First phase: OpenAI Realtime voice. Spoken output remains English; the transcript panel shows bilingual text where available.