whisper-large-v3-turbo-ro — Romanian Fine-tune
Fine-tuned openai/whisper-large-v3-turbo
on a merged Romanian speech corpus (128-mel input features).
Performance
| Metric | Normalized | Raw |
|---|---|---|
| WER | 0.063 | 0.0659 |
| CER | 0.0208 | 0.0214 |
Normalized = lowercased, punctuation stripped (comparable to Whisper papers). Evaluated on 2000 validation samples, beam=5.
Metrics read from eval_results.json.
Usage
from transformers import WhisperProcessor, WhisperForConditionalGeneration
import torch
processor = WhisperProcessor.from_pretrained("IonGrozea/whisper-large-v3-turbo-ro")
model = WhisperForConditionalGeneration.from_pretrained("IonGrozea/whisper-large-v3-turbo-ro")
inputs = processor(audio_array, sampling_rate=16000, return_tensors="pt")
with torch.no_grad():
ids = model.generate(inputs.input_features, language="romanian",
task="transcribe", num_beams=5)
print(processor.batch_decode(ids, skip_special_tokens=True)[0])
This card was auto-generated at upload time because the source folder had no
README.md (e.g. a raw checkpoint). For the full training-config card, upload
from the final_best/ directory instead.