KI-Untertitel-Generator

Whisper AI transcription — free, local, private

OpenAI Whisper is one of the best automatic speech recognition models ever released. It understands 99 languages, handles accents gracefully, and produces subtitle-quality output including punctuation and casing. Services that offer Whisper as an API charge $0.006 per minute — for a 60-minute interview, that's $0.36, and it goes to their server. This tool runs Whisper directly in your browser via WebAssembly, so it costs you nothing and your audio never leaves your device.

How it works

1. Drop your file. Any format your browser can decode: MP3, MP4, WAV, M4A, OGG, WebM, FLAC, MOV.
2. Load a model. We download the Whisper model weights from HuggingFace CDN (75–466 MB). This is a one-time download — the browser caches it, so subsequent runs work instantly, even offline.
3. Transcribe. The model runs in a Web Worker thread so the page stays responsive. A typical 5-minute audio clip takes 1–3 minutes to transcribe, depending on your CPU.
4. Export. Download as SRT (for video editors and social platforms), VTT (for web players), plain text, or JSON with raw timestamps.

Which model to choose?

• Tiny (75 MB) — quick drafts, strong English. Useful for getting timestamps even if accuracy isn't perfect.
• Tiny EN-only (75 MB) — same size as Tiny but faster on English because it skips the language-detection step.
• Base (145 MB) — the sweet spot for most use cases. Better multilingual accuracy than Tiny at acceptable speed.
• Small (466 MB) — near-professional quality. Recommended for interviews, podcasts, legal/medical content.

Export formats

• SRT — supported by YouTube, Vimeo, CapCut, DaVinci Resolve, Premiere, Final Cut Pro and every social platform that accepts caption files.
• VTT (WebVTT) — the native caption format for HTML5 <video> elements and web streaming players.
• Nur Text — the transcript without any timing markers. Useful for documentation, meeting notes or full-text search.
• JSON — raw transcript data with start and end timestamps per segment. Ideal if you want to process the transcript programmatically.

Privacy

Your audio file is decoded and processed in your browser tab using WebAssembly. Nothing is uploaded. We don't store audio, we don't store transcripts, we don't log file names. The only network requests this page makes are the one-time model downloads from HuggingFace. After that, the model runs completely offline.

Common use cases

• Subtitle a podcast episode or YouTube video without paying for Rev or Otter.ai.
• Transcribe an interview for a journalism piece or research paper.
• Generate captions for accessibility compliance.
• Create timed subtitles for a Shorts/Reel from a longer video.
• Transcribe meeting recordings for written notes.
• Create an SRT file to upload to TikTok, YouTube or Instagram as native captions.

Limitations

• First load is slow — 75–466 MB model download. Fine on broadband, slow on mobile data.
• CPU-bound — transcription happens on your CPU. A modern laptop transcribes ~10× real time with the Tiny model. Older hardware will be slower.
• Background noise — like all Whisper deployments, quality degrades significantly with heavy background noise or multiple speakers talking over each other.
• Long files — files over 30 minutes should be split first. Very long files may exhaust browser memory.