Developers · SDK · CLI · self-hosted API

Add voices to your app. Pay nothing.

Hosted voice APIs bill you per character because their GPUs do the work. With Kelados your own machine (server, laptop or the user's browser) does it, so it's free at any scale. Swap it in behind the TTS API you already use.

Node SDK

One package. Downloads the 82M-parameter Kokoro model on first run (~90 MB, cached), then generates speech offline on CPU.

bash

npm install kelados

javascript

import { createKelados } from "kelados";

const kelados = await createKelados();          // runs locally, no key needed

const clip = await kelados.speak("Hello from my app!", { voice: "af_heart", speed: 1.05 });
await clip.save("hello.wav");                   // or clip.toWav() → Buffer

// Stream long text sentence by sentence
for await (const chunk of kelados.stream(longChapter, { voice: "bm_george" })) {
  player.enqueue(chunk.audio, chunk.sampleRate);  // Float32Array @ 24 kHz
}

Add an API key to use the presets and pronunciation rules from your console:

javascript

const kelados = await createKelados({
  apiKey: process.env.KELADOS_API_KEY,   // kel_...
  baseUrl: "https://kelados.vercel.app",
});
await kelados.speak("Kelados ships SQL tooling", { preset: "Narrator" });  // lexicon applied automatically

CLI

bash

npx kelados speak "The quick brown fox" -v am_fenrir -o fox.wav
npx kelados speak -f chapter-01.txt -v bf_emma -s 0.95 -o chapter-01.wav
npx kelados voices

Drop-in API server

Run a local server that speaks both the /v1/audio/speech and /v1/text-to-speech/:voice dialects. Point your existing SDKs at it and you're done.

bash

npx kelados serve --port 8880 --key my-secret
# Docker / VPS / Raspberry Pi / your laptop: anything with Node 20+

bash

curl http://localhost:8880/v1/audio/speech \
  -H "Authorization: Bearer my-secret" -H "content-type: application/json" \
  -d '{"model": "tts-1", "voice": "nova", "input": "Same code, zero bill."}' -o out.wav

bash

curl http://localhost:8880/v1/text-to-speech/af_bella \
  -H "xi-api-key: my-secret" -H "content-type: application/json" \
  -d '{"text": "Same request, Kelados-shaped invoice: $0."}' -o out.wav

Common voice names map to Kokoro voices (alloy, echo, fable, onyx, nova, shimmer…). Output is WAV or raw PCM.

In your users' browsers

Want your web app to talk without any backend? Do what this site does: load the model in a Web Worker and let each visitor's GPU do the work.

javascript

// worker.js  (type: "module")
import { KokoroTTS, TextSplitterStream } from "https://cdn.jsdelivr.net/npm/kokoro-js@1.2.1/dist/kokoro.web.js";

const tts = await KokoroTTS.from_pretrained("onnx-community/Kokoro-82M-v1.0-ONNX",
  { dtype: navigator.gpu ? "fp32" : "q8", device: navigator.gpu ? "webgpu" : "wasm" });

self.onmessage = async ({ data }) => {
  const s = new TextSplitterStream(); s.push(data.text); s.close();
  for await (const { audio } of tts.stream(s, { voice: data.voice }))
    self.postMessage({ audio: audio.audio, rate: audio.sampling_rate });
};

Our full worker lives in frontend/public/workers/tts.js in the repo. Copy it.

API keys

Create keys in the console. Send them as x-api-key: kel_… or Authorization: Bearer kel_…. Keys are stored as SHA-256 hashes, shown once, and revocable any time.

No creditsNo per-character billingText never leaves your machine

REST API

Base URL https://kelados.vercel.app. Every endpoint accepts an API key except the public stats endpoint.

GET	/api/v1/workspace	Your presets + pronunciation lexicon (what the SDK syncs)
POST	/api/usage	{ characters, seconds, source } logs usage to your dashboard
GET	/api/usage?days=30	Daily usage rows and totals
GET/POST	/api/projects	List / create long-form projects
GET/PUT/DELETE	/api/projects/:id	Read, replace or delete a project
GET/POST	/api/presets	List / create voice presets
PUT/DELETE	/api/presets/:id	Update / delete a preset
GET/PUT	/api/lexicon	Read / replace pronunciation rules
GET	/api/stats	Public counters (no auth)

bash

curl https://kelados.vercel.app/api/v1/workspace -H "x-api-key: kel_..."

FAQ

Is it really unlimited?

Yes. Generation happens on your device, so there is no meter to run out. The only limit is how long you're willing to wait.

How good are the voices?

Kokoro-82M ranked #1 on the community TTS Arena at launch, beating models 10-100× its size. The A-grade voices hold up on long-form narration.

Which languages?

Speech: American and British English today. Transcription: 99 languages via Whisper.

Can I use the audio commercially?

Yes. The model is Apache-2.0 and Kelados is MIT. Output is yours.

Why do I need an account at all?

You don't. Accounts only sync projects, presets and pronunciation rules, and issue API keys.

Open source

Frontend, backend and SDK are MIT-licensed. Fork it, self-host it, rebrand it. Voices come from Kokoro-82M (Apache-2.0) by hexgrad; the browser runtime is transformers.js by Hugging Face.

Try the studio