Real Estate Voice Agent

Configuration

TTS — Sarvam Bulbul v3 · Rumik Silk Model Speaker Custom speaker (overrides dropdown) Language Pace 1.0 Temperature 0.01 0.01 = maximum voice consistency across the call (both Sarvam and Rumik re-sample their delivery per reply — this is the tightest either allows). Raise toward 0.2–0.3 only if the voice sounds too flat.

STT — Sarvam Model Language hint

Reasoning — LLM Model

Turn-taking & edge cases End-of-sentence silence 800 ms Higher = speaker gets more time to finish their sentence before the bot replies. Clearly complete sentences still commit early; an extra ~900 ms is added automatically when the sentence sounds unfinished. Mic sensitivity (min) 0.012 Lower = picks up quieter speech (but more background noise). Noise filter — high-pass/low-pass + compressor on the mic before STT (applied at call start) Adaptive threshold — auto-raises sensitivity above ambient noise (ambient: –) Allow barge-in (user can interrupt the bot; off = bot always finishes speaking) Debug: PCM direct output — bypass the hidden-<audio> AEC route. If the bot's voice stops "thickening" with this ON, the element playout is the cause. (Barge-in degrades while on.) Barge-in profile Barge-in sensitivity 1.6 Higher = you must speak louder/closer to interrupt (ignores TV & background talkers). Barge-in confirm time 300 ms Bot audio ducks while confirming; too short = false interrupts, too long = bot talks over you. No-response nudges before hang-up

⏱ Latency debug

wait (silence + hold) STT LLM thinking TTS synth network

No turns yet.

"reply in" = your last word → agent's first audio. Per turn: how long the endpointer waited, transcription, the LLM's first sentence, TTS synthesis, and network overhead.

📼 Call history

No calls yet.

📞 Lead: Website enquiry — 3BHK, Kalyan West

00:00

Idle

Press Start Call. The agent (Priya) speaks first, then it's your turn — speak in Hindi, English, or Hinglish. Pause when you finish a sentence and the agent will reply.