Text to speech API with voice cloning and a TypeScript SDK

An OpenAI-compatible speech endpoint

If your code already sends an OpenAI-compatible speech request, point it here: change the base URL to https://app.voicelabs.now/v1 and the key to a VoiceLabs key. POST /v1/audio/speech holds the request open until the audio is ready and returns the bytes, so there is nothing to poll.

curl https://app.voicelabs.now/v1/audio/speech \
  -H "x-api-key: $VOICELABS_API_KEY" \
  -H "content-type: application/json" \
  -d '{"model":"chatterbox","voice":"Narrator","input":"Welcome back. The briefing is ready.","response_format":"mp3"}' \
  --output speech.mp3

The TypeScript SDK

@voicelabs/sdk is the typed client for the same API. Install it with npm install @voicelabs/sdk. It has no runtime dependencies, ships ESM and CommonJS with types, and runs on Node 18 and later, Bun, Deno, Cloudflare Workers and in browsers. Field names are the wire names, so what you read here, in the OpenAPI document and in a curl response is one set of names.

import { writeFile } from "node:fs/promises";
import { VoiceLabs } from "@voicelabs/sdk";

const voicelabs = new VoiceLabs({ apiKey: process.env.VOICELABS_API_KEY! });

// Clone a voice from a short recording (the key needs the voice:clone scope).
const voice = await voicelabs.cloneVoice({
  name: "Narrator",
  reference_text: "The words spoken in the recording.",
  audio_url: "https://example.com/recording.wav",
});

// Speak in it. generateSpeech starts the job and waits for the audio.
const generation = await voicelabs.generateSpeech({
  text: "Welcome back. The briefing is ready.",
  voice_id: voice.id,
});

// The stored file is WAV.
const { bytes } = await voicelabs.downloadAudio(generation);
await writeFile("speech.wav", bytes);

Errors are typed: a spent monthly allowance, a rate limit and a missing scope each have their own class, so a retry loop does not hammer a limit that only the plan can lift. The package is MIT licensed and published on npm; the API docs say more.

Voice cloning over the API

POST /v1/voices/clone creates a cloned voice from a 2 to 30 second recording of clear speech, given as base64 audio or as the address of a public https file, together with the words spoken in it. It needs a key with the voice:clone scope, which is its own scope on purpose: a key made to generate speech was not made to build a voice from a recording. The response carries the new voice’s id, which you pass as voice_id to the speech endpoint. The voice lives on your account, so the studio, the MCP server and the browser extension list it too. Clone only a voice you have the right to use: the responsible use policy says what that means.

Every endpoint

The API reference documents every parameter and response, generated from the same OpenAPI 3.1 document the server publishes at https://app.voicelabs.now/openapi.json. Prefer an AI assistant to HTTP? The MCP server exposes the same voices as tools.

Limits, errors and the Pro gate

• Pro only. On an account without an active subscription or trial, every request, reads included, is refused with 403 pro_plan_required. Keys are not revoked: the same key works again on the next request once the account is back on Pro.
• Rate limits. Each key may make 600 requests an hour that do work, such as starting a generation. Reads are counted separately, at 300 requests a minute, so polling cannot lock you out of the next create. Responses carry RateLimit and RateLimit-Policy headers.
• Errors you can branch on. Every error is an RFC 9457 problem document with a stable code. Branch on it, not on the status: rate_limit_exceeded clears by waiting, while quota_exhausted means the plan’s monthly audio is spent. Each code has a page of its own.
• Safe retries. Send an Idempotency-Key header on a write and a retry replays the original response instead of generating twice.

Engines and languages

All seven text-to-speech engines are built on open-source models and hosted by us, and each has an id you can send as model to the compatible endpoint. Together they cover 23 languages.

The engines page lists the upstream project behind each one, and the language index says which engines speak which language. For the case for hosted rather than local models, see cloud text to speech.

What it costs

One plan: VoiceLabs Pro, US$8 a month or US$59 a year, with a 7-day trial that asks for a card up front. It includes API access, voice cloning, all seven engines and 2 hours of generated audio a month (fair use). It is a flat subscription: no per-character fee and no usage bill. The pricing page has the detail.

Questions about the text to speech API