metadata
license: mit
language:
- de
metrics:
- wer
pipeline_tag: automatic-speech-recognition
Model Description
Fine-tuned Whisper-tiny on SwissDial-ZH dataset for Swiss German dialects.
Model Details
- Model Name: nizarmichaud/whisper-tiny-swiss-german
- Base Model: Whisper-tiny-v3
- Dataset: SwissDial-ZH (8 Swiss German dialects): https://mtc.ethz.ch/publications/open-source/swiss-dial.html
- Languages: Swiss German
Training
- Duration: 4 hours
- Hardware: NVIDIA RTX 3080
- Batch Size: 32
- Train/Test Split: 90%/10% (specific sentence selection)
Performance
- WER: ~37% on test set
Usage
from transformers import WhisperForConditionalGeneration, WhisperProcessor
model_name = "nizarmichaud/whisper-tiny-swiss-german"
model = WhisperForConditionalGeneration.from_pretrained(model_name)
processor = WhisperProcessor.from_pretrained(model_name)
audio_input = ... # Your audio input here
inputs = processor(audio_input, return_tensors="pt", sampling_rate=16000)
generated_ids = model.generate(inputs["input_features"])
transcription = processor.batch_decode(generated_ids, skip_special_tokens=True)
print(transcription)