcripts293 hours of Swiss German parliamentary speech with Standard German transcripts13 hours with a dialect distribution intended to be close to real-world Swiss GermanThe training run improved steadily up to checkpoint-750, then degraded afterward. The uploaded model is therefore the best checkpoint from the run, not the final checkpoint.
Validation trajectory during the successful run:
| Step | WER | Normalized WER |
|---|---|---|
| 250 | 41.05 | 40.21 |
| 500 | 39.63 | 38.86 |
| 750 | 37.96 | 37.25 |
| 1000 | 42.92 | 42.20 |
| 1250 | 43.64 | 42.92 |
| 1500 | 43.64 | 42.89 |
This is why checkpoint-750 is the shipped model.
The tuned model was compared against the base openai/whisper-large-v3-turbo on a large random sample from the same private training corpus regime used for this retraining run.
Comparison setup:
16,384Results:
| Model | WER | Normalized WER |
|---|---|---|
Base openai/whisper-large-v3-turbo | 45.71 | 44.52 |
| This model | 39.18 | 38.48 |
Absolute improvement over base on that sampled training slice:
-6.54-6.04This model is intended for Swiss German ASR workloads where the target transcription is Standard German text.
It is the right version to try if:
checkpoint-750; later checkpoints were worse.import torch
from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
model_id = "Flurin17/whisper-large-v3-turbo-swiss-german"
device = "cuda:0" if torch.cuda.is_available() else "cpu"
torch_dtype = torch.float16 if torch.cuda.is_available() else torch.float32
model = AutoModelForSpeechSeq2Seq.from_pretrained(
model_id,
torch_dtype=torch_dtype,
low_cpu_mem_usage=True,
use_safetensors=True,
)
model.to(device)
processor = AutoProcessor.from_pretrained(model_id)
pipe = pipeline(
"automatic-speech-recognition",
model=model,
tokenizer=processor.tokenizer,
feature_extractor=processor.feature_extractor,
torch_dtype=torch_dtype,
device=device,
)
result = pipe("path/to/audio.wav")
print(result["text"])
adamw_torch_fused1e-55This model is distributed under Apache 2.0, following the base model license declared for openai/whisper-large-v3-turbo.
Based on how Systran structures their official CT2 model cards and community fine-tune CT2 conversions, here's the adapted card for `jayr23/whisper-large-v3-turbo-swiss-german-ct2`:[^1][^2][^3][^4]
***
```markdown
---
language:
- de
- gsw
license: apache-2.0
pipeline_tag: automatic-speech-recognition
library_name: ctranslate2
tags:
- whisper
- automatic-speech-recognition
- speech-to-text
- swiss-german
- schweizerdeutsch
- ctranslate2
- faster-whisper
base_model:
- Flurin17/whisper-large-v3-turbo-swiss-german
model-index:
- name: whisper-large-v3-turbo-swiss-german-ct2
results:
- task:
type: automatic-speech-recognition
name: Automatic Speech Recognition
dataset:
name: Private Swiss German validation split
type: private-swiss-german-validation
split: validation
metrics:
- type: wer
value: 37.96315900910438
name: Word Error Rate
- type: wer
value: 37.250335238901826
name: Normalized Word Error Rate
---
# Whisper Large V3 Turbo Swiss German — CTranslate2
This is a CTranslate2 conversion of
[Flurin17/whisper-large-v3-turbo-swiss-german](https://huggingface.co/Flurin17/whisper-large-v3-turbo-swiss-german)
for use with [faster-whisper](https://github.com/SYSTRAN/faster-whisper) and
other CTranslate2-based tools (noScribe, Subtitle Edit, faster-whisper-GUI, etc.).
The original model is a fine-tuned version of `openai/whisper-large-v3-turbo` for
Swiss German ASR, transcribing Swiss German speech into Standard German text.
All training details, metrics, and intended use are documented in the
[original model card](https://huggingface.co/Flurin17/whisper-large-v3-turbo-swiss-german).
## Usage
### faster-whisper (Python)
```python
from faster_whisper import WhisperModel
model = WhisperModel("jayr23/whisper-large-v3-turbo-swiss-german-ct2",
device="cuda", compute_type="float16")
segments, info = model.transcribe("audio.wav", language="de")
for segment in segments:
print("[%.2fs -> %.2fs] %s" % (segment.start, segment.end, segment.text))
model = WhisperModel("jayr23/whisper-large-v3-turbo-swiss-german-ct2",
device="cpu", compute_type="int8")
The original Transformers model was converted to CTranslate2 format with:
ct2-transformers-converter \
--model Flurin17/whisper-large-v3-turbo-swiss-german \
--output_dir whisper-large-v3-turbo-swiss-german-ct2 \
--copy_files tokenizer.json preprocessor_config.json \
--quantization float16
Metrics are identical to the original model (conversion is lossless at float16). See the original model card for the full validation trajectory and comparison against base Whisper Turbo.
| Model | WER | Normalized WER |
|---|---|---|
Base openai/whisper-large-v3-turbo | 45.71 | 44.52 |
Flurin17/whisper-large-v3-turbo-swiss-german | 39.18 | 38.48 |
Apache 2.0, following the base model license. See Flurin17/whisper-large-v3-turbo-swiss-german.