Dynamic Accent Duplex experimental cascaded voice agent

Speak English, Nigerian Pidgin, Igbo, Yoruba or Hausa, mixed however you like. Ada answers in the same mix. You can interrupt her. Use headphones (otherwise her voice can trigger barge-in).

Pipeline: Silero VAD → Omnilingual ASR (CTC) → per-word language ID → Qwen3 LLM with mix validator → Dynamic Accent Flash TTS. Research demo; replies can be wrong, and Igbo/Yoruba/Hausa beyond simple phrases are weak.