Shagufta/whisper-small-km-indomain5-withoutac

Whisper Small - Kreol Morisien (In-domain 5 hours, without AC)

Fine-tuned openai/whisper-small on 5 hours of in-domain Khmer speech data.

Training Details

• Base model: openai/whisper-small
• Language: Kreol Morisien (km)
• Training data: 5 hours in-domain
• Best step: 400 / 630 (epoch ~19 of 30)
• Best WER: 20.33%
• Best validation loss: 0.5536
• Learning rate: 3e-5
• Scheduler: cosine
• Batch size: 8 × 4 (effective 32)
• Weight decay: 0.05
• Warmup steps: 100
• Precision: bf16

Usage

from transformers import WhisperForConditionalGeneration, WhisperProcessor, pipeline

processor = WhisperProcessor.from_pretrained("Shagufta/whisper-small-km-indomain5-withoutac")
model = WhisperForConditionalGeneration.from_pretrained("Shagufta/whisper-small-km-indomain5-withoutac")

pipe = pipeline("automatic-speech-recognition", model=model, tokenizer=processor.tokenizer, feature_extractor=processor.feature_extractor)
result = pipe("audio.wav")
print(result["text"])