wingetCollections
  • Auto
  • Search
  • Awesome Collections
  • Awesome Programs
  • My Favorite Collections
  • My Collections
  • Account / Settings
  • Terms of Service
  • Privacy
  • About
  • What is WinGet?

© 2026 winget.ragerworks.com

Loading...
  1. Go back
  2. Packages
  3. VoiceWave Local Core
VoiceWave Local Core logo

VoiceWave Local Core Rithwik Posani

accessibility dictation privacy speech-to-text transcription voice whisper
Use this command to install VoiceWave Local Core with WinGet:
winget install --id=VoiceWave.LocalCore -e

VoiceWave Local Core is a privacy-first desktop dictation application designed for seamless speech-to-text transcription on Windows. It captures audio locally, processes it using whisper.cpp models without sending data to external servers, and inserts transcribed text directly into applications with reliability.

Key Features:

  • Local Transcription: All processing occurs on-device, ensuring no audio leaves your machine.
  • Global Hotkey Control: Use push-to-talk or toggle dictation with customizable hotkeys for efficient input.
  • Direct Text Insertion: Reliably inserts transcribed text into the active application, with clipboard and history fallbacks to ensure accuracy.
  • Verified Model Management: Ensures Whisper models are validated through checksums for consistent performance.

Audience & Benefit: Ideal for privacy-conscious users, developers, and professionals on Windows seeking reliable, offline transcription. VoiceWave Local Core offers a secure alternative to cloud-based dictation tools, prioritizing user data protection while maintaining high accuracy and responsiveness.

Install VoiceWave Local Core via winget with ease.

README

VoiceWave

VoiceWave hero

> 🎙️ Offline, privacy-first dictation for Windows · no cloud · no subscription trap · Whisper accuracy · < 500 ms cold start

winget install VoiceWave.LocalCore

GitHub Stars GitHub Downloads Release License Platform winget

Overview Live Demo

VoiceWave in action: pill activation, live capture, and instant text insertion — all on-device.

Install

Via Microsoft winget (one line, recommended):

winget install VoiceWave.LocalCore

winget upgrade VoiceWave.LocalCore keeps you current on every future release.

Or grab the latest installer directly and run it. Once installed, VoiceWave keeps itself current via signed in-app auto-update.

Why VoiceWave?

VoiceWaveWispr Flow ProDragon ProfessionalWindows Speech
PriceFree$15/mo~$500 one-timeFree
Works Offline✅ Always❌ Cloud-only✅ Yes✅ Yes
Audio stays on device✅ Never leaves❌ Sent to servers✅ Local✅ Local
Accuracy engineWhisper (local)Whisper (cloud)ProprietaryProprietary
Open Source✅❌❌❌
Windows✅✅✅✅
Free tier✅ (open source)2,000 words/week❌✅ (limited)
Latency< 500 msNetwork-dependent< 1 sVaries
Customizable✅ Rust/React❌❌❌

Wispr Flow Pro pricing source: wisprflow.ai/pricing (April 2026 — $15/mo, $12/mo billed annually).

What's New in 0.6.0

Hands-free dictation, dark mode, and deeper insights

  1. Double-tap the push-to-talk shortcut to dictate hands-free. Pause when you are done or press the shortcut again; Escape cancels the active dictation.
  2. Hands-free capture keeps a short release tail when stopped manually, so quiet final syllables are not clipped.
  3. VoiceWave now has complete Light, Dark, and System themes across the app shell, dashboard, settings, and floating recording pill.
  4. Statistics now include insertion-success-aware totals, a GitHub-style streak heatmap, and richer insight panels.

What's New in 0.5.9

Usage visibility and dashboard polish

  1. Optional anonymous activated-install counting records one first successful dictation and at most one active marker per day. It is off by default and never sends audio, transcript text, account identity, filenames, or settings.
  2. Updates keep the same random installation identifier, so updating VoiceWave does not inflate the active-install count.
  3. The Home greeting now uses the signed-in profile name or a general guest welcome instead of a hardcoded name.
  4. The website now labels and counts installer transfers only, excluding updater metadata checks and unrelated release assets.

What's New in 0.5.8

Crash safety

  1. VoiceWave could refuse to launch after an unclean shutdown. If the machine lost power or the app was force-killed while saving one of its data files, that file could be left with its size recorded but no contents — and the app would then fail to start, with no window and no message. Every store now recovers: it finishes the interrupted save, falls back to a backup, and if neither can be read, sets that file aside and starts from defaults.
  2. Saves are crash-safe. Data is written to a temporary file and flushed to disk before replacing the original, so losing power mid-save can no longer destroy your settings, dictionary, snippets, history, or stats.
  3. A damaged encryption key is set aside and regenerated instead of blocking startup, and VoiceWave tells you once if anything had to be reset.

What's New in 0.5.7

First-run reliability

  1. A failed speech-model download no longer traps setup. If the one-time download fails — no connection, a captive network, a dropped link — VoiceWave now says so and offers Retry instead of showing a progress bar that never moves. Whatever downloaded before the failure is kept.
  2. Cancel actually cancels a model download, and keeps the bytes already fetched so Install resumes rather than restarting. Pause and Resume, which never applied to these models, are no longer shown.
  3. Quitting during a download no longer leaves a background helper process running after the app is closed.
  4. Setup shows downloaded size ("213 MB of 465 MB") alongside the percentage, so a slow connection reads as slow rather than stuck.

What's New in 0.5.6

Voice snippets

  1. Say a phrase you choose and VoiceWave inserts your saved text exactly as you wrote it — casing, punctuation, line breaks, and URLs preserved byte-for-byte. Triggers work standalone or inline, and repeat within one dictation.
  2. Saved expansions are protected through deterministic formatting and the on-device polish pass, so neither the formatter nor the language model can reword content you own.
  3. Encrypted on device, works signed out and offline, replicates across your devices when signed in. Own page in the app with search, edit, and delete.

Dictionary fixes and local-first storage

  1. Removed a plaintext dictionary backup file left on disk by an older migration, and stopped unapproved suggestions from influencing transcription before you approve them.
  2. Dictionary storage was rebuilt so the encrypted on-device store is the single source of truth in every account state, with optional cloud storage as replication on top — tombstoned deletes that propagate, and local writes that never roll back on a remote failure. Groundwork rather than a user-visible change: cloud sync ships disabled in release builds.

What's New in 0.5.5

Mode-aware polish profiles

  1. Coding and Writing use separate, compact on-device prompts tuned for the bundled Qwen2.5-1.5B model. A Rust fidelity gate accepts the result or falls back to deterministic formatting when the model is slow or uncertain.
  2. Literal never calls the language model and keeps fillers and repeated words; Standard preserves immediate insertion and can offer an optional polished version afterward.
  3. Selecting an AI-shaped profile prepares the local model automatically, and the active profile is warmed after launch. (The four-thread production CPU cap was listed here in error — it landed after this build was cut and ships in 0.5.6.)

What's New in 0.5.0

Zero-setup offline transcription

  1. The installer now ships a self-contained CPU faster-whisper runtime — a fresh install transcribes the default model immediately, with no Python install and no manual setup, fully offline.

On-device AI polish (opt-in)

  1. An optional local LLM pass cleans up filler, capitalization, and punctuation after dictation and offers the polished text on the pill. It runs entirely on-device and uses a fail-closed fidelity validator; rejected output falls back to deterministic text.

Productivity

  1. Dictionary export / import — back up and move your custom-term dictionary between machines.
  2. Interactive pill actions — one-tap actions on pill notices (copy the transcript, add a term).
  3. Mic-volume guard — a Dynamic-Island-style pill notice warns when your input level is too low, so quiet-mic dictations don't silently fail.

Insertion

  1. Terminals are now detected by process, not window title — dictation pastes correctly into the Codex and Claude desktop apps (previously misread as CLIs and forced to clipboard-only).
  2. The "Polished version ready" pill surfaces occasionally rather than after every dictation; fallback and error notices still show every time.

What's New in 0.4.0

Updates

  1. In-app auto-update — VoiceWave checks GitHub on launch and offers a one-click "Install & Restart" when a newer signed release is available. Updates are cryptographically signed and verified before install.

What's New in 0.3.1

Audio pipeline

  1. Soft word endings ("s", "th", "f", drifted "e") no longer clipped — post-release capture window lifted to 300 ms, with a matching 300 ms trim pad on both capture and inference layers.
  2. Volume-adaptive trim thresholds — quiet speakers and post-pause resumptions now land fully instead of being truncated as silence.
  3. All aggressive DSP (pre-emphasis, gain normalization, noise attenuation, hum notch, soft limiter) defaults to off. Whisper receives audio as captured.

Inference

  1. Cold-start prewarm at app launch — first dictation drops from 2–5 s to ~500 ms.
  2. Extra hallucination guard: log_prob_threshold = -1.0 on the primary decode alongside the existing no_speech and compression_ratio floors.
  3. Opt-in whisper.cpp models: wcpp-small.en (~466 MB) and wcpp-large-v3-turbo (~1.6 GB). Vulkan backend available behind the whisper-vulkan cargo feature.

Reliability

  1. Push-to-talk release no longer drops silently on key bounce.
  2. Clipboard paste keeps the dictated text — no more "old content pasted" after delayed Ctrl+V.
  3. SendInput refuses to type into Windows security dialogs (UAC, Credential Manager, PIN prompts).
  4. Worker stdout drained between requests to prevent ID-mismatch stalls after cancel/retry.
  5. Decode threads reserve one CPU core for the UI to stop taskbar stutter during transcription.

Overview Product At A Glance

AreaSummary
RuntimeTauri 2 shell + Rust core + React/Tailwind frontend
ASR (default)faster-whisper via a bundled CPU Python runtime (fw-small.en / fw-large-v3), CUDA auto-detected when present — works offline out of the box
ASR (opt-in)whisper.cpp via whisper-rs (wcpp-small.en / wcpp-large-v3-turbo), Vulkan feature flag
Privacy PathNo outbound audio transport, no cloud transcription path in v1
UX ContractExplicit state model: idle -&gt; listening -&gt; transcribing -&gt; inserted/error
Insertion ReliabilityDirect insert -> clipboard fallback -> history fallback (blocked on Windows security dialogs)
Platform ScopeWindows implementation/validation active since 2026-02-10

Capabilities Core Capabilities

IconCapabilityTechnical Detail
Local-onlyLocal-Only PrivacyProduction path has no outbound audio transport and no cloud transcription rewrite path for v1.
RuntimeDeterministic Runtime FlowAudio capture, inference orchestration, insertion, and persistence are handled in Rust services behind Tauri commands/events.
InsertionFallback-Safe InsertionInsertion engine prioritizes direct insertion and degrades to clipboard/history fallback to preserve user text.
ModelsVerified Model LifecycleModel install/switch includes cataloging, checksum verification, and recommendation logic.

Architecture Runtime Architecture

VoiceWave runtime pipeline

Primary architecture boundary is defined in docs/rfc/0001-system-architecture.md.

Core modules:

  1. desktop-shell (Tauri host)
  2. audio-pipeline (capture, resample, buffering, VAD)
  3. inference-worker (whisper.cpp integration + cancellable jobs)
  4. insertion-engine (direct/clipboard/history reliability chain)
  5. hotkey-manager (global binding lifecycle)
  6. model-manager (catalog/download/checksum/health)
  7. persistence (local settings/history/stats)
  8. diagnostics (redacted export, opt-in)
  9. experience-state (shared state contract for UX)

Status Status Snapshot (Repository Baseline)

As documented in this branch:

  1. Phase 0 complete.
  2. Phase 1 complete for Windows runtime integration.
  3. Phase 2 implemented and validated in Windows execution scope.
  4. Phase 3 implemented and validated for Windows rescue baseline.
  5. Phase 4 and Phase 5 gate automation is implemented (phase4, phase5, and reliability checks).
  6. Release gate automation is implemented (npm run release:gate).
  7. Phase 6 not started.

Open release blockers called out in current docs:

  1. Artifact freshness policy now requires fresh evidence (&lt;= 7 days) for release-candidate decisions.
  2. Reliability exit thresholds are stricter (insertion &gt;= 98%, correction &lt;= 12%, crash-free &gt;= 99.5%, TTFSD &lt;= 3).
  3. Legal/compliance checklist and risk register are now release-gate inputs and must stay current.

References:

  1. docs/PHASE3_IMPLEMENTATION.md
  2. docs/PHASE4_READINESS.md
  3. docs/PHASE5_READINESS.md
  4. docs/phase3/artifacts/windows-manual-acceptance-2026-02-11.md

Stack Stack

  1. Frontend: React 18 + Tailwind + Vite
  2. Desktop shell: Tauri 2
  3. Core runtime: Rust
  4. ASR (default): faster-whisper (CTranslate2) via Python subprocess; CUDA auto-detected
  5. ASR (opt-in): whisper.cpp via whisper-rs; Vulkan behind whisper-vulkan cargo feature
  6. Local storage/ops: encrypted billing files + local runtime state/history artifacts

Quick Start Quick Start

  1. Install dependencies:
npm install
  1. Run frontend dev mode:
npm run dev
  1. Optional: enable cloud auth + sentence sync (Firebase):
Copy-Item .env.example .env

Then fill the VITE_FIREBASE_* keys in .env from your Firebase project settings. When configured, VoiceWave enables:

  1. Email/password sign-up + sign-in
  2. Per-user cloud storage of only the latest 5 sentences
  3. Per-user synced approved dictionary terms
  4. Recent sentence list in Home and Dictionary views

Firestore rules template is provided at:

  1. docs/firebase/firestore.rules

  2. Run tests:

npm run test -- --run
  1. Build frontend:
npm run build
  1. Run desktop app (requires Rust/Tauri prerequisites):
npm run tauri:dev
  1. In app: open Models, install fw-small.en (recommended default, ~466 MB) or fw-large-v3 for highest accuracy. Optional: wcpp-* whisper.cpp variants live alongside them. Then run dictation from Home.

Validation Validation and Gates

CommandPurpose
npm run phase1:validatePhase I validation suite
npm run phase1:batteryPhase I battery/thermal run
npm run phase2:validatePhase II validation suite
npm run phase3:validatePhase III validation suite
npm run phase4:prepPhase IV readiness report
npm run phase4:gatePhase IV blocking gate
npm run phase5:prepPhase V readiness report
npm run phase5:gatePhase V blocking readiness gate
npm run phase5:reliabilityPhase V reliability evidence report
npm run phase5:reliability:gatePhase V blocking reliability gate
npm run release:gateEnd-to-end release blocking gate (phase4 + phase5 + risk/compliance)
npm run test:coverageFrontend test + coverage thresholds gate
npm run security:secrets -- -EnforceEnforced secret leakage scan
npm run security:deps -- -EnforceEnforced dependency vulnerability gate
npm run security:firestore-rules -- -EnforceFirestore rules schema/policy contract gate
npm run quality:frontend:gateFrontend quality gate (coverage + build)
npm run quality:backend:gateBackend quality gate (tests + inventory threshold)
npm run phaseA:cpuCPU latency sweep
npm run phaseB:gpu:checkGPU readiness check
npm run phaseB:gpuGPU latency sweep
npm run phaseB:fwFaster-whisper latency sweep

Contract Runtime Contract Surface (High-Level)

Phase I baseline commands:

  1. start_dictation(mode)
  2. cancel_dictation()
  3. load_settings()
  4. update_settings(settings)
  5. get_voicewave_snapshot()

Phase II additions include hotkey, permission, and insertion command surfaces.

Phase III additions include model manager, benchmark/recommendation, history, and dictionary command surfaces.

Phase V additions include diagnostics status and export command surfaces.

See full contract list in docs/rfc/0001-system-architecture.md.

Monetization Monetization and Entitlements

VoiceWave is free to use. Download, install, and run — no subscription required.

Reference: docs/monetization.md

Security Privacy and Security Guardrails

  1. No outbound audio transport in production path.
  2. Local-only ASR and deterministic local post-processing in current monetization architecture.
  3. Model/update verification paths are documented in phase evidence.
  4. Diagnostics export is user-triggered and revocable.
  5. Anonymous activated-install analytics is off by default. If explicitly enabled, VoiceWave sends only a random installation identifier, app version, first-success marker, and once-per-day active marker — never audio, transcript text, account identity, filenames, or settings.

References:

  1. docs/security/threat-model-v1.md
  2. docs/risk/risk-register.md
  3. docs/phase4/evidence/update-signing-verification.md

Evidence Latest Recorded Validation Evidence

Latest recorded run: 2026-07 (against 0.4.0 baseline).

  1. Rust library test suite: 218 pass / 0 fail.
  2. Frontend test suite (npx vitest run): 34 pass / 0 fail.
  3. npm run build: recorded pass.
  4. cargo check --no-default-features --release: recorded pass.

Reference artifact trail:

  1. docs/phase3/artifacts
  2. docs/phase4/artifacts
  3. docs/phase5/artifacts
  4. docs/testing/hardware-tier-recommendation-windows.json

CI CI

Workflow: .github/workflows/ci.yml

Current CI baseline includes:

  1. Docs formatting checks
  2. Markdown lint
  3. Secrets scan
  4. Phase 0 artifact integrity checks
  5. Frontend tests and build
  6. Rust tests and compile paths
  7. Release gate job (npm run release:gate) on Windows runner

Local pre-commit guard (recommended):

  1. git config core.hooksPath .githooks
  2. chmod +x .githooks/pre-commit (macOS/Linux only)

Repo Repository Map

src/                   # React app and UI runtime bridge
src-tauri/             # Rust core runtime + Tauri shell
voicewave-website/     # Marketing website
docs/                  # RFCs, phase plans, evidence, security, testing
scripts/               # Validation, readiness, benchmark, tauri utilities
vendor/                # Local whisper-rs / whisper.cpp vendored deps

Docs Source of Truth Docs

  1. Product requirements: docs/prd/v1-prd.md
  2. Architecture RFC: docs/rfc/0001-system-architecture.md
  3. Test strategy: docs/testing/test-strategy.md
  4. Hardware tiers: docs/testing/hardware-tiers.md
  5. Release thresholds: docs/testing/release-thresholds-windows.json
  6. Legal/compliance checklist: docs/testing/legal-compliance-checklist.md
  7. Phase recovery plan: docs/PHASE_RECOVERY_PLAN.md
  8. Implementation ledger: Idea.md
  9. Phase I implementation: docs/PHASE1_IMPLEMENTATION.md
  10. Phase II implementation: docs/PHASE2_IMPLEMENTATION.md
  11. Phase III implementation: docs/PHASE3_IMPLEMENTATION.md
  12. Phase III remaining: docs/PHASE3_REMAINING.md
  13. Phase IV readiness: docs/PHASE4_READINESS.md
  14. Phase V readiness: docs/PHASE5_READINESS.md
  15. Changelog: CHANGELOG.md
  16. Contributing: CONTRIBUTING.md

Verify Local Verification Utility

# Phase 0 artifact checks
&amp; "C:\Windows\System32\WindowsPowerShell\v1.0\powershell.exe" -ExecutionPolicy Bypass -File .\scripts\ci\check-phase0-artifacts.ps1
Versions
0.6.0
0.5.1
0.3.1
0.2.0
Website
github.com
License
Last updated
8/20/2026
Download latest version
Loading...