Product · Voice platform & API

A voice API for developers.

Speech recognition, dialogue and synthesis — three HTTP calls. A familiar OpenAI-compatible format, sentence-by-sentence response streaming and natural barge-in. Voice in your product — in one evening.

Free in beta — 2,000 requests a day, no card.

Talk to fg-voice-1 — by voice, right now.

Model

One model. Everything voice needs.

fg-voice-1beta · free

The flagship voice model for business assistants: real-time dialogue with natural barge-in.

Response
instant opening line
Response streaming
by sentences, SSE
Interruption (barge-in)
echo-robust
Languages
Russian · English
Company knowledge
up to 5 documents per persona
Price in beta
$0
More about the model
Quickstart

From audio to answer —
three calls.

A familiar HTTP API

Send audio, get a stream of sentences, speak them. Each step is a separate endpoint with your code in between. Or grab it all at once from the ready web client.

  • POST /v1/audio/transcriptions
  • POST /v1/chat/completions · SSE, by sentences
  • POST /v1/audio/speech
# Assistant voice in one request
curl https://fg-platform.com/v1/audio/speech \
  -H "Authorization: Bearer $FG_API_KEY" \
  -d '{
    "model": "fg-voice-1",
    "voice": "sergei",
    "input": "Sir, the report is ready to send."
  }'
stream · SSE · by sentences
Migration

OpenAI-compatible.
Two lines change.

Same SDK, same format

The format matches OpenAI: the same messages array, the same chat.completion response, the same chunked streaming. In the official SDK you only change the base URL and the key.

  • gpt-4o → fg-voice-1
  • voice: "alloy" → voice: "sergei"
  • Realtime API → three calls + an opening line
import OpenAI from 'openai';

const client = new OpenAI({
  baseURL: 'https://fg-platform.com/v1',
  apiKey: process.env.FG_API_KEY,  // fg-…
});

// the rest of your code stays the same
const chat = await client.chat.completions.create({
  model: 'fg-voice-1',
  messages: [{ role: 'user', content: 'Hello!' }],
});
Features

Built for production

Sentence streaming
The response arrives as SSE deltas of whole phrases — each can go straight into speech synthesis.
Interruption (barge-in)
Echo-robust: the model tells the user's voice apart from its own sound coming from the speaker.
Languages RU/EN
Russian and English out of the box; language is detected and synthesized automatically.
Voices
A library of natural voices; by name or by ID from the catalog (700+ voices).
Company knowledge
Up to 5 documents per persona connect without training the model.
Keys and limits
Dashboard: API keys, call stats, daily quota. Free in beta.
Getting started

Add voice today

Build it yourself

  • An API key in a minute, no card
  • Docs and code samples
  • A ready web client to embed
Get an API key

We'll integrate together

  • We'll help design the voice scenario
  • We'll connect your knowledge bases and CRM
  • Persona and voice for your brand
Contact us