Loading tool…
Typing up an hour of conversation by hand takes most people four to six hours. Drop in a recording and Whisper turns it into an editable, timestamped transcript on your own computer, so the audio never touches a server.
Processing locally on your device
Overview
Whisper is OpenAI's open speech recognition model, published under the Apache-2.0 licence. It turns audio into a log-Mel spectrogram, reads it with a transformer encoder and writes text token by token with a decoder. Because it was trained on hundreds of thousands of hours of varied, multilingual audio, it copes well with accents, background noise and technical vocabulary, and it recognises 99 languages. This tool runs the ONNX versions of Whisper Tiny (39 million parameters), Base (74 million) and Small (244 million) through Transformers.js.
Your file is decoded by the browser, mixed to mono and resampled to 16 kHz, which is what Whisper expects. Long recordings are cut into 30-second windows that overlap by 5 seconds on each side, so words at a boundary aren't lost, and the pieces are stitched back together using Whisper's own timestamp tokens. If you leave the language on automatic, the tool asks the model which language it hears in the loudest of the first few windows and reports its confidence. Word-level timings come from the model's cross-attention, aligned with dynamic time warping, so you can click any word to play from that point.
Speed depends on your hardware. With a WebGPU graphics card the encoder runs on the GPU in 16- or 32-bit precision with a 4-bit decoder; Tiny is about 100–120 MB, Base 165–205 MB and Small 410–590 MB to download. Without WebGPU, Tiny (41 MB) and Base (77 MB) run on the CPU with 8-bit weights in WebAssembly. Small is only offered with WebGPU because it would be too slow and memory-hungry on a single CPU thread. On a typical laptop CPU, Base processes speech a little faster than real time; on a GPU it is several times faster. The model is downloaded once and cached, after which the tool also works offline.
Whisper can also translate: switch on Translate to English and it writes English text for speech in any supported language. Accuracy is very good for clear speech in widely spoken languages and drops for crosstalk, heavy music, rare languages and very poor recordings. Always proofread transcripts you plan to quote or publish; the editor makes that quick, because every line is linked to its place in the audio.
| Model | Parameters | CPU download (8-bit) | WebGPU download | Best for |
|---|---|---|---|---|
| Tiny | 39 M | 41 MB | ~100–120 MB | Quick drafts, clear speech |
| Base | 74 M | 77 MB | ~165–205 MB | Everyday transcription |
| Small | 244 M | Not offered | ~410–590 MB | Hard audio, more languages |
Drop in an audio or video file (MP3, WAV, M4A, OGG, FLAC, MP4, MOV, WebM).
Pick the Whisper size and the spoken language, or leave language detection on.
Optionally turn on Translate to English, word-level timings or a vocabulary hint.
Click Transcribe and follow the progress window by window.
Click timestamps or words to check the audio and edit any line that needs it.
Download TXT, SRT, VTT or JSON, or copy the text.
Use cases
Journalists and researchers can transcribe recorded interviews without sending confidential conversations to a cloud service, then click through the timestamps to check quotes before using them.
Turn a recorded lecture, webinar or team call into searchable notes. The paragraph-style TXT export reads cleanly, and the timestamped version makes it easy to find the moment something was said.
Export SRT or VTT and upload it to YouTube, Vimeo or your video player. With word timings on, long sentences are split into readable two-line cues automatically.
Use Translate to English to get the gist of a voice message, podcast or clip in another language, all on your own device.
Pro tips
Answers to common questions about the speech to text.
Tiny is fastest and fine for clear, close-miked speech. Base makes noticeably fewer mistakes and is the best default. Small is the most accurate here, but needs a WebGPU graphics card and a download of roughly 410–590 MB.
Yes. Drop in an MP4, MOV or WebM file and the tool decodes its soundtrack. The video plays next to the transcript so you can check each line.
Usually within a tenth or two of a second for clear speech. They come from the model's attention, so fast speech, music and translation mode make them less exact. Segment timestamps are always available.
No. Whisper produces text and timings but doesn't separate speakers. You can add speaker names while editing the transcript.
The first run downloads the model you picked. It's stored in your browser's cache, so later runs start as soon as the model is loaded into memory, even without an internet connection.
No. ToolEnhancer processes files on your own device inside the browser. They are never sent to, stored on or seen by our servers.
Yes. The speech to text is free to use with no sign-up, no watermarks and no daily limits.