Early access · self-hosted voice AI

Voice AI for 15 languages, Hindi to Spanish.

Build conversational IVR and WhatsApp agents across 11 Indian languages, Hinglish and Spanish — with real speech-to-text, text-to-speech, and an LLM router with automatic provider failover. Self-host it, or request an early-access account.

No payment required Self-hosted or managed Tenant-isolated by default
Illustrative example Not a live call
Caller

Namaste, mujhe 100 unit ka quotation chahiye.

Vaani

Sure! Aap kis city mein hain? Main aapko abhi WhatsApp pe quotation bhej deta hoon.

15
Languages understood
4
Speak back in their own voice
3
LLM providers with failover
Self-hosted
Or managed early access

What's built so far.

One platform for telephony, AI and messaging. Bring your own SIP trunk, number, or Meta credentials.

Conversational IVR

State-machine IVR with slot-filling. Reachable over WebRTC (Pipecat) or your own SIP trunk (Jambonz).

STT + TTS

AI4Bharat IndicConformer STT for 21 Indian languages (bring your own endpoint), Groq Whisper for English and Spanish. Piper TTS bundled for English, Hindi, Marathi and Spanish.

LLM router with failover

Groq → OpenRouter → Cerebras, in order. Free-tier friendly, with no single point of failure.

WhatsApp & Instagram

Meta Cloud API integration for WhatsApp and Instagram messaging — bring your own Meta app credentials.

Quotation drafting

The agent extracts structured quotations from conversation. Pricing and catalog data must be confirmed by you before it's ever shown to a customer — nothing is guessed.

Own your data

Self-host with one docker-compose up. Every tenant's data, keys, and knowledge base are isolated.

Built for Bharat — and beyond.

Fifteen languages understood end to end: speech-to-text, intent, and quotation extraction — including Hinglish code-switching.

Speaks back in its own voice

English हिन्दी Hindi मराठी Marathi Español Spanish

Understood — answered in the English voice

Hinglish বাংলা Bengali தமிழ் Tamil తెలుగు Telugu ગુજરાતી Gujarati ಕನ್ನಡ Kannada മലയാളം Malayalam ਪੰਜਾਬੀ Punjabi ଓଡ଼ିଆ Odia অসমীয়া Assamese اردو Urdu

The split above is deliberate, not marketing: the first group has a real bundled Piper voice, so replies are spoken in that language. The second group is transcribed and understood (AI4Bharat IndicConformer for the Indic languages, Whisper for Spanish) and answered in the English voice until a real voice model exists for it. Hinglish is romanized Hindi read with the English voice by design.

Developer-first.

One REST API for chat, STT, TTS and translation. Works from curl, Python, Node — or your own client.

  • OpenAPI 3.1 spec at /api/docs
  • Cookie-session or bearer-token auth
  • Self-host with one docker-compose up
View API docs →
# Ask Vaani a question in any of the 15 supported languages
# (GET /api/languages returns the list)
curl $VAANI_BASE_URL/api/chat \
  -H "Authorization: Bearer $VAANI_KEY" \
  -d '{"lang":"hien","message":"Delhi ka weather?"}'

# Synthesize speech (English, Hindi, Marathi or Spanish voice)
curl $VAANI_BASE_URL/api/tts \
  -d '{"text":"Hello there","voice":"en"}' \
  -o hello.wav

Try Vaani today.

Self-host it, or create an early-access account — no payment required right now.

Create your account →