Whisper Transcriber
Turn recordings into editable text without sending them anywhere. Audio is decoded, resampled, and transcribed by a Whisper model that runs inside a Web Worker in your own browser — there is no transcription server behind this utility.
What it does
Whisper Transcriber converts speech into text on your own machine. You give it an audio or video file — or record straight from the microphone — and it produces a full transcript plus timestamped segments you can edit and export as TXT, SRT, or WebVTT subtitles.
Everything happens in the browser tab: the media is decoded and downmixed to mono 16 kHz, the OpenAI Whisper model runs through Transformers.js on ONNX Runtime Web, and the result never leaves the page. Nothing is uploaded, queued, or stored on a server.
The app is built as a standalone micro-application under static/utility-apps/whisper-transcriber. The CAD AutoScript shell embeds it in a same-origin iframe, so the transcriber can evolve in its own repository while sharing this site's catalog, article, reactions, comments, and access layer.
First run: the model download
The first time you press Start Transcription, the browser downloads the multilingual Whisper Tiny weights from the Hugging Face Hub and caches them. This is the only network traffic the utility generates, and it carries no audio — only the model files.
How much depends on which engine runs, and the badge at the top of the utility always shows the figure for your device before you start:
| Engine | Weights | Typical device |
|---|---|---|
| WASM | ~41 MB | Phones and tablets, and any browser without WebGPU |
| WebGPU | ~77 MB | Desktop browsers with GPU support |
- Later runs start from that cache and need no network at all.
- Clearing site data, using a private window, or browser storage pressure can evict the cache and trigger a fresh download.
- The model revision is pinned to a fixed commit, so the same version is fetched every time.
When to use it
- Writing up meeting, supplier-call, or design-review notes from a recording.
- Turning site walkthrough and inspection videos into searchable text.
- Dictating punch lists or QA observations instead of typing them on site.
- Producing SRT or WebVTT subtitles for internal training and handover videos.
- Transcribing material you are not allowed to upload to a third-party service.
Workflow
- Drop an audio or video file into the drop zone, or press Record Audio and stop when you are done.
- Wait for the status line to report the media as ready — it shows duration, and whether the windowed pipeline will be used.
- Choose language, timestamps, and engine if the defaults do not suit the recording.
- Press Start Transcription. On the first run, the progress bar shows the model download; after that it goes straight to inference.
- Correct the transcript in place — the text area is editable, and the exports below read from it.
- Copy the text, or export TXT, SRT, or WebVTT.
- Press Clear File to drop the media, transcript, and loaded model from memory.
Settings
| Setting | Options | What it changes |
|---|---|---|
| Model | Whisper Tiny (multilingual) | The weights used for recognition. The size shown is the one your device will actually fetch. |
| Language | Automatic Detection, English Only, Russian Only | Automatic lets Whisper detect the language; picking one explicitly is more reliable on short or noisy clips. |
| Timestamps | Show Timestamps, Text Only | Timestamps produce the segment list that SRT and WebVTT export needs. |
| Engine Runtime | Auto (GPU/WASM), WASM Only, WebGPU Preferred | Auto uses the GPU on desktop and WASM on phones. WASM is slower but works everywhere, and the utility falls back to it by itself whenever the GPU path is unavailable. |
Long recordings
Short clips are decoded in one pass. Longer media is probed first and then run through a conveyor of mono windows, so only one window of audio sits in memory at a time while the model stays loaded between windows. Desktop uses 30-second windows with 5 seconds of overlap; phones use 15 and 3, which halves the peak memory. The transcript fills in window by window, and Cancel stops the run without losing what has already been produced.
On a phone
The utility runs on mobile browsers, with a few adjustments made for you: the WASM engine is the default, the model download is the smaller 41 MB one, and audio is windowed more tightly. Two things are still worth knowing.
- Keep the tab in the foreground. The utility asks to keep the screen awake while it works, but that does not stop the operating system suspending the page if you switch to another app. A long transcription needs the tab visible.
- Expect it to be slower than a laptop. The work is real inference on your own device; a phone simply has less to give it.
Privacy
- Audio, video, filenames, transcript text, and your edits stay in the page. They are never sent to CAD AutoScript, to analytics, or to any transcription API.
- The only outbound requests are the pinned model files from the Hugging Face Hub. There are no fonts, scripts, or trackers loaded from anywhere else, and the build is checked for that automatically before it can be published.
- The microphone is only accessed after you press Record Audio, and this is the only utility on the site that is granted microphone permission at all.
- Clear File releases the microphone stream, the audio buffers, and the loaded model. Closing the tab does the same.
- Nothing is written to your account, and no transcript history is kept — save or export anything you want to keep before you leave the page.
Requirements and limits
- A current Chromium, Firefox, or Safari release with WebAssembly support, on desktop or mobile. WebGPU is optional and only adds speed.
- 41 MB of free browser cache for the model (77 MB on the WebGPU path), plus working memory proportional to the recording length.
- Whisper Tiny favours speed over accuracy: expect to proofread technical vocabulary, part numbers, and proper names.
- Heavy accents, crosstalk, and noisy site audio degrade the result — a closer microphone helps far more than any setting here.
- Speaker diarization ("who said what"), live rolling transcription, and larger Whisper models are not part of this version.
- ⚠️ Transcripts are a drafting aid. Verify them against the recording before quoting them in a report, a minute of meeting, or any contractual document.
Related tools
Feedback / bug report
- Open a GitHub issue
- Email or DM with the slug
whisper-transcriberso we can reproduce the issue.