SigmaVoice

Standalone, system-wide, local-first voice→text dictation. Global hotkey → capture → on-device Whisper → paste into any focused app (editor, terminal, Slack, browser, anywhere). Runs in the system tray. Free, local, private — no cloud, no account.

Inspired by BridgeMind's BridgeVoice — but on-device by default and not paywalled.

Features (v0.4)

• Global hotkey dictation — push-to-talk (hold a modifier combo like ⌘⇧ or a base-key shortcut) and toggle modes, set with a "press your keys" shortcut recorder.
• On-device Whisper transcription (tiny/base/small/medium) via whisper.cpp — fully offline; in-app model download (progress + Cancel); macOS Speech.framework fallback.
• Paste into anything — clipboard + macOS Accessibility keystroke (CGEvent).
• Focus-preserving recording HUD overlay (never steals focus from your target app).
• Always-visible floating pill — click-to-dictate, drag-to-move (position remembered), idle/listening/processing states; switchable Logo & text ↔ Logo only.
• Apple-grade settings — macOS sidebar, system light/dark + vibrancy, an Overview dashboard (Words / Sessions / WPM), hotkey shown as ⌘⌥⎵ keycaps.
• Dictionary & verbal macros — literal substitutions (e.g. "new line" → ⏎, "bridge mind" → "BridgeMind") applied to every transcript.
• Usage stats — words, sessions, average WPM, recent activity.
• Single-instance, tray-only, persistent settings.

Install (macOS · Apple Silicon)

One-line install — curl doesn't set the Gatekeeper quarantine flag, so it just works (no "unidentified developer" dialog):

curl -fsSL https://raw.githubusercontent.com/s1gmamale1/SigmaVoice/main/scripts/install-macos.sh | bash

Pin a version: append -s v0.4.0. Prefer to inspect first:

curl -fsSL https://raw.githubusercontent.com/s1gmamale1/SigmaVoice/main/scripts/install-macos.sh -o install-sigmavoice.sh
less install-sigmavoice.sh && bash install-sigmavoice.sh

Or grab the DMG manually from Releases and, after dragging to Applications, run xattr -cr /Applications/SigmaVoice.app (or System Settings → Privacy & Security → Open Anyway).

SigmaVoice runs in the menu bar. First use prompts for Microphone + Accessibility (paste), and Input Monitoring for push-to-talk — System Settings → Privacy & Security. (Windows installer coming in a follow-up.)

Architecture — the voice engine lives in a submodule

SigmaVoice is the app shell. The voice engine (capture state machine, output router, Whisper/CLI engines, model registry) and the native bindings are shared with SigmaLink and consumed here via a git submodule at sigmalink/ — single source of truth, no duplicated voice code.

SigmaVoice/                         (this repo — the standalone app)
├── src/                            main process: main.ts, hotkey-manager.ts,
│                                   hud-window.ts, text/kv/settings helpers, preloads
├── renderer/                       settings.html + hud.html (no framework)
├── scripts/build.cjs               esbuild bundler (natives kept external)
├── electron-builder.yml            DMG (mac arm64+x64) / NSIS (win x64), unsigned
├── package.json                    link: deps → sigmalink/app/{packages,native}/*
└── sigmalink/                      git submodule → SigmaLink (the voice engine)
    └── app/
        ├── packages/voice-core/    @sigmalink/voice-core  (engine, TS)
        └── native/voice-{mac,win,whisper}/   N-API bindings (+ whisper.cpp)

The four @sigmalink/* dependencies use pnpm link: (not file:) so the natives' built build/Release/*.node stay resolvable through the submodule.

Where development happens

App-shell development happens in THIS repo (src/, renderer/, scripts/, configs). The voice engine + native modules are authored in SigmaLink and consumed here via the ./sigmalink/ submodule — fix engine/native code in SigmaLink, then bump the submodule pointer:

git -C sigmalink fetch origin && git -C sigmalink checkout <sigmalink-sha>
git add sigmalink && git commit -m "bump engine submodule → <sha>"

The two open bugs (Windows build, quit-time abort) are in the engine/natives → fixed in SigmaLink. Full orientation for maintainers: docs/HANDOFF.md. Conventions: CLAUDE.md. (SigmaLink/sigma-voice/ is the historical dev copy, pre-2026-05-29; this repo is authoritative for the app shell now.)

Develop / build

git clone --recurse-submodules https://github.com/s1gmamale1/SigmaVoice.git
cd SigmaVoice
pnpm install                 # links the engine from ./sigmalink
pnpm typecheck               # tsc --noEmit
pnpm build                   # esbuild → sigma-dist/{main.js,preload.cjs,hud-preload.cjs}
pnpm dev                     # run from source (tray app)

First the natives must be built for Electron's ABI (CI does this; locally, build them once in the submodule, e.g. cd sigmalink/app && pnpm install then node-gyp rebuild --target=<electron-version> --dist-url=https://electronjs.org/headers in each native/voice-*).

Installers

pnpm run pack:mac    # macOS DMG (arm64 + x64), ad-hoc signed
pnpm run pack:win    # Windows NSIS (x64), unsigned

Do not run pack:* against a SigmaLink checkout you develop in. With link: deps + npmRebuild, electron-builder can recompile the natives over the symlink into the shared engine tree. This repo ships npmRebuild: false (the natives are built once, then copied) to avoid that; packaging is meant for CI / a clean checkout.

macOS permissions

System-wide dictation needs, on first use (granted in System Settings → Privacy & Security):

• Microphone — capture.
• Accessibility — paste the transcript into the focused app (CGEvent Cmd+V).
• Input Monitoring — only for push-to-talk (the global key-UP that ends hold-to-talk). If denied, SigmaVoice degrades to tap-to-toggle and tells you.

Unsigned/ad-hoc build → Gatekeeper shows "unidentified developer" (recoverable): xattr -cr /Applications/SigmaVoice.app or System Settings → Open Anyway.

Distribution posture

Internal-use, unsigned (macOS ad-hoc / Windows no Authenticode). macOS arm64 + Windows x64 only. When a Developer ID / Authenticode cert is acquired: drop the afterSign ad-hoc hook, set mac.identity, hardenedRuntime: true, notarize: true.

Release CI

.github/workflows/release.yml triggers on v* tags: checks out the submodule recursively, builds the natives for Electron's ABI in sigmalink/app, bundles, and produces the unsigned DMG / NSIS installer. First-run validation pending on real GitHub runners (multi-arch native build) — until then, SigmaLink's release-sigma-voice.yml remains the proven release path.

Deferred

• Windows keystroke-inject (sendPasteKeystroke equivalent) + Tier-3 typing fallback.
• AI cleanup / custom-instructions reformatting; optional cloud transcription.
• Wake-word ("Hey Jorvis") listening mode (engine code exists, off by default).
• Electron auto-updater; Developer-ID signing + notarization; DMG background art.