Podcast transcription

Transcribe a podcast episode in your browser, free: Whisper writes a timestamped transcript on your own computer, as text or as SRT and VTT for Apple Podcasts and your host. Nothing uploaded.

AudioRinse · transcriberOn-device · 0 B uploaded

Drop a recording here

MP3 · WAV · M4A · MP4 · MOV · WEBM — up to 3 hours. Nothing is uploaded.

SpeakersProLanguage

Presets

The model downloads once and stays in this browser, so later transcripts start at once and work with Wi-Fi off. Your recording is never uploaded. Free: Fast transcripts without speaker labels, as text, SRT or VTT, no limit. Pro: Accurate, speaker labels, Word, NVivo and MAXQDA files. Your first Pro download is free with an account.

How it works

Drop the finished episode onto the box above: MP3, WAV, M4A, or the video version. Fast is free; Accurate makes fewer mistakes (Pro).
Choose how many people speak to have each line labelled, or leave labels off. Whisper transcribes on your computer; the episode isn't uploaded.
Read it through: click a line to hear it, fix names in place, and type each speaker's name once.
Download VTT or SRT for Apple Podcasts and your host, text for show notes, or Word (Pro).

Your own transcript in Apple Podcasts

Apple Podcasts writes its own transcripts, but shows yours instead when your RSS feed has one. Most podcast hosts have a transcript upload for each episode that puts the file in the feed; Transistor, for one, takes SRT, VTT, TXT or JSON. Apple reads VTT and SRT. A VTT can also name the speaker of each line, and Apple shows the name whenever the speaker changes: the VTT you download here names them, as WebVTT voice tags, once you've labelled the speakers.

Then, in Apple Podcasts Connect, set the episode to display the transcripts you provide. Apple checks them first: a transcript with spelling mistakes, or one that leaves out part of the episode, isn't shown. Read it through before you upload.

Transcribe the finished episode

Many recording platforms transcribe the raw recording. Once you've edited it in Audacity or another editor and added the intro, the ads and the outro, those times no longer match what listeners hear, and Apple wants the whole episode covered. Export the final mix and transcribe that: every line's time then matches the published episode.

Hosts and guests

Choose the number of people who speak, or "Not sure" to let it count them. Lines come out as Speaker 1, Speaker 2 and so on; type a name once and every line by that person changes.
On three recorded hour-long interviews, fewer than 2 words in 100 went to the wrong speaker. Hosts talking over each other cost the most.
Speaker-labelled downloads, Word files and the Accurate model are Pro. A Fast transcript without names is free as text, SRT or VTT, with no limit.

Getting a good transcript

Pick the language if you know it: detection reads the first 30 seconds, which on many shows is the intro music.
Music under the voice costs accuracy, and a theme song can come out as stray lines. Check the start and the end.
Names of guests, shows and sponsors are spelled the way they sound. Click the line, listen, and fix it in place.
Clean verbatim: most "um" and "uh" are left out, which is what a published transcript usually wants.
A guest on a noisy line? Run their recording through the background noise remover first: cleaner audio means fewer wrong words.

Questions

Is my recording uploaded anywhere?

No. OpenAI's open Whisper model runs inside your browser, on your own computer. The recording and the transcript never leave it; the only download is the model, the first time. The data-flow statement lists every request the page makes.

Is it free?

Yes, with the Fast model: transcripts as text, SRT or VTT, with no limit and no account. Pro ($7.99 a month or $39.99 a year) adds the Accurate model, speaker labels, Word files and videos with burned-in subtitles. The same Pro covers every AudioRinse tool.

Which file does Apple Podcasts want?

VTT or SRT. VTT is the better choice: it can name the speakers, and Apple shows the names. Upload it to the episode in your podcast host, which adds it to your RSS feed; uploading in Apple Podcasts Connect itself is only for subscriber episodes. Apple shows transcripts in English, Spanish, French, German, Italian, Portuguese, Dutch, Danish, Finnish, Norwegian and Swedish.

Can I transcribe a video podcast?

Yes. Drop the MP4 or MOV and its sound is transcribed; the SRT loads into YouTube Studio as subtitles. The video isn't uploaded, however large it is.

Can I transcribe someone else's podcast?

Yes, from the episode's audio file: download it from the show's website or a podcast app, then drop it here. The page can't fetch links, because it never sends anything to a server. A show's words belong to the show, so keep the transcript for your own use.

What else is a transcript good for?

Show notes and episode descriptions, quotes for social posts, timestamps for chapters in a YouTube description, captions for the video version or for short clips (with auto captions), and a searchable record of everything said on the show.

How accurate is it?

Accurate (Whisper large-v3-turbo) got 8.3% of words wrong on three recorded hour-long interviews with professional reference transcripts, on a par with desktop Whisper tools on the same files. Fast (Whisper small) makes more mistakes, mostly on names, accents and noisy recordings. Clear speech close to the microphone helps both.

How long does it take?

Fast runs on your processor at about twice real time: a 30-minute recording takes about 15 minutes on a recent laptop. Accurate runs on the graphics chip through WebGPU, about as long as the recording on a 2020 laptop's built-in graphics and faster on newer ones. The first time also downloads the model.

Which languages does it understand?

Whisper transcribes around 100 languages, most accurately English and the major European languages. Pick the language in the menu, or let it detect it from the first 30 seconds.

How long can the recording be?

Up to 3 hours, audio or video. On a computer with little memory the page takes less, and says so.