This is a lightweight transcription MVP: upload audio, we keep only the first 2 minutes, remove silence, and generate a transcript.