Get Weeve

How browser transcription works

Weeve is a static web application. Your browser decodes the selected recording, downloads speech and speaker models from Weeve’s CDN, and runs inference in local workers. Your recording and transcript stay on this page.

Drag audio files here

or

Compatibility at a glance

A recent Chrome, Edge, or Safari release is required. WebGPU is used when available, with a WebAssembly fallback. Files are limited to 60 minutes.

  • Local decoding
  • Model cache
  • Web workers
  • WebGPU or WASM

How it works

Four clear steps from local file to reviewed text.

  1. 01

    Decode and resample

    The browser reads the chosen file and creates 16 kHz mono audio in local memory.

  2. 02

    Download and cache models

    A first WebGPU run downloads about 1.3 GB. WASM downloads about 0.94 GB.

  3. 03

    Transcribe and separate speakers

    Speech recognition and diarization run concurrently in browser workers.

  4. 04

    Align and export

    Timed words and speaker segments are aligned locally for review and export.

Product details

Practical details for processing explained.

The network boundary

The app shell, speech models, and WebAssembly runtime arrive over the network. Models come from models.getweeve.io and the diarization runtime is served under the same /tools application.

The recording bytes, decoded waveform, transcript text, and speaker names are not sent to Weeve for processing. This is the precise privacy boundary behind “Your recording never leaves this page.”

Caching and browser storage

Model files are stored in the browser Cache API after download when the browser permits it. A later run can reuse those files, avoiding another full model download.

Browser storage is not guaranteed forever. Private browsing, storage pressure, manual site-data clearing, or browser policy can remove the cache.

Workers, fallback, and failure behavior

The browser tries WebGPU for faster speech inference when an adapter is available. WebAssembly with SIMD is the fallback. Separate workers keep model work away from the main interface thread.

Speaker diarization is optional to a successful word transcript. If speaker processing fails, Weeve keeps the completed transcript and explains that speaker labels are unavailable.

Analytics and open components

Optional PostHog analytics are cookieless, memory-only, and disabled when no key is configured. They record coarse routes and product states, not file names, transcript text, speaker names, or arbitrary query parameters.

The model, runtime, and font licenses are listed on the licenses page. Weeve Tools remains a static export with no API route or server transcription endpoint.

Your recording stays here

Recording bytes, decoded audio, transcript text, and speaker names stay on this page. Speech models download from Weeve’s CDN.

A transcript you can review

Timestamps and speaker labels stay editable before export.

Maya

Let’s confirm the launch checklist before we finish.

Speaker 02

The help article is ready. I will send the final link this afternoon.

Maya

Great. I renamed your speaker label, so the exported notes will use your name.

CopyMarkdownTextPDF
Transcribe now

Frequently asked questions

Practical answers about files, processing, privacy, and results.

Related transcription guides

Continue with product details that match your file.