spaghettiwires calc

Audio & video transcription

Upload a recording (a meeting, lecture, interview, podcast or video) and get the text back, split by speaker with a timestamp on every turn. If you like, the text is also cleaned up and split into sections, with a summary on top. You only pay for what you use, about €0.12 an hour.

You can use them in Claude too: connect it to your account.

[[ t.off_text ]]

[[ t.gate_title ]]

[[ t.gate_text ]]

[[ t.upload_title ]]

[[ t.duration ]]
[[ measuring ? '…' : fileDuration ? duration(fileDuration) : t.duration_unknown ]]
[[ t.language_hint ]]
[[ phaseText ]]

[[ t.keep_open ]]

[[ t.credit_title ]]

[[ credit ? euro(credit.balance) : '…' ]]

[[ t.list_title ]]

[[ fmt(t.list_range, { from: list.offset + 1, to: list.offset + list.items.length, total: list.total }) ]]

[[ t.loading ]]

[[ t.list_empty ]]

[[ viewLoading ? t.loading : t.viewer_empty ]]

How it works

From file to text

The file travels from your browser straight to our Amazon S3 storage, in 64 MB pieces (up to 5 GB). There we pull out just the audio track, measure how long it is and hand it to a speech recognition service that also tells voices apart. The original file is deleted as soon as the job ends.

Speakers and timestamps

Each voice gets a letter (A, B, C…) in the order it first speaks, up to 15 voices. We don't know names, so replace the letters yourself if you need to. Telling voices apart is accurate for recordings up to 4 hours 50 minutes, the most that can be transcribed in a single pass. Beyond that the recording is cut into equal parts and the letters restart from A in each one, so the same person may get a different letter before and after a cut; the text marks where that happens. With clean-up on, the cleaned text tries to work out who is speaking from what they say and to keep labels consistent across parts, but it can get it wrong: in recordings that long, double-check who says what around the cuts.

Cleaning up the text

With "Clean up the text" ticked, once the transcript is ready a language model works through it in chunks. It splits the text into numbered sections and topics, drops fillers, false starts and repetitions, mends broken sentences and fixes recurring transcription mistakes such as names and acronyms. Speakers get a label describing their role, and the top of the document carries a summary, a participants table and editorial notes. Nothing is added: every sentence can be traced back to whoever said it, and the timestamps still point into the recording. Here too you pay for the tokens used (€0.25 per million in, €1.25 per million out): usually about €0.02 more per hour, so roughly €0.14 an hour in total. Only the cleaned text is kept. If the clean-up fails you get the raw transcript and pay for the transcription alone.

Price and credit

You pay for the tokens the transcription actually uses, not a flat rate per minute. Audio is counted by length (about 30,000 tokens an hour, pauses included) and the transcribed text on top, so an hour of mostly silence costs less than an hour of non-stop talk. The rates are €1.70 per million audio tokens and €4.00 per million text tokens, which for an ordinary conversation comes to about €0.12 an hour; the real bill depends on how much is said. Before starting we measure the length and set aside the most it could cost (about €0.15 an hour): if your credit is short, the job doesn't start and you pay nothing. At the end you pay for the tokens actually used and get the rest back; if something goes wrong you get it all back.

How long we keep it

The text is saved as Markdown and kept for 6 months from the day it was transcribed; the list shows each one's deletion date. You can download it, as Markdown or PDF, or delete it at any time.

From Claude

Connect Claude, or any other assistant that speaks MCP, and your transcripts become context for a conversation: it lists them, reads them a part at a time and searches them, without changing anything or spending credit. How to connect it.

New here?