Speech recognition in the browser
Speech recognition turns spoken audio into written text. Here is how Whisper Web does that on your device, and what local processing means for your files.
The recording is processed on your device. The browser reads the audio, runs a Whisper model and produces editable text with timestamps. The application and model still need to download over the network.
From recording to transcript
Read the audio
Choose a file, record your microphone or provide a direct media URL that allows browser access. The browser decodes the media into audio samples.
Run the speech model
A Web Worker runs Whisper Tiny, Base or Small on the current device. Tiny Q8 with WebAssembly is the default; WebGPU is an optional backend on supported hardware.
Review the result
Edit timestamped segments, replay the recording and export TXT, JSON, SRT or VTT. Whisper Web also tries to save the text and timestamps in this browser's local history.
Local and cloud processing
The main difference is where the speech model runs. Local processing uses your device; a cloud service sends the work to a provider's infrastructure. Features and limits vary by service.
| Question | Whisper Web | Typical cloud service |
|---|---|---|
| Where is audio processed? | On your device | On a remote server |
| Is media uploaded? | No | Usually |
| What provides compute? | Your CPU or supported GPU | Provider infrastructure |
| What sets the limits? | Device, browser and tool limits | Provider, plan and network |
What local processing still depends on
Network access
The site and selected model download over the network. A pasted media URL also contacts its source server. No media upload does not mean a fully offline website.
A supported recording
The standard tool accepts up to 300 MB and 20 minutes. The large-file tool accepts a local file up to 1 GB and 1 hour. The browser must also support the audio codec inside the file.
Available device resources
Model size, memory, browser support and other running tasks affect processing time. Start with the default WebAssembly backend; WebGPU is optional.
Treat the transcript as a draft
Check names, numbers, quotations and specialist terms against the recording. Whisper Web does not identify speakers or write summaries; add those during review if you need them.
Export a copy before you leave. Local history stores text and timestamps, not the original recording. If saving fails, you can still export the current result. To replay a saved transcript later, choose the original media again.
Common questions
Does browser speech recognition upload my audio?
Whisper Web processes selected media on the current device. The application and model still download over the network, and a direct media URL contacts its source server. Other speech-recognition services may use a different data path.
Is WebGPU required for local speech recognition?
No. WebAssembly is the default in Whisper Web. WebGPU is optional and depends on browser and hardware support.
What happens after the transcript is created?
You can edit and export the result in the current workspace. Whisper Web also tries to save text and timestamps in this browser's IndexedDB. Export a copy if saving fails; saved records remain until deleted or browser storage is cleared.