When Audio Separation Hurts Zero-Shot ASR: Evaluating SAM-Audio with Whisper on Bengali and English Speech

arXiv cs.AI · 18 hours ago

Researchers tested SAM-Audio, an audio separation tool, as preprocessing for OpenAI Whisper's zero-shot speech recognition on Bengali and English datasets. On English speech, SAM-Audio improved audio quality metrics (PSNR increased from 32.28 dB to 35.99 dB) but Whisper's word error rate rose from 10.53% to 21.66% for the base model. The study shows that cleaner audio does not guarantee better transcription accuracy, suggesting audio preprocessing can harm zero-shot ASR performance.