From the upload to the export: what you get back, how it is edited and searched, which engines run it, and where your files go. Prefer to just try it? Transcribe a file free.
Last updated:
Same product, two places to run it. Pick by how sensitive the recording is.
The convenient option: nothing to install. Your file is uploaded to the Scriptorium server and transcribed there, under the hosted privacy policy. Best for non-sensitive recordings.
One file up to 15 minutes and 25 MB, 3 tries a day, with no account. Preview files are deleted after 24 hours.
Try it freeRecordings, transcripts and search stay on infrastructure you control. Best for recordings under legal privilege, medical confidentiality or internal security policy.
How self-hosting works
00 · PRIVATE
Private meeting transcription. Run TryScriptorium on infrastructure you control and the recording never leaves it — no third-party speech API, no copy on somebody else's disk. The hosted workspace on this site is the other option, and it says so plainly before you upload anything.
01 · RECORD
Drop in what you already have. Video or audio, phone recording or conference export — the container is unwrapped, the audio track is pulled out, and the picture is discarded. A 4K file and a voice memo of the same meeting cost the same.
02 · TRANSCRIBE
The engine writes out every line with a start and an end time, and separates the turns by voice when more than one person is speaking. Nothing is summarised away: what you get back is the meeting, in full, with the clock still attached.
03 · UNDERSTAND
Click any line to correct it. Rename a speaker once and the change runs through the whole transcript. It behaves like a document you own rather than a machine output you have to accept.
04 · SEARCH
Search runs across every transcript in your library, and it searches your corrections rather than the engine's first guess. Each result keeps its timestamp, so the sentence you half-remember takes you back to the second it was said.
05 · YOUR KNOWLEDGE
One meeting becomes a searchable record. A year of them becomes the thing nobody wrote down. It leaves in the shape you need — plain text, timestamped text, SRT, VTT, Markdown, JSON with word timings, or CSV.
Free for one file up to 15 minutes, no account. No card. No installation.
The transcription engine runs beside the files it processes. In a self-hosted deployment, audio, transcripts, search and speaker analysis remain on infrastructure you control. The hosted preview on this site is different: uploads are processed on this server and are covered by the hosted privacy policy.
Use the hosted workspace for a quick, non-sensitive transcription. Self-host for recordings governed by legal privilege, medical confidentiality or internal security policy.
Explore self-hosted transcription
A scriptorium was originally a room where a manuscript was copied by hand and corrected in the margin — a word struck through, a replacement written above it. Every line in your transcript can be corrected the same way; it just takes a click instead of a quill, and the correction is what gets searched from then on.
Sample transcript · nothing is uploaded
This is how a finished transcript behaves. Search it, and matches highlight as you type. Hover or select a line to see where it sits in the recording.
Speaker labels
When more than one voice is in a recording, the turns are separated and labelled, so an interview comes back as an exchange instead of one undifferentiated block. Up to eight speakers.
one recording
Labels start generic and you rename them once each. This is worked out from the sound of each voice, so it is reliable at hearing that the speaker changed and imperfect at deciding exactly how many people there are — similar voices and heavy crosstalk are where it struggles.
Word-level timing
Not just every line — every word carries a start and an end. That is what makes a transcript you can jump around in, and it is in the JSON export in full.
00:15 · Yusuf
Shall I share the numbers from last quarter?
transcript.json
Editing
Automatic transcription gets names and jargon wrong. Click the line, correct it, and the correction is re-indexed — so searching later finds what you wrote, not the mistake it replaced.
before
Shall I share the nummers from last quarter?
after
Shall I share the numbers from last quarter?
Corrections are saved against the segment and the search index is rebuilt for it, so the transcript and what you can find in it never drift apart.
Search
Search runs across your whole library rather than the file you happen to have open — which is the difference between a transcript and an archive you can actually use.
Board meeting
18 Sep
Q3 planning call
02 Sep
Investor update
27 Aug
Team retro
14 Aug
Pick the language, or let it work the language out for itself. Speech in any of these can also be translated into English text. See every language, with notes on each.
English, Arabic, Spanish, French, German, Portuguese, Italian, Dutch, Turkish, Russian, Ukrainian, Polish, Hindi, Urdu, Bengali, Chinese, Japanese, Korean, Vietnamese, Thai, Indonesian, Malay, Tagalog, Persian, Hebrew, Swedish, Danish, Norwegian, Finnish, Greek, Czech, Romanian, Hungarian
Exports
A transcript is only useful in the shape the next tool expects. Every one of these is produced by the same exporter that writes your download — this is its real output.
Amina: Let's get started — can everyone hear me okay?
Yusuf: Yes, loud and clear. Shall I share the numbers from last quarter?
Amina: Go ahead. I want to look at retention first.[00:03] Amina: Let's get started — can everyone hear me okay?
[00:11] Yusuf: Yes, loud and clear.
[00:15] Yusuf: Shall I share the numbers from last quarter?
[00:22] Amina: Go ahead. I want to look at retention first.1
00:00:03,200 --> 00:00:06,900
Amina: Let's get started — can everyone hear me okay?
2
00:00:11,000 --> 00:00:13,400
Yusuf: Yes, loud and clear.
3
00:00:15,100 --> 00:00:18,800
Yusuf: Shall I share the numbers from last quarter?
4
00:00:22,000 --> 00:00:25,600
Amina: Go ahead. I want to look at retention first.WEBVTT
00:00:03.200 --> 00:00:06.900
<v Amina>Let's get started — can everyone hear me okay?
00:00:11.000 --> 00:00:13.400
<v Yusuf>Yes, loud and clear.
00:00:15.100 --> 00:00:18.800
<v Yusuf>Shall I share the numbers from last quarter?
00:00:22.000 --> 00:00:25.600
<v Amina>Go ahead. I want to look at retention first.# Board meeting
**Duration:** 12:04 · **Language:** en · **Model:** small · **Words:** 1,180
_Source: `board-meeting.mp4`_
## Transcript
Amina: Let's get started — can everyone hear me okay?
Yusuf: Yes, loud and clear. Shall I share the numbers from last quarter?
Amina: Go ahead. I want to look at retention first.{
"title": "Board meeting",
"source": "board-meeting.mp4",
"language": "en",
"model": "small",
"duration": 724.0,
"created_at": "2026-09-18T10:04:00Z",
"speakers": {
"1": "Amina",
"2": "Yusuf"
},
"segments": [
{
"idx": 0,
"start": 3.2,
"end": 6.9,
"text": "Let's get started — can everyone hear me okay?",
"speaker": "1",
"highlight": false,
"words": [
{
"w": "Let's",
…index,start,end,start_clock,speaker,highlight,text
0,3.200,6.900,00:03,Amina,0,Let's get started — can everyone hear me okay?
1,11.000,13.400,00:11,Yusuf,0,"Yes, loud and clear."
2,15.100,18.800,00:15,Yusuf,0,Shall I share the numbers from last quarter?
3,22.000,25.600,00:22,Amina,0,Go ahead. I want to look at retention first.Engines
Six speech models, from one that keeps up with a long recording to one that takes its time and gets more of it right. The bar is download size; the note beside each is the trade it makes.
tiny
75 MB
Fastest. Rough draft quality.
base
145 MB
Fast. Fine for clear speech.
small
free tier
480 MB
Recommended. Good accuracy at ~3x realtime.
medium
1.5 GB
More accurate, about 3x slower than small.
large-v3-turbo
1.6 GB
Near best accuracy, much faster than large.
large-v3
3 GB
Best accuracy. Slow on CPU.
Accuracy
We do not publish an accuracy percentage, because the honest answer depends entirely on your audio. What we can tell you is what costs you accuracy, roughly in the order it does the most damage — and that none of it needs equipment to fix.
Two people talking at once
Overlapping speech is the hardest case there is, and no tool solves it well.
Distance from the microphone
A phone across the room picks up more room than voice.
Echo
Hard walls and bare floors smear every word into the next.
Background noise
Cafés, traffic and air conditioning all eat consonants.
Specialist vocabulary
Names, drugs and case numbers are guessed from context, and context is often not enough.
An ordering, not a measurement. Moving the microphone closer and asking people not to talk over each other will do more than any setting in any product.
If a media player can open it, it can almost certainly be transcribed.
.txtPlain text.txtTimestamped text.srtSubRip subtitles.vttWebVTT captions.mdMarkdown report.jsonJSON with word timings.csvOne row per segmentWhile you wait
There is one transcription worker, shared by everyone, so a busy moment means a wait. The page tells you which of the two you are in — and your position in line — rather than showing a spinner that implies work that has not started.
Queued
Waiting for the worker. You are shown your position in line.
Transcribing
Running now, with progress through the recording.
Done
The transcript opens, editable and exportable.
Transcription runs faster than real time, so a recording is generally done in a fraction of its own length — but it is a queue, and at a busy moment you wait behind other jobs.
Retention
Said plainly here, in full on the privacy page. Deleting a transcript removes the upload, the extracted audio and the database rows in one go — there is no copy left behind.
Uploaded
The file arrives and is queued. It is not sold or shared with third parties.
Transcribed
The transcript is yours. Every account is isolated from other users; staff can open a file only with a special permission, and every access is logged.
Free preview: 24h
A transcript made without an account has nobody to own it, so it and its files are deleted automatically after 24 hours.
Deleted when you say
Deleting removes the upload, the audio proxy and the database rows together.
Deployment
Scriptorium is one program that can run in two places, and which one you want depends entirely on the recording.
Hosted — this site
Upload and get a transcript in a minute with nothing to install. The file is processed on this server under the hosted privacy policy. Right for a quick, non-sensitive recording.
Read the hosted privacy policy →Self-hosted — your hardware
The engine runs beside the files it processes, on infrastructure you control. Right for recordings governed by legal privilege, medical confidentiality or an internal policy that does not allow a third-party server.
Explore self-hosted transcription →Accounts
The preview exists so you can see a real transcript of your own file before deciding anything. An account — also free, also no card — is what makes it somewhere to keep them.
The preview
Free
no account needed
A free account
Also free
no card
Paid plans add a larger monthly allowance and priority in the queue — see pricing.
Every transcription tool has edges. Ours are listed here rather than discovered after you have uploaded something.
Your upload goes to the Scriptorium server, is transcribed there, and stays with your account until you delete it. Every account is isolated from other users. Staff can open a file only with a special permission, and every access is logged.
Files from the free preview, where no account exists to own them, are deleted automatically after 24 hours.
Deleting a transcript removes it from your library immediately. Read the privacy page for the full detail.
Simple tools for audio, video, images and documents. The image and PDF tools run in your browser where they can; see all tools.
No account, no card. See what the transcript looks like before you decide anything.