Get started
Overview
What Vaakini is, how it speaks a reply, and the ways to connect to it.
What Vaakini is#
Vaakini is a text-to-speech API for voice agents. It has one model, vaakini-tts-1, which speaks 12 Indian languages.
| Vaakini | |
|---|---|
| Model | vaakini-tts-1 |
| Languages | Assamese, Bengali, English, Gujarati, Hindi, Kannada, Malayalam, Marathi, Odia, Punjabi, Tamil, Telugu |
| Voices | 14 of ours, and your own, each made from a short recording |
| Audio | 24000 Hz, mono: 16-bit PCM, MP3 or WAV |
| Price | ₹0.30 per 1,000 characters spoken (Pricing and billing) |
How a reply is spoken#
Sentence by sentence, as the text arrives.
Send text whole or as your LLM writes it. Vaakini splits it into sentences and speaks each one as soon as it is complete, in order. Audio is streamed back while it is being made, so playback can start with the first sentence while the rest of the reply is still being written.
A sentence ends at its end mark (., ?, !, ।), never at a comma. Each sentence is sent to the model as you wrote it, apart from surrounding whitespace.
Ways to connect#
Three clients, and the API they are built on.
- LiveKit pluginlivekit-plugins-vaakini, as your agent's TTS.
- Pipecat servicepipecat-vaakini, in your pipeline.
- Python SDKvaakini, for sync and asyncio code.
- HTTP APIPOST /v1/audio/speech: a text in, its audio out.
- WebSocket API/v1/stream: text in as it is written, audio out per sentence.
- QuickstartFrom sign-up to the first sentence in a few minutes.
The API#
One host, one key.
| Surface | URL |
|---|---|
| HTTP | https://api.vaakini.com |
| WebSocket | wss://api.vaakini.com/v1/stream |
Every request carries an API key from the dashboard’s API keys page. Keys start with vk_; the clients read one from VAAKINI_API_KEY.
Authorization: Bearer vk_…A key belongs to an org, which owns its keys, custom voices, usage and credit. The API reference lists every endpoint.
Good to know#
- The first audio follows the first sentence’s end. A sentence is spoken once it is complete, so while text is streamed in, nothing is spoken until the first end mark arrives (or the reply ends).
- The same text can sound slightly different each time. The model samples, so two takes of one sentence are not identical. A lower temperature makes takes steadier.
- A custom voice created seconds ago can be refused for a few seconds. Wait about five seconds before the first request with a new voice.
- Limits. A text, or a reply on the WebSocket, may be up to 20,000 characters. A WebSocket that sends nothing for 30 minutes is closed.