Building a Tamil voice companion: STT/TTS choices, long-term memory, scaling concurrency
intrepidkarthi · reddit · 2026-08-20
A developer is building a Tamil voice companion for 5–10 minute long conversations (not a task bot). Current stack: Sarvam Saaras STT, own LLM, TTS, all over LiveKit, with 2s voice-to-voice latency. The quality bar is ChatGPT's Tamil voice conversation (naturalness and turn-taking), but they need a cascade pipeline since safety gates and memory live at the text seam.
Three open questions:
- Tamil stack: Google Chirp3 HD sounds better than Sarvam Bulbul but has no pitch control or Tamil custom pronunciation; ElevenLabs Flash lacks Tamil and Deepgram Nova-3 is 68% WER on Tamil — any third option?
- Memory: using structured extraction into SQLite (facts with validity windows) instead of RAG to keep the prompt cache warm; curious how Graphiti/Zep or Mem0 hold up for non-English voice agents.
- Scaling: self-hosted LiveKit Agents vs Pipecat — where did it break? Suspects are batch STT and non-streaming TTS; moving to Sarvam's streaming websocket endpoints next, with benchmark numbers promised.
Related event: Devs Debate Tech Stack for Tamil Voice Companion App(2 posts)→
More from coding & agent
- Visualizing a week of Claude Code work: expanding file changes chronologically — repligate · 2026-08-20
- Connect Claude Code to Grok via OAuth Without API Keys — socialwithaayan · 2026-08-20
- Turn Grok Bot into an always-on job hunter that applies to 20 roles a day — bigaiguy · 2026-08-20
- I ran the 'Claude solves SEO' loop for weeks—here's why those viral posts are BS — ayushtweetshere · 2026-08-20
- As AI agents plug into Gmail and Drive, prompt injection flaws demand strict access controls — emmanuelvivier · 2026-08-20
- Temporal knowledge graph memory engine cuts latency by 90% and beats MemGPT — anirbanbandyo · 2026-08-20