CTranslate2-optimized Whisper implementation — 4× faster inference, lower memory than OpenAI's original