Guide
Transcribe short English recordings with Whisper in your browser, or turn SRT, VTT and pasted captions into clean, downloadable text.
Choose the fastest source you already have
If your video editor or caption provider has already produced SRT or VTT, import that file instead of transcribing the audio again. The cleaner removes cue numbers, display settings and caption markup while preserving the spoken text. You can export a continuous paragraph, line-by-line text or timestamped lines. A selected subtitle file takes precedence over the paste box. Plain text is also accepted, but it cannot produce meaningful subtitle timing. The tool will never invent timings for a text-only source.
Run Whisper locally with realistic expectations
Whisper tiny.en is an English speech-recognition model. The first transcription downloads model files and a runtime, then processes decoded audio on your device using WebAssembly. Short, clear clips are the best starting point. This interface limits files to 60 MB and recordings to ten minutes to reduce memory pressure. A fast desktop may finish much sooner than a phone. Browser support for video audio decoding varies; if MP4 or another container fails, export a WAV or MP3 track from your editor. Cancel stops the worker, though cached model downloads can remain in browser storage.
Review words before reusing them
A transcript is a draft, especially when a recording contains music, accents, overlapping voices or silence. Check names, technical terms and numbers against the original media. Whisper can generate plausible words during non-speech sections, so delete anything you cannot hear. Overlap cleanup only removes repeated prefix words when neighboring cues overlap in time; it does not try to rewrite grammar. Edit the text in the result before downloading the edited version. SRT and VTT exports keep the cleaned cue text and original or machine-estimated timestamps, independent of free-form text edits.
How the result is made
Imports use a local SRT/VTT parser with explicit timing validation. Audio is decoded and resampled to 16 kHz mono, then transcribed in a worker using Transformers.js 3.7.3 and Xenova/whisper-tiny.en at a pinned model revision. The model is based on OpenAI Whisper; this site is independent. Review machine output for errors.