Thanks — I’m responding about how to use custom TTS and STT models with a realtime LLM for transcription. Short answer: • You can run a realtime LLM for live reasoning/audio and pair it with a separate STT pipeline for timestamped transcripts, or run the realtime LLM in text-only mode and route its output through a custom TTS. Achieve this by overriding the agent’s

stt_node()

/

tts_node()

or by using an

inference.STT()

/

inference.TTS()

plugin for your chosen provider. Quick guidance and next steps: 1. Which architecture do you want? ◦ Hybrid (realtime LLM + separate STT for transcripts and timestamps) ◦ Half-cascade (realtime LLM in

modalities=["text"]

+ custom TTS for audio output) 1. Which environment / SDK are you using? (Python, Node.js, or other) 2. Which custom STT/TTS provider or model do you plan to use? (e.g., Deepgram, AssemblyAI, ElevenLabs, Cartesia, a self-hosted model) If you tell me the answers above I’ll provide a small code example and the exact node override or plugin configuration for your environment. Sources: Pipeline vs. Realtime - Which is the better Voice Agent Architecture? | Liv... | Realtime models overview | LiveKit Documentation | Pipeline nodes and hooks | LiveKit Documentation | Speech-to-text (STT) models overview | LiveKit Documentation | Text-to-speech (TTS) models overview | LiveKit Documentation, with interactive elements