Audio and video to text, free and private
Drop a recording, podcast, voice memo or video. OpenAI’s Whisper speech model runs on your own device and gives you an editable transcript with timestamps.
Public-domain LibriVox recording. Load it, then press Transcribe to try the full tool.
First use downloads the speech model and engine (about 97 MB, of which 72.5 MB is the model); later visits load it from your browser cache. Runs on your CPU; nothing is uploaded. Up to 100 MB and 2 hours per file; under 30 minutes is the comfortable range on a laptop.
To transcribe audio or video for free without uploading it, drop the file above: Whisper base runs inside your browser, detects the language, and returns an editable transcript with timestamps as TXT, SRT or VTT. Accuracy depends on the language: about 9% word errors for English and Spanish in the Whisper paper, far worse for some.
Checked by Koldflux · updated 2026-09-25 · accuracy figures read from the Whisper paper’s tables; tool tested on public-domain recordings
What the free tool runs
| Detail | |
|---|---|
| Model | Whisper base (OpenAI), multilingual, 74M parameters; ONNX export by onnx-community, MIT-licensed weights |
| Download | About 97 MB on first use (72.5 MB model + ONNX Runtime WebAssembly engine), then cached by your browser |
| Languages | 99 language settings, plus automatic detection from the first speech |
| Output | Editable transcript with timestamps; TXT, SRT and VTT downloads; English translation option |
| Files | MP3, M4A, WAV, OGG, Opus, FLAC, AAC, MP4, MOV, WebM, MKV and most other audio or video your browser or ffmpeg can open; up to 100 MB |
| Privacy | Audio is processed in your browser tab and never uploaded |
Accuracy by language
Word error rate on the FLEURS read-speech test in the Whisper paper (lower is better). For Chinese and Japanese it is effectively a character error rate. The free tool runs Whisper base; large-v2 is the largest model the paper measures.
| Language | Whisper base (this tool) | Verdict | Whisper large-v2 | Verdict |
|---|---|---|---|---|
| English | 8.9% | Good | 4.2% | Good |
| Spanish | 9.9% | Good | 3% | Good |
| Portuguese | 13% | Good | 4.3% | Good |
| Italian | 17.9% | Usable, needs a proofread | 4% | Good |
| German | 17.9% | Usable, needs a proofread | 4.5% | Good |
| Russian | 20.5% | Usable, needs a proofread | 5.6% | Good |
| Japanese | 22.8% (CER) | Usable, needs a proofread | 5.3% | Good |
| Korean | 27.8% | Usable, needs a proofread | 14.3% | Good |
| French | 28.5% | Usable, needs a proofread | 8.3% | Good |
| Turkish | 27.5% | Usable, needs a proofread | 8.4% | Good |
| Polish | 30.8% | Rough draft only | 5.4% | Good |
| Dutch | 33% | Rough draft only | 6.7% | Good |
| Chinese | 34.1% (CER) | Rough draft only | 14.7% | Good |
| Indonesian | 33.1% | Rough draft only | 7.1% | Good |
| Vietnamese | 40.5% | Rough draft only | 10.3% | Good |
| Arabic | 48.8% | Rough draft only | 16% | Usable, needs a proofread |
| Ukrainian | 37.7% | Rough draft only | 8.6% | Good |
| Urdu | 52% | Rough draft only | 22.6% | Usable, needs a proofread |
| Tamil | 58.7% | Rough draft only | 17.5% | Usable, needs a proofread |
| Hebrew | 61.7% | Not reliable | 27.1% | Usable, needs a proofread |
| Hindi | 101.1% | Not reliable | 21.5% | Usable, needs a proofread |
| Marathi | 100.3% | Not reliable | 38.3% | Rough draft only |
| Bengali | 100.6% | Not reliable | 104.1% | Not reliable |
| Telugu | 105.2% | Not reliable | 99% | Not reliable |
| Kannada | 107.2% | Not reliable | 37% | Rough draft only |
| Malayalam | 107.4% | Not reliable | 100.7% | Not reliable |
| Gujarati | 101.4% | Not reliable | 102.7% | Not reliable |
How it runs on your device
The page reads the audio track with your browser’s own decoder (ffmpeg.wasm steps in for formats the browser cannot open), mixes it to mono and resamples it to 16 kHz, the input Whisper expects. Long recordings are cut into windows of at most 30 seconds, the length Whisper was trained on, and each cut is placed at the quietest moment between 20 and 30 seconds so words are not split. Silent windows are skipped, because Whisper tends to invent text on silence.
The model is Whisper base, OpenAI’s 74-million-parameter multilingual model, run by Transformers.js in a background worker: a 4-bit encoder that uses your graphics card through WebGPU when available, and an 8-bit decoder on the CPU. If Whisper starts repeating itself, which small models do, that window is decoded again with fallback settings, the same idea as the original Whisper software. The first run downloads about 97 MB (the model is 72.5 MB); after that it loads from your browser’s cache.
How to transcribe a file
- Drop the fileAny audio or video up to 100 MB. It stays on your device.
- Pick the language, or let it detectAuto-detect listens to the first speech. For mixed-language talk or closely related languages, choose it yourself.
- Transcribe and correctText appears window by window. Click a timestamp to hear that part, and fix words in place.
- DownloadTXT for notes and scripts, SRT or VTT for subtitles. Your edits are in every file.
Limits and when to use something else
- Speed depends on your device. A laptop handles a 30-minute file comfortably; phones may run out of memory on long files. Hard limits: 100 MB and 2 hours per file.
- No speaker labels: Whisper writes one stream of text. It also does not mark music, laughter or background sounds reliably.
- Timestamps follow Whisper’s phrases, not individual words, so check sync before publishing.
- Files only: no links from YouTube or TikTok, and no live microphone (microphone access is disabled on this site).
- Several languages are not usable with a browser-sized model: Hindi, Marathi, Bengali, Telugu, Kannada, Malayalam and Gujarati all score 100% or worse with Whisper base in the paper.
Frequently asked questions
Is it really free, with no sign-up?
Yes. The speech model runs on your own device, so there is no server cost per minute: no account, no watermark, no daily limit. Files up to 100 MB and 2 hours.
Is my file uploaded anywhere?
No. The audio is decoded and transcribed inside your browser tab. The only downloads are the model and engine files, once, which your browser then caches.
How accurate is it?
It depends on the language. The Whisper paper measures this model (Whisper base) at 8.9% word errors for English and 9.9% for Spanish on read speech, around 18 to 30% for German, Japanese and French, and 100% or worse for several Indian languages. The table on this page lists each language.
How long does it take?
On one CPU thread of our test server, 30 seconds of English took about 6 seconds. Laptops vary; with WebGPU the heavy part runs on the graphics card. A 30-minute file is comfortable; long files mostly need patience and a tab left open.
Does it label speakers?
No. Whisper produces one stream of text with timestamps. If you need speaker names, add them while editing.
Can I paste a YouTube or TikTok link?
No. The tool works on files on your device. Download your own video first, then drop the file.
Can I record from my microphone?
Not on this site: microphone access is switched off for all Koldflux pages. Record with your phone or computer’s recorder app and drop the file.
Language pages
Each page has the tool preset to that language, measured accuracy, our own test on a real recording and subtitle rules.
Turn one idea into a week of posts
Koldflux Studio (paid, from $19/month) transcribes your upload on our servers with Whisper large-v3-turbo and automatic language detection, then turns one recording into Shorts and carousel posts you review, approve and export. It does not publish or schedule for you.
- 1 · Recording, idea or script
- 2 · Pick the angles
- 3 · Review and approve
- 4 · Export Shorts and carousels
