Audio to Text

There is an HTTP API for exactly this: POST a media URL, get text, SRT and VTT back as JSON. Same limits, same engine, same transcripts — no upload, no email step, and a free key in one request. Without a key you still get one short file a day, so a first try costs nothing at all.

Most free transcribers stop somewhere between ten and thirty minutes. That number is not chosen around what people actually record; it is chosen around what is cheap to give away. So we measured our own work instead.

Across 96 transcripts made between February and September 2026 — our own customers’ podcast episodes, interviews and recorded calls, measured on 16 September 2026:

• Median length: 47.3 minutes. Half of everything we transcribe is longer than that.
• 96% were 90 minutes or under. The longest was 93.6 minutes.
• Only 17% were 15 minutes or under. A fifteen-minute free tier would have been useless for five files in six.
• Average 7,775 words, at about 172 words a minute. 94% were audio rather than video.

Ninety minutes is where those two facts meet: it covers the overwhelming majority of real episodes, and it is a length we can afford to give away. These are aggregate figures from our own account — no customer file, name or content is in them.

How this compares

We read the free tiers of six other transcription tools off their own pricing pages and put them in one table — longest free file, whether they label speakers, whether you need an account, and what happens when the free part runs out. It includes where we lose, because a comparison that does not is an advertisement.

Who is speaking

Most transcription tools put speaker labels behind the paywall, and we did too until September 2026. It was the wrong line to draw. Most of what people send this tool is an interview or a two-hander, and a conversation transcribed without turns is not a cheaper document — it is a worse one, because the reader ends up doing the work in their head.

So the free tool diarises. Each turn is labelled in the text and in the subtitle files, the label appears where the speaker changes rather than on every line, and a file with one voice in it comes out with no labels at all. What separates free from paid here is the model, the ninety-minute cap and the 2 files a day — not whether the result is readable.

What it costs us to run

Transcription is about 15 cents an hour of audio, which we pay. That is why there are limits, and why we would rather print them than let you find them: 2 files a day per person, ninety minutes each, and a monthly budget for the tool as a whole. If a month runs out, the tool says so and offers to email you when it resets. It does not quietly get worse or start asking for a card.

What happens to your file

It is uploaded to private storage that is not readable from the web, passed to AssemblyAI to be transcribed, and deleted after 7 days. The text is kept for 30 days so the download links in your email keep working, then that is deleted as well. Nothing is used to train anything.

Where the free tool stops

It gives you the words and who said them. It does not cut clips, burn captions onto video, or find the good bits — those are what Slice’s paid tools do, and pretending otherwise would waste your time. It is also tuned for English: it will transcribe other languages and tell you which one it detected, but we are not going to promise the accuracy.