Thanks — I’m responding to your question about which models to self-host for a full STT → TTS → LLM pipeline. Short answer • For STT: NVIDIA Riva • For TTS: NVIDIA Riva • For LLM: Ollama Why these recommendations • NVIDIA Riva supports self-hosted streaming STT and TTS, multilingual voices, and deployment on your own servers—so one vendor handles both speech directions and simplifies integration and latency tuning. • Ollama runs locally and exposes an OpenAI-compatible endpoint (you point the base_url to your local Ollama server), letting you run modern LLMs on-prem without relying on cloud API keys. Next practical steps (high-level) 1. Set up NVIDIA Riva server for STT and TTS. ◦ Deploy Riva on a machine with a compatible NVIDIA GPU. ◦ Configure the Riva URIs in your application (e.g.,

riva_uri

) and tune language/voice models you need. 1. Install Ollama locally for LLM. ◦ Run Ollama on your host and expose its API (for example `http://localhost:11434/v1`). ◦ Choose the model you want (e.g., Llama-family models) and point your app’s OpenAI-compatible client to Ollama’s base URL. 1. Integrate the pipeline. ◦ Use Riva for real-time STT to transcribe audio into text. ◦ Send transcriptions to Ollama for intent, dialogue management, or text generation. ◦ Send Ollama output to Riva TTS to synthesize audio responses. 1. Test latency and scale. ◦ Measure end-to-end latency and tune batching, model sizes, and GPU resources. ◦ If you need higher throughput, scale Riva and Ollama instances horizontally or choose smaller/faster models. If you want, I can: • Give specific deployment steps or a checklist to install NVIDIA Riva and Ollama. • Recommend hardware sizing (GPU/CPU/RAM) based on expected concurrency. • Help pick particular Riva language/voice models or Ollama model variants for accuracy vs latency trade-offs. Which of those would you like next? Sources: NVIDIA Riva STT plugin guide | LiveKit Documentation | NVIDIA Riva TTS plugin guide | LiveKit Documentation | Ollama LLM plugin guide | LiveKit Documentation