Google's API is solid and probably still wins on raw latency for short clips, but it's cloud-only, and that was the dealbreaker given what we're transcribing. AssemblyAI had nice extra features (summarization, speaker labels) bolted on top, but we didn't need the extra product we needed the core transcription to be accurate and to run somewhere we control. Whisper being open-weight meant we could self-host it instead of trusting a vendor's retention policy with early, unfiltered voice notes.