# Optional: transcribe audio/video into Memory (downloads a small speech model once)
faster-whisper>=1.0
