AI employees
Voices & languages
Each AI employee picks a voice model — the engine that listens and speaks — and a voice. Set both on the Voice tab of the AI employee's page.
Voice models
Gemini Live
Google's real-time voice model. The recommended choice — fully streaming, with the lowest delay (around 300–600 ms).
Ultravox
Real-time streaming voice model.
Hume EVI
Expressive voice that responds to the caller's emotion.
Sarvam AI
Built for Indian languages — Hindi and regional languages, with Indian voices.
Cascaded (BYO)
Build your own pipeline from three parts — a transcriber, a language model and a voice — using your own provider keys.
When you switch model, Daffy Voice also switches to one of that model's own voices — a voice from one model does not exist in another.
Indian languages
Through Sarvam voices, an AI employee can speak:
- Hindi · English (India) · Tamil · Telugu · Kannada · Malayalam · Bengali · Gujarati · Marathi · Punjabi · Odia
Two ways to use them
- Sarvam AI model — pick a Sarvam voice (Priya, Neha, Rahul, Aditya and others). It speaks the language your workspace is set to, which is Hindi by default.
- Cascaded (BYO) with Sarvam (Hindi) as the voice provider — here you choose the Language per AI employee from the list above.
Tip
Cascaded (bring your own keys)
A cascaded pipeline has three stages you choose separately:
- Transcriber (STT) — Deepgram, with a choice of models and an auto-detect or multilingual setting.
- Model (LLM) — Google Gemini, Groq, DeepSeek, or any OpenAI-compatible service with your own base URL.
- Voice (TTS) — ElevenLabs, OpenAI, Deepgram Aura or Sarvam; you can also paste a custom voice ID.
Keys go in Settings
Advanced settings
Under Advanced · LLM settings: Temperature (lower is more predictable — about 0.3 for support, 0.7 for sales), Max response tokens (keeps replies short), Interruption sensitivity (0 lets the caller finish, 1 stops the moment they speak) and Top P. Not every model uses every setting.