Most speech APIs are built for speed: a sentence back in a few hundred milliseconds, for a voice assistant or a phone menu. The Everlit TTS API is built for the other job, the listen that lasts twenty minutes. It is the same voice model that narrates for newsrooms every day, on its own, behind one endpoint.
One request
Send text and a voice. You get back a job with a stable audio address, and you collect the finished file by polling or by webhook, as MP3, WAV or Opus.
curl -X POST https://api.everlit.audio/v1/speech \
-H "Authorization: Bearer api_YOUR_KEY" \
-H "Content-Type: application/json" \
-d '{"text": "Hello from Everlit.", "voice": "669"}'
# => 202 { "id": "spch_01jx...", "status": "queued",
# "audio_url": "https://api.everlit.audio/v1/speech/spch_01jx.../audio" }Voice 669 is Lana Clark, one of the stock voices. The rest of the catalog is one call away, and so are your own cloned voices.
Slow on purpose
A job usually finishes in 15 to 60 seconds, because every file gets a full render and an automatic quality check before you see it. That makes the API a good fit for articles, archives, audiobooks, courses and anything narrated at the moment it is published. It is the wrong choice for a live conversational agent or a phone system that needs speech in milliseconds, and we would rather say so here than have you find out in testing.
A clock on every word
A finished job can return word-level timings: each word you sent, with a start and an end. They are built from your text, not from a speech recognizer’s guess at it, so “84%” comes back as one entry whose span covers the spoken “eighty-four percent.” The same timings come as SRT or WebVTT captions. That is what makes read-along text, captions and subtitled video possible without a second service. Timings are available in 25 of the API’s 69 languages, at no extra cost, and you can watch a read-along play on the API page.
Direction inside the text
Long-form narration needs direction, and the direction goes in the text itself. Inline tags add a pause, bring a music bed in or duck it under a passage, and hand a quotation to a second voice. Speed and the pause between paragraphs are options on the request. A pronunciation library fixes a name once for every future job. And you can clone a voice by API from a 15-second sample, with the speaker’s consent on record; the clone is private to your account and costs the same as ours.
What it costs
$8 per million characters for Latin, Cyrillic and Greek scripts, prepaid, with no monthly fee. Other scripts are weighted so that every language costs about the same per finished hour of audio. In English, that works out to roughly:
- $0.06 for a 1,200-word article
- $0.46 for a 10,000-word long read
- $4.12 for a 90,000-word audiobook
Word timings, captions, every format and cloned voices are included at that rate. Failed jobs are never charged, and credit never expires. If you are on a paid Everlit Studio plan, speech is already included.
Made for pipelines
- Signed webhooks with retries, so you do not have to poll.
- Idempotency keys, so a request you send twice is only made once.
- An audio address that is known the moment you submit and stays stable for 30 days.
- An OpenAI-compatible speech endpoint: point the OpenAI SDK at Everlit’s base URL and keep your code.
Getting a key
There is no self-serve signup. Request a key and a person reads the request, usually within one business day. The API documentation is open in the meantime. And if you would rather not write code at all, the same voice is available to AI assistants through the Everlit MCP server.