Product · Voice API

Real-time voice, three HTTP calls.

fg-voice-1 hears, answers and speaks over an OpenAI-compatible API. Point your existing SDK at our base URL; the reply streams back sentence by sentence, each one ready to synthesize.

Get an API key Read the docs

Free in beta — 2,000 requests a day, no card. Paid plans will be announced in advance.

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

01 · The loop

Hear. Think. Speak.

Three endpoints, your code in between. No WebRTC, no sockets, no SDK to install — plain HTTP with one SSE stream in the middle.

Hear

POST /v1/audio/transcriptions

Multipart audio — webm/opus, mp3, wav, m4a, up to 32 MB — comes back as text. Silence and recognition noise return an empty string, so the assistant simply keeps listening.

Think

POST /v1/chat/completions

With stream: true every SSE delta is a finished sentence, not a token. Hand it straight to synthesis instead of buffering the whole reply.

Speak

POST /v1/audio/speech

Text up to 4096 characters becomes mp3 at 44.1 kHz, sent back as a chunked stream you can start playing before it ends.

        
        
      

        
# text → mp3, one call
curl https://fg-platform.com/v1/audio/speech \
  -H "Authorization: Bearer $FG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "fg-voice-1",
    "voice": "sergei",
    "input": "Sir, the report is ready to send."
  }' --output reply.mp3
      

        
import OpenAI from 'openai';
import fs from 'node:fs';

// the same SDK — the base URL and the key change
const fg = new OpenAI({
  baseURL: 'https://fg-platform.com/v1',
  apiKey: process.env.FG_API_KEY,     // fg-…
});

const heard = await fg.audio.transcriptions.create({
  file: fs.createReadStream('question.webm'),
  model: 'fg-voice-1',
});

const stream = await fg.chat.completions.create({
  model: 'fg-voice-1', stream: true,
  messages: [{ role: 'user', content: heard.text }],
});

for await (const part of stream) {
  const line = part.choices[0].delta.content; // a whole sentence
  if (!line) continue;
  const speech = await fg.audio.speech.create({
    model: 'fg-voice-1', voice: 'sergei', input: line,
  });
  play(Buffer.from(await speech.arrayBuffer()));
}
      
stream · SSE · by sentences

02 · Numbers

Latency, measured.

Median time each call spends on our side, taken from live platform traffic rather than a benchmark rig.

213msSpeech → textA spoken turn of about 2.5 s recognised and returned.

724msModel starts answeringTime to the first byte of the streamed reply.

135msSentence → audioOne spoken sentence rendered to mp3.

Median of live platform traffic, 14–19 August 2026 (n = 109 / 346 / 352 calls). Our side of the call only — the network between you and us is not ours to promise.

Model: fg-voice-1 — one id for all three calls; listed by GET /v1/models.
Streaming: SSE. Each delta is a finished sentence, ready to synthesize — not a token you have to reassemble.
Voices: By name, sergei — also what you get when voice is left out. By id, any of 900+ voices in the catalog.
Languages: Speech in and out: English and Russian. Text: any language the visitor writes in. Recognition takes the language from the audio — nothing to set per request.
Audio out: mp3, 44.1 kHz, mono, 128 kbps, chunked. Input up to 4096 characters per call.
Audio in: webm/opus, mp3, wav, m4a — up to 32 MB per file, multipart/form-data.
Compatibility: OpenAI request and response shape, including errors: error.type, error.code — null when the failure came from a provider rather than from the request.
Context: max_tokens 1–4096, 1024 by default. Your own system message sets the persona; the last 40 messages of messages are kept.
From the browser: CORS is open; the key travels in the Authorization header. Keep it server-side in production.
Limits: 120 requests a minute per key, 2,000 a day per account in beta.

03 · Start

A key in a minute. $0 in beta.

No card. Sign up, confirm your email, create a key — the next line you write is a working call.

Build it yourself

Sign up and confirm your email — an unconfirmed account can't call the API
Create a key in the console — fg-… is shown once, save it
Point your SDK at https://fg-platform.com/v1

# check the key works curl https://fg-platform.com/v1/models \ -H "Authorization: Bearer $FG_API_KEY"

Get an API key Read the docs

Free in beta — 2,000 requests a day, no card. Paid plans will be announced in advance. Existing keys keep working.

What beta gives you

2,000 requests a day per account
120 requests a minute per key
900+ voices by id, English and Russian speech
Keys, usage and the daily counter in the console

Platform plans Volume or SLA — talk to us