AI Transcription

AI transcription with Whisper: audio and video to text, free

Upload an audio or video file and SpeakScribe transcribes it with Whisper large-v3-turbo, OpenAI's open speech recognition model. You get a clean transcript, a timestamped version and SRT and VTT subtitles, and you can translate the result into 46 languages. Try 20 minutes free without an account; a free account gives 120 minutes a month.

New uploads are paused until it finishes, so the service stays fast for everyone. You can leave this page — the result will be waiting in your queue.

Open my queue →

MP3, M4A, WAV, FLAC, OGG, MP4 · up to 20 MB as a guest

Transcribing your audio…

Need it in English? Transcribe and translate in one go →

Free
20 min, no account
120 min a month with a free account · files up to 30 min
Model
Whisper large-v3-turbo
99.6% of 2,140 files in 1–30 October 2026; backup engine for the rest
Speed
2 s for a 5–15 minutes file
Median after upload, real jobs
Output
TXT · SRT · VTT
Clean and timestamped text, subtitles · MP3, M4A, WAV, FLAC, OGG, WMA, MP4, WEBM in

What happens to your file

  1. 1 · Check

    Your browser checks the file's size and, where it can, its length before uploading. If it is over your plan's limit you are told right away, before anything is uploaded.

  2. 2 · Prepare

    On the server the real duration is read from the file — that is what counts toward your minutes. Large files are re-encoded to compact speech audio first.

  3. 3 · Recognise

    Whisper large-v3-turbo converts speech to text in segments, each with a start and end time. If that service is busy, a backup engine finishes the job instead.

  4. 4 · Deliver

    You get a clean transcript, a timestamped one and subtitle files (SRT and VTT), ready to download, copy or send to the next step below.

What you get back

Every job returns all of these — take whichever fits the job.

Clean transcript for reading and editing
Thanks everyone for joining. Let's start with last month's numbers. I'll share my screen so you can follow along. The headline is that renewals came in ahead of forecast for the third quarter running.
Timestamped for subtitles and citation
[00:00 → 00:04] Thanks everyone for joining.
[00:04 → 00:09] Let's start with last month's numbers.
[00:09 → 00:14] I'll share my screen so you can follow along.
Subtitles .srt and .vtt
1
00:00:00,000 --> 00:00:04,000
Thanks everyone for joining.
2
00:00:04,000 --> 00:00:09,000
Let's start with last month's numbers.

An illustration of the output formats, not the result of a particular recording. Subtitle files are ready for YouTube and video editors.

How accurate is it? Tested in 40 languages

Native speakers reading the same news-style sentences (Google's FLEURS dataset), 800 recordings in total, transcribed with the same setup this site uses on 30 October 2026 and checked against what was said. How much editing to expect:

English Quick read-through
Spanish Quick read-through
Arabic Quick read-through
Persian Light editing
Urdu Moderate editing
Hindi Moderate editing
Bengali Heavy editing
Indonesian Quick read-through
Malay Quick read-through
Filipino Quick read-through
Turkish Quick read-through
Russian Quick read-through

Of 40 languages measured: 20 quick read-through, 2 light editing, 5 moderate editing, 5 heavy editing, 8 not usable yet. See every language →

How long it takes

Time from a finished upload to a ready transcript, across 2,140 real files in 1–30 October 2026. Uploading a large file over a slow connection usually takes longer than transcribing it.

Recording lengthFilesMedian wait9 in 10 done within
Under 5 minutes1,473 1 s2 s
5–15 minutes545 2 s3 s
15–30 minutes119 4 s27 s

Only 3 files were longer than 30 minutes in this period, too few to report.

Formats and limits

Supported files: MP3, M4A, WAV, FLAC, OGG, WMA, MP4, WEBM. MP4 video is accepted directly — the audio track is used. Other video formats (MOV, MKV, WEBM) need converting to MP4 or MP3 first.

PlanPriceMinutes per monthLongest fileResults kept
Guest (no account) $0 20 in total 10 min —
Free $0 120 30 min 7 days
Starter $4.99/month 300 60 min 30 days
Pro $9.99/month 1,200 90 min 90 days
Studio $19.99/month 2,000 120 min 180 days
Agency $39.99/month 3,000 150 min 365 days

Guest minutes come out of the Free allowance — creating a free account unlocks the rest. Full plan details are on the pricing page.

Where your recording comes from — and how to upload it

Most apps already save audio in a format SpeakScribe reads. Save or export the file to your phone or computer, then upload it — on a phone, straight from the browser.

SourceWhat to upload
WhatsApp voice noteSave it from the chat; it is an .opus file — rename the ending to .ogg and upload. Step by step
Telegram voice messageSave to downloads; Telegram voice files are .ogg and upload as they are.
Signal and other messengersSave the attachment. .m4a uploads directly; if it ends in .aac, convert it to MP3 or M4A first.
Zoom, Google Meet, Microsoft TeamsDownload the meeting recording (MP4, or M4A audio from Zoom) and upload it. Meeting guide
iPhone Voice Memos, Samsung Voice Recorder, Google RecorderShare or export the recording as M4A (or WAV) and upload. Voice memo guide
VoicemailUse your phone's "save" or "share" option on the message to get an audio file, then upload it.
Screen recordings, YouTube, TikTok, Instagram, TwitchUpload the video file (MP4) — links are not accepted. For your own YouTube videos, download the original from YouTube Studio.
iPhone video (MOV), WEBM, MKVConvert to MP4 or MP3 first.
Cassette tapes and old recordingsDigitise them to MP3 or WAV first; older tapes with hiss give a rougher transcript.

When a transcript looks wrong

The same sentence repeats, or text appears where nobody spoke

Whisper can "fill" long silences or music with repeated or invented lines. Cut music-only and silent stretches out of the file and transcribe again.

It came out in the wrong language or script

Auto-detect listens to the first seconds. Choose the language yourself — for example Hindi rather than auto-detect, so Hindi is not written in Urdu script, and Nepali so it is not written as Hindi.

Whole words are missing

Usually quiet speech, people talking over each other, or a TV or wind in the background. Record closer to the speaker; there is no setting that recovers speech the microphone did not catch clearly.

Names, numbers and jargon are wrong

These are the hardest words in every language. Search the transcript for each name once and replace it everywhere; check every number against the audio.

What decides how accurate it is

  • The recording. One speaker close to a microphone is the best case. Room echo, phone-call audio and wind noise cost the most accuracy.
  • Music and overlap. Background music under speech, and people talking over each other, are the most common causes of missing or merged words.
  • The language. In our test Spanish needed only a quick read-through, while Bengali needed heavy editing — and some less widely spoken languages are not usable yet.
  • Names and numbers. Product names, people's names and numbers are the words most likely to need fixing, in any language.

How to get a cleaner transcript

  • Choose the language instead of auto-detect when you know it. Detection listens to the opening of the file, so an intro song or a greeting in another language can fool it.
  • Trim long silences and music-only parts. Speech models sometimes write stray or repeated lines over long stretches without speech.
  • Upload the original file, not a re-recorded or heavily compressed copy.
  • Split very long recordings at natural breaks if they exceed your plan's file length — each part is transcribed separately.

After the transcript

In the transcriber, a finished transcript can go through one more AI step — pick any combination:

Fix spelling and punctuation

Cleans up punctuation, capitalisation and diacritics so the text reads like it was typed, not dictated.

Translate the transcript

Into one of 46 languages. The audio translator does both steps at once and shows measured English quality per language.

Study notes

Turn a lecture or class recording into notes: a summary, key points, an outline, key terms and review questions.

A publish package

Title ideas, a description and tags for YouTube, written from what you actually said. See YouTube SEO from transcripts.

Translate a recording to English: Hindi, Urdu, Bengali, Tamil, Telugu, Nepali, Arabic, Persian, Spanish.

Need captions? Every job also includes SRT and VTT files — see the subtitle generator guide. Going the other way — from a script to a voiceover — is text to speech, with 361 voices.

What SpeakScribe does not do (yet)

So you can decide quickly whether it fits your job:

  • —No speaker names. Lines are timestamped but not labelled by speaker.
  • —No live transcription or recording in the browser — upload a finished file.
  • —No links. Upload the file instead of pasting a YouTube or social media link.
  • —No Word, CSV or JSON export. You get TXT, timestamped text, SRT and VTT.
  • —No online editor. Edit the downloaded text in any editor.
  • —No translated subtitle files. Translations come as text.
  • —No offline app. It runs in the browser and needs an internet connection.
  • —Monthly or yearly plans only — no one-time purchase; the Free plan has no time limit.

Guides for your recording

Detailed advice for your language, file type and kind of recording — each with a working transcriber.

AI transcription questions

Yes. The Free plan has no end date and needs no credit card: 120 minutes of audio every month, files up to 30 minutes, and the minutes start again each month. You can also try 20 minutes without creating an account. Paid plans only add more minutes and longer files — the transcription itself is the same.
No. Free transcripts are complete, plain text with no watermark, and the SRT and VTT subtitle files are the same ones paid plans get. The only differences between plans are the monthly minutes and the longest file.
No. Upload a file and you get up to 20 minutes as a guest, no email needed. A free account adds the rest of the monthly allowance and keeps your transcripts in your history.
Yes. Files are transcribed with Whisper large-v3-turbo, OpenAI's open speech recognition model, running on fast cloud hardware — nothing to install. In 1–30 October 2026, 99.6% of files were transcribed by it; when that service is busy, a backup engine finishes the job.
We tested it on native speakers reading the same sentences in 40 languages and checked every transcript against what was said. How much editing to expect: English — quick read-through; Spanish — quick read-through; Arabic — quick read-through; Hindi — moderate editing. Noise, music, people talking over each other and phone audio make any language harder — the audio to text page rates every language we tested.
Seconds, once the file is uploaded. Across 2,140 real jobs in 1–30 October 2026, a 5–15 minutes file was ready in a median of 2 seconds; uploading a large file usually takes longer than transcribing it.
Yes, in the phone's browser — Android or iPhone, no app to download. Pick a voice note or recording from your files and upload it. You can also add SpeakScribe to your home screen so it opens like an app.
Yes. Upload it as MP4 and the audio track is transcribed. Other video formats (MOV, MKV, WEBM) need converting to MP4 or MP3 first; a WhatsApp voice note (.opus) only needs its ending renamed to .ogg.
No — SpeakScribe transcribes files, not links. For your own videos, download the original from YouTube Studio and upload it as MP4; you also get SRT subtitles you can add back to YouTube.
Not yet. You get the full text and a timestamped version with one line per spoken segment, but speakers are not named. For interviews, add the names while you review the timestamped file.
You download plain text (.txt), a timestamped text and SRT/VTT subtitles; there is no Word export or online editor yet. Copy the text into Word or Google Docs to edit and format it, and fix subtitle lines directly in the .srt file before you upload it to a video platform.
30 minutes per file on the Free plan and up to 150 minutes on paid plans; guests can upload files up to 10 minutes. Split longer recordings at natural breaks and transcribe the parts one after another.
Partly. Each file is transcribed as one language, so choose the main one — English words inside a Hindi, Urdu or Arabic recording are usually kept, but long passages in a second language come out weaker. If a recording switches languages for long stretches, transcribe those parts separately.
Pick it when you know it. Whisper guesses the language from the opening of the recording, so a file that starts with music, silence or a greeting in another language can be detected wrongly.
Yes — into English or 45 other languages, right after transcription. The audio translator page does both steps in one go, and the translation step does not use extra minutes.
Yes. After transcribing, choose Study notes: you get a summary, key points, an outline, the key terms the speaker explained and review questions, written only from what was said in the recording.
No. SpeakScribe does not train any model on your files or transcripts. A file is sent to the speech-recognition service only to produce your transcript, and it is stored privately for your account.
Files are stored privately and download links expire after an hour. Results are deleted automatically after 7 days on the Free plan, and after 30 days on Starter, 90 days on Pro, 180 days on Studio and 365 days on Agency. To have a file removed sooner, email support from the address on your account.

Try it on your own recording

Upload a file and compare the transcript with what was said — guests get 20 free minutes.

Open the transcriber