Whisper Desktop is an Electron + React desktop app for local speech transcription with Whisper running through Transformers.js in a Web Worker.

It is derived from Xenova's whisper-web project and adapts that browser-first implementation into a desktop workflow with Electron packaging, file input, microphone recording, and local model execution.

This project is usable, but it should be treated as an early public alpha.

What works today:

• Local transcription in a desktop app
• Audio input from file, URL, or microphone recording
• Model download and caching through Hugging Face
• Transcript export as TXT and JSON
• Speaker diarization via pyannote.audio (Python sidecar) with optional speaker-count hint and word-level speaker assignment
• Inline rename / reassign / add-speaker controls in the transcript view

What is not shipped yet:

• Polished cross-platform installers
• Production hardening for broad non-technical distribution
• Persistence of speaker names across sessions or files

The packaged installer is currently Windows-first. The development setup should still run anywhere Electron and Node are supported, but the release flow in this repository is aimed at Windows builds.

npm install
npm run electron:dev

This starts Vite on port 5174 and launches Electron against the dev server.

npm install
npm run electron:build

This builds the renderer and creates a Windows installer with electron-builder.

• React renders the desktop UI
• Vite builds the renderer
• Electron hosts the app shell
• A Web Worker loads Whisper via @xenova/transformers
• Audio is resampled to 16 kHz before inference
• Model files are downloaded on first use and then cached locally

See docs/ARCHITECTURE.md for the implementation overview.

• First run may take time because Whisper model files need to be downloaded
• Performance depends heavily on local CPU and available system resources
• Large audio files can take noticeable time to decode and transcribe
• Diarization runs on CPU via Python (pyannote.audio) and roughly takes 0.3–1× the audio length after the first run; the pyannote model (~500 MB) is downloaded into the Hugging Face cache on first use

• Audio cleanup pass: optional ffmpeg-based high-pass + spectral denoise + loudness-normalize step that runs before both Whisper and pyannote see the audio. Toggle in Settings; helps noisy phone/laptop-mic recordings. ffmpeg is now bundled with the installer.
• Inline rename / reassign / add-speaker controls in the transcript view (carried over from 1.1.2). Click a speaker name to rename them everywhere; click the ▼ to reassign just one line; pick "+ Add new speaker" when pyannote merged two real people.
• Word-level diarization assignment (carried over from 1.1.1) — speaker labels are derived from per-word Whisper timestamps instead of coarse sentence chunks, fixing boundary mis-attributions.
• Number of speakers dropdown in Settings; setting an exact count (e.g. 2) noticeably improves accuracy.
• Probe-race fix: hitting Transcribe immediately after launch no longer silently skips diarization while the Python backend is still warming up. The transcribe request now awaits the probe.
• Python stderr is streamed to the main process console with a [py …] prefix so model-download / inference progress is visible while pyannote runs.
• .env loading at startup (carried over from 1.1.1): drop a file containing HUGGINGFACE_TOKEN=… next to the installed .exe (or in the repo root in dev) and it gets picked up automatically.

• Rename a speaker everywhere: click the speaker label (e.g. Speaker A) above any line to rename them (e.g. “Alex”). Every line tagged with that speaker updates instantly. Press Enter to save, Escape to cancel.
• Reassign a single line: click the small ▼ next to a speaker label to change just that one line to a different speaker. Useful when diarization mis-attributes a sentence at a speaker boundary.
• Add a new speaker: from the reassign menu, choose “+ Add new speaker” when pyannote merged two real speakers into one. The new speaker can then be renamed like any other.
• Renames and reassignments are reflected in the TXT and JSON exports.
• Edits live for the current transcript only; starting a new transcription resets them.

• Added a Number of speakers dropdown in Settings (visible when “Diarize speakers” is enabled). Set it to the exact count (e.g. 2 for an interview) to noticeably improve speaker accuracy; leave on Auto to let pyannote decide.
• Diarization now assigns speakers using word-level Whisper timestamps, which fixes the common case where a sentence straddling two speakers got attributed to the wrong person.
• The transcript view now shows an “Identifying speakers…” indicator while pyannote is still running after Whisper has finished, so it no longer looks like the app is idle.
• A .env file dropped next to the installed .exe (or in the repo root in dev) is now loaded at startup, so users can supply HUGGINGFACE_TOKEN without setting a system environment variable.

This project builds on the original whisper-web work by Xenova:

• Upstream project: https://github.com/xenova/whisper-web
• Transformers.js: https://github.com/xenova/transformers.js

The repository keeps the original MIT license and should be understood as a desktop adaptation of that foundation.