Saydraft is a free, open-source offline dictation app for Windows. Local Parakeet/Whisper speech-to-text and on-device cleanup keep processing on your PC.
Speak messy. Paste polished.
Desktop UI v2: implementation notes, with a dedicated redesign reference and searchable screen gallery. The earlier app-ui/ snapshot is retained unchanged.
The latest repairs prepare speech recognition before the first shortcut, restore tray clicks, improve tray artwork, and separate optional browser text reads from the required insertion checks. First-run Models now includes optional local AI cleanup. Changes and verification limits. Control focus rings appear during keyboard navigation; mouse clicks do not leave them behind. The current portable also fixes a frozen Windows field-reader initialization failure; packaging now checks the real reader. App/tray icons use the landing-page star. Current unsigned build and verification.
View the Bar animation (5.76-second loop)
Public sample animation, including optional English live preview. No recording or model inference. Collapse this disclosure to hide motion.
The static image is shown first; animation is an explicit choice.
See minimum preview acceptance for the current unsigned portable and verification scope, and the delivery handover for known issues and future work.
Download Saydraft for Windows — unsigned preview
Development preview · 0.1.1. Download the Windows x64 portable ZIP, extract the whole folder and run Saydraft.exe. Quit an older instance from the tray first. The build is unsigned; Windows may show an unknown-publisher prompt. Release notes and checksums include the exact source commit and SHA-256. Models are installed explicitly during setup. Broad accuracy and physical Windows acceptance remain separately qualified.
Press F10 to start/stop a voice note, or use Record note in Hub → Notes. Notes save locally with a History recovery copy and preserve your open Scratchpad. This preview also repairs literal-comma cleanup regressions and CPU cleanup startup/admission. Verification and limits.
Hold a shortcut, speak, then release to transcribe, clean up and attempt delivery into the original editable field. History saves the result before delivery. Installed speech and cleanup models run locally, with no cloud fallback or subscription.
Highlighted speech languages, in alphabetical order: Arabic, Bengali, Brazilian Portuguese, Chinese (Mandarin), Filipino / Tagalog, French, German, Hindi, Indonesian, Japanese, Korean, Malay, Persian (Farsi), Spanish, Tamil, Telugu, Thai, Urdu and Vietnamese, with compatible local models. English is also supported. See the language guide for coverage, shared-model setup and measured accuracy limits.
Website · User guides · Documentation · Contributing
Explore the offline dictation guide, developer workflows and dated Wispr Flow comparison. The website also covers Typeless, Superwhisper, Windows voice typing and Dragon; its changelog lists published releases only. The Home page includes actual Hub/Bar screenshots, source-linked comparisons and alphabetical language guides. Notes guidance separates local Scratchpad notes from experimental meeting features. See the Home implementation and verification for capture provenance and limits. The prepared-example demo is included in the current site source. It uses canned text and no microphone or uploads; demo documentation describes its behavior and publication limits.
Install from source · First launch · Models and languages · Daily use · Troubleshooting
Start with Ctrl + Win to hold-to-speak, Ctrl + Win + Space for hands-free capture, and Shift + Alt + Z to paste your last words. Review and rebind shortcuts in Settings if another app uses them. Install compatible models once, choose tone/cleanup per app group, then review a short sample in your own workflow.
See the source-backed Wispr Flow migration guide and Typeless comparison. These comparisons do not establish equivalent accuracy, speed or complete parity.
Product facts checked 2026-11-14; features and plans can change. Sources are linked in every cell. This is a development preview with qualification limits, not an accuracy or speed ranking.
| Feature | Saydraft | Wispr Flow | Typeless | Windows voice typing |
|---|---|---|---|---|
| Dictation connection | Offline with compatible models installed. Project privacy guide | Internet required for transcription. Flow documentation | Internet required for transcription. Typeless FAQ | Standard Win+H uses online recognition. Microsoft documentation |
| Windows availability | Windows 10/11 development source. Project setup status | Windows app. Flow platforms | Windows app. Typeless platforms | Built into Windows. Microsoft feature guide |
| Free use | Free, open source; no subscription quota. Project license and limits | Limited weekly words; paid unlimited plan. Flow plans | Limited weekly words; paid unlimited plan. Typeless plans | Included with Windows. Microsoft feature guide |
This compares standard online Win+H voice typing. Copilot+ Fluid enhancements and the separate offline Voice Access feature have different requirements. No affiliation with Wispr Flow, Typeless or Microsoft is implied. Saydraft has a configurable per-dictation recording cap, 20 minutes by default, separate from its lack of a subscription word quota.
Does setup need internet? Yes: review model sizes/licenses and start a Download/Resume action. Later launches reuse installed compatible models offline. Selecting a language starts no transfer.
Are Bengali, Arabic and the other languages separate downloads? No. Languages outside the default model's coverage share the optional All languages model; there are no Bengali-only or Arabic-only packs. See language setup.
Is accuracy the same in every language? No. Coverage is distinct from measured quality. The public multilingual evaluation reports its methods and substantial Bengali, Tamil and Telugu errors. Review names, numbers and your own speech before relying on results; owner accuracy acceptance is deferred.
Is it completely offline? Installed recognition and cleanup run locally. Model downloads are explicit; optional GitHub update notices are off by default. The privacy boundary explains Python guard and native-helper firewall limits, as well as Windows clipboard history/sync.
Does it work in every text field? Delivery is guarded. Protected/elevated fields, unsupported selections or focus changes can refuse insertion. Review the saved History result and delivery recovery.
Are Command, Translate, Ask and meeting notes ready for production? They are experimental Pro hardware-profile features with explicit setup/opt-in and separate qualification limits. Pro is a hardware tier, not a paid subscription. See their implementation guides.
What can I install today? Use the unsigned Windows 10/11 x64 portable preview. Extract the whole ZIP and run Saydraft.exe; its Python runtime is included. Developers can use the source instructions with Python 3.13.
Does the web demo record me? No. The current site's prepared-example demo uses canned text, no microphone, models, clipboard or text insertion. Hosted publication is a separate step from a source push.
Saydraft is in active development. The current unsigned preview
publishes the owner-confirmed working build from 12bc27f, including the frozen
field-reader repair and updated app/tray icons. All 1,440 packaged files and 89
embedded modules verify. The owner reports insertion working; full clean-machine
and all-language qualification remain separate. Build and scope.
Historical packaging verification produced an approximately
104 MB unsigned installer with resource, notice and install/uninstall checks. Its
retained verification
covers 38 components and 506 notice files.
The draft-release verification
retains a historical real dry run; signing and publication remain separate.
Development checks and builds run locally before direct pushes. GitHub Actions workflows are manual-only and are not a delivery gate. See the local verification checklist.
App update checks are off by default. Enable and save them in Settings → Privacy for startup/daily notices and a release-page link. Install updates manually from GitHub.
Packaged builds also offer an optional Windows Firewall block
in Settings → Privacy. Confirm the installation path, then approve Windows UAC.
Remove the rule before downloading models or checking for updates. The rule
covers that Saydraft.exe; separate helper executables need their own coverage.
Source Python is never blocked by these controls.
Core dictation and the Hub are available from source. Owner accuracy acceptance is deferred. Implementation and remaining qualification are tracked separately:
| Available from source | Still in progress |
|---|---|
| Local dictation, rules cleanup, optional local AI cleanup | Current-model quality and release latency qualification |
| Hub, History, Dictionary, Snippets, Style, Scratchpad and Notes | Full Command qualification |
| Experimental Pro meeting notes: reviewed setup, consented Start, speaker transcript and local summary | Live-device, packaged, 30-minute meeting and broader accuracy acceptance |
| Experimental Pro Translate with per-group Style targets | Broader translation quality and physical workflow acceptance |
| Experimental Pro Ask about selected text, with separate answer Notes; native read-only browser workflow verified | Packaged/physical Ask acceptance; broader answer quality |
| Six-page setup, verified downloads, model performance check | Fresh-profile and physical Windows acceptance |
| Public website, setup guide, published unsigned portable and opt-in update notices | Signed installer and package-manager distribution |
Command Mode's selection guard, local edit client, replacement checks and separate Copy recovery windows are implemented. The experimental model now has separate Pro admission and exclusive ownership, preserving cleanup defaults. The experimental Labs setup and held instruction route are connected for readable selections. Read-only results are offered for Copy. The held-flow verification includes full CI/package checks, controlled native worker scenarios and both-theme setup/recovery renders. A separate Labs opt-in now connects clipboard-copy fallback for unsupported safe fields; Windows clipboard history/sync settings apply. It offers Copy when the original selection cannot be verified for replacement. Full Command acceptance remains open; see the implementation status and connected-flow verification.
Experimental Translate while speaking uses the normal dictation shortcut with an English, Spanish or Arabic target saved per app group in Style. Enable it separately in Labs after explicitly installing the shared 4B editing model. Successful translations are saved before guarded delivery; failures preserve recognized speech for Copy without automatically pasting it. Raw and Code bypass translation. Native text smoke checks and controlled worker tests are distinct from physical Windows and broad quality acceptance. See the retained verification for native smoke cases, CI/package results and their scope.
See the documentation index for individual features and the parity plan for the complete acceptance status.
The latest cleanup measurements retain all ten CPU/RTX model comparisons, complete policy replays, and the earlier CPU failure. The CPU policy replay keeps rules throughout, so combined hybrid qualification remains open. Benchmark completion alone does not establish app latency or a model default.
The release-latency matrix now covers the specified 3/10/30/90-second cold/warm runs through the actual worker, History and checked simulated paste, using Whisper base for Lite and Parakeet for Standard/Pro. The full Lite baseline completes all 160 cases, but misses the latency targets: 3-second warm p50/p95 is 1.270/1.330 seconds. Standard/Pro and performance qualification remain open; the earlier warm fixture results retain their original scope.
- Local dictation: Parakeet or Whisper recognition, hold-to-talk and hands-free capture, a floating Bar, cancellation, and History before delivery.
- Cleanup and style: rules for fillers, repetitions, self-corrections, punctuation, and lists; per-app styles; optional local AI cleanup with validation and rules fallback. Normal AI cleanup also checks content-word identity and order after rules; validation details describe the allowed edits and the broader Polished setting. The Hub's Style preview includes staged Dictionary edits and lets you preview a resolved language without changing speech settings.
- One Hub: Home statistics and recent dictations; searchable History with raw/cleaned changes, Markdown/CSV export, and retry when audio was retained.
- Scratchpad and Notes: a floating plain-text draft with local autosave, searchable notes, and revision conflict recovery. A shortcut also opens the last dictation's changes. See the Notes guide.
- Personal vocabulary: searchable Dictionary terms, protected terms, and replacement pairs, with conservative spelling correction in resolved-English dictation; Snippets with date/clipboard variables and import/export. See dictionary behavior and limits.
- Translate while speaking · Labs: optional Pro translation to English, Spanish or Arabic with per-group targets, local generation and Copy recovery. See the setup and limits.
- Guided first launch: microphone selection and a level-only test, focused shortcut learning, chat/email dictation practice with before/after review, optional local AI model setup, and a Style quick pick before opening Home.
- Explicit model setup: provisional PC recommendations, disclosed sizes/licenses, resumable downloads with per-file verification progress, tier overrides, and model storage controls.
- Whisper mode · experimental: optional gain, AGC, local speech enhancement and more sensitive speech detection for quiet speech. Its whispered-speech accuracy target remains unverified. See setup and scope.
- Local data controls: no telemetry, audio retention off by default, optional retention limits, a read-only firewall-rule check, and explicit firewall controls for packaged installations.
Use Windows 10 or 11, 64-bit Python 3.13, Git, a microphone, and enough free disk space for your selected models. In PowerShell:
git clone https://github.com/tanzir71/saydraft.git
cd saydraft
py -3.13 -m venv .venv
.venv\Scripts\python -m pip install -r requirements.txt
.venv\Scripts\python -m saydraftKeep the console available while diagnosing startup problems. The current download page also links to GitHub Releases. The current build is a portable ZIP; source instructions above are optional for developers.
The setup wizard has six pages: Welcome, Microphone, Set up models, Shortcut, Try it, and Style quick pick.
- In Microphone, choose Default Windows input or select Refresh inputs to find a named device. Test input runs a ten-second, level-only check; it does not transcribe or save a recording. Check microphone privacy reads Windows permission status without opening an input; it provides guidance when access is blocked or unknown.
- In Set up models, optionally select Check this PC, then Use speech recommendation. This is a provisional capability recommendation, not a measured performance tier. It does not start a download or enable AI cleanup.
- Review the speech model, size, and licenses, then select Download / Resume. Setup shows package progress and individual file transfer/checksum phases. Installed-file checks run in the background, including pinned dependencies. After installation, Check model performance measures the selected speech model using bundled audio. It opens no microphone, downloads nothing, and keeps successful measurements separately from saved setup preferences.
- In Shortcut, keep the preset, enter a binding, or choose Learn shortcut and press/release your keys. Review any conflicts before continuing.
- In Try it, practice in the window's own chat or email field. Preparation opens no microphone. Hold your shortcut or choose Record, then release or choose Stop and preview to review the transcript, cleaned text, and edit chips. Each practice recording is limited to one minute. Audio stays in memory; practice creates no History entries and uses no external paste or clipboard. You can continue without recording.
- In Style quick pick, choose Casual or Formal for Messages and Email. The rules preview uses your last practice transcript or a labeled example. Finish rechecks the selected model, saves changed choices, and opens Home after setup work has stopped. Existing intensity and AI settings stay unchanged. Cancel preserves saved preferences; completed downloads and resumable partials remain separate.
In Set up models, Next requires a ready speech model. Catalog packages must pass their file checks; an existing local Whisper model needs a successful offline load check through Check model performance. Finish checks the current model again before saving. Cancel, hiding setup, or going Back revokes a pending Finish. Capability recommendations and measured speed remain distinct. Fresh-profile manual acceptance is still open. See the first-run guide.
At least one compatible speech model is required to dictate. Model weights are downloaded during explicit setup rather than bundled with the source. Review the recommended speech model, its download size, and its license, then select Download. Progress and pause/resume are available. Later launches reuse installed files. Choosing a language or model never starts a download by itself; you need an installed compatible model before dictating.
| Speech option | Approximate download | Purpose |
|---|---|---|
| Default: Parakeet v3 int8 + language detector | 773.4 MB | Recognition in 25 European languages |
| Parakeet v2 int8, English max | 661.3 MB | English recognition |
| All languages · Whisper large-v3-turbo | 1,621.7 MB | Thai, Chinese, Korean, Bengali, Arabic and other supported languages |
| Whisper base / small / medium | Shown in setup | Alternative multilingual models |
Thai, Chinese, Korean, Bengali and Arabic share the optional All languages
model. Download it once; there are no separate language packs or language-specific
download buttons. Portuguese, including Brazilian Portuguese, is also covered by
the default Parakeet v3 model. Choose Portuguese (including Brazilian Portuguese)
in Languages; both engines use pt, without a separate Brazil-only decoder setting.
Installed models can be selected through language pins and automatic routing;
language coverage and routing details
explain setup and its limits. Language coverage is not an accent-specific accuracy claim.
Optional AI cleanup is disabled by default. Explicit setup offers pinned Qwen3.5 0.8B or 2B Q4_K_M weights with a local llama.cpp CPU or Vulkan runtime. Recommendations depend on the selected tier and available measurements; they do not enable or download AI cleanup. Rules-only dictation needs no cleanup model. See model details, cleanup setup, and storage controls. Model weights and dependencies retain their own licenses.
- Click the field where you want text.
- Hold Ctrl + Win, speak, then release.
- Saydraft transcribes and cleans locally, saves History, and attempts delivery into the original target. Review History if delivery is skipped or fails.
Fresh profiles use these shortcuts. Existing custom bindings are preserved; change them in Hub → Settings → Shortcuts.
| Shortcut | Action |
|---|---|
| Ctrl + Win | Hold to dictate |
| F9 | Alternate hold-to-talk shortcut |
| XButton1 | Configurable mouse hold-to-talk shortcut |
| Ctrl + Win + Space | Toggle hands-free dictation |
| Shift + Alt + Z | Paste the last dictation |
| Win + Alt + S | Open Scratchpad |
| Ctrl + Win + D | View the last dictation's changes |
| F10 | Start/stop a voice note; saves to Notes and History |
| Esc | Cancel the current recording |
Two short taps of a recording shortcut also lock hands-free mode; a single short tap does not. The default recording cap is 20 minutes, with a warning at 19 minutes. Saved custom limits stay unchanged. See shortcut behavior and recording limits.
The dictation path is microphone → local recognition → cleanup → History → delivery. Models are fetched only through explicit setup actions. Python network access is restricted to loopback, requested model downloads and saved opt-in GitHub update checks. Packaged firewall controls add an OS rule for that executable, including its native libraries. Effective Windows policy and separate helper processes have their own verification limits. See the privacy boundary and Verify offline.
| Data | Default location |
|---|---|
| Configuration | %LOCALAPPDATA%\Saydraft\config.json |
| History and Notes | %LOCALAPPDATA%\Saydraft\history.db |
| Logs | %LOCALAPPDATA%\Saydraft\logs\saydraft.log |
| Audio, only when retention is enabled | %LOCALAPPDATA%\Saydraft\audio |
| New model installations | %LOCALAPPDATA%\Saydraft\models\<id>\<version> |
Audio retention is off by default. You can enable Keep forever or Keep N days; expiry preserves History text and statistics. Retry needs retained audio. Compatible legacy model installations remain readable in place. See audio retention and data migration.
- No input: check Windows microphone permissions, then open Settings → Microphone, refresh inputs, and run Test input. A missing saved microphone does not silently switch to another device.
- Next is unavailable in Set up models: finish installing the selected package and its dependencies. For an existing local Whisper model, run Check model performance. If files change later, return to this step and check them again before finishing setup.
- Practice is unavailable: return to Set up models to install the selected speech model, or to Microphone to check the input. After an app stop or Settings change, choose Prepare practice to start again. If setup cannot confirm that practice stopped, cancel setup and restart Saydraft as prompted.
- Shortcut conflict: review Settings → Shortcuts and rebind conflicting combinations. Another dictation app may use Ctrl + Win.
- Text did not insert: open History to review/copy the saved result. Focus changes, protected fields, and elevated targets can prevent delivery; see delivery behavior.
- Slow setup or recognition: review the model size and installed files in Settings → Models, and use the performance check before selecting a tier.
- AI cleanup kept the rules result: check that AI cleanup is enabled and its model/runtime are ready. Low-need text, insufficient time, or a rejected model reply can all keep rules output. See the cleanup gate and validator before treating fallback as a failure.
For development setup and checks, read CONTRIBUTING.md. Documentation links to implementation notes, evaluation methods, Windows acceptance checks, and the website workflow. The website component guide covers 26 component groups, 94 Storybook states and responsive visual checks. The owner-supplied landing v2 design defines the current homepage. The documentation guide covers the dated user-guide routes, shared model-size tables and local content verification. Website changes are validated locally and pushed; hosted publication is a separate step. The SEO guide covers per-page social images, sitemap/readable exports, owner verification tags and the remaining rich-result eligibility limits. Report reproducible problems in GitHub Issues with redacted logs and your source version; keep private recordings and transcripts out of public reports.
The automated gates cover the offline test suite, 556 cleanup golden cases, and 100 authored adversarial validator cases. These are distinct from native model measurements and owner accuracy acceptance. The retained native baseline exposed a word-substitution regression that is now guarded; it predates that fix and is not current qualification. See cleanup evidence for the source and corpus checks required of replacement reports.
Saydraft's code is licensed under MIT. Models and dependencies keep their own licenses, documented in third-party notices. Thanks to NVIDIA Parakeet, Whisper/faster-whisper, sherpa-onnx, Qwen, llama.cpp and PySide6. The bundled fonts keep their OFL notices. Built by Tanzir Rahman. You can support Saydraft.
