Speak naturally.Keep it private.
Turn your voice into polished text in any app — processed 100% locally on your device with No audio uploads, Zero cloud transcripts, and 100% offline & private.
“hey team quick update on the latency benchmarks real time factor dropped to zero point two please test and report any bugs in eng channel”
Works instantly anywhere you can place a cursor
No browser plugins, no custom integrations, and zero clipboard overwrites. HushWrite simulates native OS keyboard events directly into the frontmost window.
Yourvoiceisyourfastesttool.Turnunfilteredthoughtsintopristinewriting,code,anddecisionswithouttouchingthecloud.
Whisper AI inference runs directly on your Apple Silicon Neural Engine or Windows DirectML GPU with zero cloud subscriptions, zero telemetry, and complete offline privacy.
100% On-Device Compute
All acoustic feature extraction and token decoding run directly in local GPU/Neural Engine memory. No cloud roundtrips, no API keys, and no rate limits.
Zero Telemetry or Audio Logs
Your microphone audio never touches a remote server. Verified by packet inspection tools like Wireshark and LuLu. 0 bytes outbound network egress.
Universal OS Injection
Dictate into Cursor, Slack, Notion, Xcode, or Terminal at the OS level. HushWrite simulates native key events directly without overwriting your clipboard history.
Test HushWrite in your browser.
Experience how HushWrite turns unstructured messy speech into formatted writing while stripping filler words with zero cloud latency.
Backend Refactor Note
Coding & Architecture“we need to refactor the payment gateway to use exponential backoff with jitter and add an idempotent key to prevent double charging on stripe webhooks”
Beat the cloud on latency.
Local AI is only compelling if it feels instantaneous. We engineered HushWrite in native Rust to outpace cloud WebSocket pipelines at every stage.
Tail Latency (Speech End → Injected)
Cloud dictation pays a mandatory network round-trip penalty (DNS + TLS handshake + WebSocket upload + cloud inference queue). HushWrite decodes inside GPU VRAM immediately.
Make HushWrite sound like you.
HushWrite adapts to how you write in different apps. Set a different style for messages, work chats, emails, and code editors—computed 100% locally on your machine.
“hey are you free for lunch tomorrow let's do 12 if that works for you”
Hey, are you free for lunch tomorrow? Let’s do 12 if that works for you.
One voice input. Formatted for every app.
HushWrite senses the frontmost active window and shapes tone, punctuation, and layout to fit the medium.
“hey mark we just merged the auth token rotation fix to main let me know if staging looks good thanks”
Own your customization.
Generic cloud dictation gives you one rigid output style. HushWrite gives you complete local automation, vocal snippets, and portable dictionaries.
Voice-Triggered Text Snippets & Macros
Define vocal expanders. When you say 'bug template', 'schedule link', or 'invoice address', HushWrite replaces the phrase with your structured schema immediately.
“insert bug template”
### [Bug]: Title - **Steps to Reproduce:** 1. Open dashboard settings 2. Toggle Air-Gap mode - **Expected:** Sockets close immediately - **Actual:** Verified 0 bytes egress - **Environment:** HushWrite v0.1.0 · macOS Sonoma
Features that solve real daily jobs.
Built for developers, lawyers, healthcare professionals, and founders who cannot—and will not—stream sensitive voice data to the cloud.
“Speak the messy first draft. Paste the polished version.”
Speak freely with false starts and filler words. HushWrite strips ums, ahs, and repetitions, injects proper punctuation, and outputs clean paragraphs ready to send.
“One shortcut works wherever you write.”
Press Option+Space (macOS) or Alt+Space (Windows) in any text field. Speak naturally and text pastes directly at your cursor via native OS input injection.
“Your client notes and private thoughts stay yours.”
Inference runs 100% on your local CPU/GPU using whisper.cpp. Audio is held only in RAM and deleted on decode. Zero cloud telemetry, zero remote databases.
“Dictate on a plane, in a hotel, or anywhere Wi-Fi fails.”
Whisper weights execute completely offline on your device. Transcribe on 35,000-ft flights, in subways, or in secure air-gapped rooms with zero internet connection.
“Teach it client names, code terms, medications, or product names.”
Add proprietary names, code symbols, technical acronyms, or medication terms. HushWrite biases recognition weights to achieve flawless accuracy.
“Private dictation in the languages you actually use.”
Automatic language detection across 99 languages. Dictate in Spanish, German, Japanese, French, or English without cloud transmission or reconfiguring settings.
How HushWrite works in three steps.
No window switching or manual copying. Speak and your thoughts become structured writing.
Press your global shortcut
Tap Option+Space on macOS or Alt+Space on Windows from any focused application to trigger the minimal pill.
Speak at your natural pace
Speak naturally in your preferred tone. Local Whisper models decode audio streaming chunks concurrently on your GPU.
Receive formatted text instantly
Release the shortcut. HushWrite clears filler sounds, structures sentences, and pastes text directly into your document.
Cloud dictation vs. Local dictation.
Cloud tools protect data with privacy policies. HushWrite protects data by keeping your dictation on your device in the first place.
The Local-First Difference: You don't have to trade speed, usability, or smart formatting to get complete air-gapped privacy.
| Capability | HushWriteLocal-First (Free MIT) | Wispr Flow | Superwhisper |
|---|---|---|---|
| Architecture & Data Boundary | 100% On-Device (Zero cloud upload) | Cloud transcription (Audio uploaded) | Hybrid (Cloud for advanced) |
| Privacy Guarantee Mechanism | Physical architecture (Zero bytes sent) | Policy & settings based | Policy based |
| Offline Dictation (Planes & Travel) | |||
| Universal Global Hotkey (Any App) | |||
| Filler Word Stripping & Punctuation | Instant local formatting | Cloud LLM | Local / Cloud |
| Custom Jargon & Phonetic Biasing | Unlimited on-device dictionary | Cloud managed | Pro tier only |
| macOS & Windows Parity | Native macOS & Windows (Metal / DirectML) | macOS & Windows preview | macOS only |
| Latency | Under 180 ms (Local GPU) | 300 to 700 ms (Network ping) | 250 ms |
| Pricing & License | Free & Open Source (MIT) | $144 / year subscription | $200 lifetime |
Calculate your speaking speed dividend.
Human speech flows at ~160 words per minute while average typing tops out at ~40 words per minute.
Select the right engine for your hardware.
HushWrite runs 100% locally on your machine. Choose between quantized Whisper GGML, ultra-fast Parakeet ONNX, and on-device GGUF LLMs for intelligent text rewriting.
Default: Whisper Small (q5_1) · Opt-In: Large v3 Turbo (q5_0) · Fast Tier: Parakeet ONNX
Whisper Small (190 MB) provides the ideal balance of sub-200ms latency and high punctuation accuracy. For specialized vocabulary, toggle Large v3 Turbo (q5_0). For sub-50ms English streaming, enable the Parakeet ONNX fast tier.
Whisper Small (q5_1)
Default RecommendedBuilt-in Defaultwhisper.cpp (GGML)Available NowThe official out-of-the-box default for all users. The ultimate balance of sub-200ms speed, low memory footprint, and high punctuation accuracy.
Why Quantization (q5_0, q8_0) Matters on Windows
Unlike macOS with its unified memory and Apple Neural Engine (ANE), Windows CPUs rely on SIMD vector units (AVX2/AVX-512) and DirectML. 5-bit (q5_0 / q5_1) and 8-bit (q8_0) quantization is your main lever to compress weights by up to 65%, avoiding memory bandwidth stalls and delivering sub-200ms latency without thermal throttling.
Parakeet ONNX Runtime: English Fast Tier
NVIDIA’s parakeet-tdt-0.6b-v2 is the architecture Windows dictation engines have converged on for ultra-fast streaming (sub-50ms). Running via ONNX Runtime with DirectML, it excels at English throughput. For 99-language coverage, Whisper remains the primary engine.
Local LLMs for Smart Cleanup & Voice Commands
Without cloud API roundtrips, HushWrite leverages llama-cpp-2 in GGUF format (Qwen 2.5 1.5B & Phi-3.5 Mini). Quantized to Q4_K_M on CPU, it executes filler removal, code formatting, and voice transformations with zero cloud egress.
Start Speaking. Stop Typing.
Free and fully functional offline. Official Microsoft Trusted Signed Windows installer (.exe) and MSI package with zero friction.
HushWrite for Windows
Full hardware-accelerated local whisper engine. Digitally signed with Microsoft Trusted Signing for instant, safe installation with zero SmartScreen blocks.
Officially code-signed by SnackforCode Aka Alex
As a newly released version, Windows SmartScreen may show a one-time prompt while Microsoft establishes reputation for this fresh build. Click "More info" → "Run anyway" to install.
HushWrite for macOS
The current build is distributed through our official GitHub Releases and requires a one-time macOS approval because it is not yet signed and notarized. We are funding Apple Developer signing from early customer revenue.
If you prefer a signed installation, use the Windows version today or join the Mac waitlist for the signed release.
Requires one-time macOS security approval. Signed/notarized release is planned.
We are funding Apple Developer signing from early customer revenue.
Transparent, Verifiable Releases
Because early-access macOS builds require manual approval, every build is published with strict cryptographic integrity and open provenance.
Expected macOS Permissions Explained
When launching HushWrite for the first time on macOS, the operating system will request the following permissions. Here is why each one is necessary:
1. Microphone Access
Required to capture your voice. Audio streams exclusively into volatile local RAM for whisper.cpp processing and is immediately zeroed upon completion.
2. Accessibility Permissions
Required for text insertion. Enables HushWrite to inject transcribed text directly at your active cursor position in any app without stealing focus.
3. Input Monitoring (Global Hotkey)
Required so HushWrite can listen for your global push-to-talk hotkey (⌥ Space) while running quietly in the background.
What Network Access is Optional vs. Zero
HushWrite has no required user account, no telemetry by default, and zero audio or transcript egress. The only optional network activities are:
- 1.Model Weight Downloads: 1-time download of Whisper GGML model files from official Hugging Face / GitHub mirrors.
- 2.Update Checks: Read-only query to the GitHub Releases API (can be toggled off in Settings).
Frequently asked questions.
Everything you need to know about HushWrite performance, privacy, and local architecture.