Transcription API

• Free: 2 a month, the same as the website. Then $2 an hour, bought in packs.
• 2 free a month per account with a key, counted together with that account’s free use of the website, resetting on the 1st (UTC); 1 a day per address without a key.
• 1 hour a file free with a key, 3 hours on API hours, 10 minutes without a key.
• API hours: $20 for 10 hours or $90 for 50 hours. Charged by the minute, rounded up; a transcript that fails is given back in full. They stay with the key you buy them with (keep it safe: no other key can use them, and a lost key cannot be recovered), never expire, and are used only once the month’s free ones are gone. Buy API hours.
• 600MB a file.
• English. The language is detected and reported; other languages are not promised.
• No speaker labels. You get the words, not who said them.
• Sign in to get a free key. Up to five active keys per account, three made a day. Every key on an account shares its free allowance, so more keys never mean more. Keys made before 30 Oct keep working; sign in to manage your keys.
• The free ones are free. No card, no trial, no watermark. We cap what the free tool can spend in a month; when it runs out the API says so and gives the date it resets. API hours are not affected.

What happens to the file

A URL is fetched by the transcription provider and by us for a few hundred kilobytes, to read its length before anything is spent. Nothing is stored except the transcript. Audio uploaded through the web tool is deleted after 7 days, text after 30, by a job that runs daily. Nothing is used to train anything.

Errors

Every refusal carries the limit, what we measured, and when it resets, so a caller can act without reading a sentence. A refusal to a key also carries buy_credit (the link to this page’s Buy section) and the pack prices.

A callback instead of polling

Send callback_url (a public https address) with the request. The 202 reply then carries a callback object with a secret starting whsec_, shown only there. When the job finishes we POST JSON to the address: { "type", "timestamp", "data" }, where type is transcript.completed or transcript.failed and data is what the GET above returns for the same job, plus its poll_url.

POST https://your-server.example.com/slice
webhook-id: msg_0182d3a8-71b0-4e14-b295-98b7d9c7a6a4_transcript.completed
webhook-timestamp: 1792224000
webhook-signature: v1,K5oZfzN95Z9UVu1EsfQmfVNQhnkZ2pj9o9NDN/H/pI4=
Content-Type: application/json

{
  "type": "transcript.completed",
  "timestamp": "2026-10-17T09:20:00.000Z",
  "data": {
    "id": "0182d3a8-71b0-4e14-b295-98b7d9c7a6a4",
    "status": "completed",
    "file_name": "episode.mp3",
    "duration_seconds": 3619,
    "word_count": 12373,
    "language": "en",
    "text": "Welcome to the show ...",
    "srt": "1\n00:00:00,720 --> ...",
    "vtt": "WEBVTT\n\n...",
    "poll_url": "https://tryslicemedia.com/api/v1/transcribe/0182d3a8-...?token=..."
  }
}

It is signed the Standard Webhooks way, so their libraries verify it as is. Three headers: webhook-id (the same on every retry, so you can skip repeats), webhook-timestamp (unix seconds) and webhook-signature, which is v1, and then the base64 HMAC-SHA256 of id.timestamp.body, keyed with the base64-decoded part of the secret after whsec_. Reject anything with a timestamp more than five minutes old.

// Node: check the signature before trusting the body
const crypto = require('crypto');
function fromSlice(secret, headers, rawBody) {
  const id = headers['webhook-id'], ts = headers['webhook-timestamp'];
  if (Math.abs(Date.now() / 1000 - Number(ts)) > 300) return false;
  const key = Buffer.from(secret.replace(/^whsec_/, ''), 'base64');
  const sig = crypto.createHmac('sha256', key).update(`${id}.${ts}.${rawBody}`).digest('base64');
  return headers['webhook-signature'].split(' ').includes('v1,' + sig);
}

Answer with any 2xx to say you have it. We try three times over about twenty seconds, waiting up to five seconds for an answer each time, and do not follow redirects. If all three fail, the poll_url still works for 30 days. A file we had already transcribed for you comes back completed straight away and is not called back.

Clips, and agents

To have Claude, ChatGPT, Cursor or Claude Code make clips as well as transcripts, connect the Slice MCP server. Slice for AI agents has the setup for each.

If you would rather not use an API

The web tool takes an upload, shows the transcript on screen and emails you the files. Same limits, same engine, same transcripts.