SAID

Say it once.
Local voice. Your text. Your words land at the caret—on this PC.

Warning

The v0.3.0 installer and portable build are not Authenticode-signed, so Microsoft Defender SmartScreen may show an unrecognized app warning. Download SAID only from this repository's release page and verify the files against the attached SHA256SUMS.txt before running them.

Warning

The first v0.4.0 macOS preview is ad-hoc signed, not Apple-notarized. After dragging SAID to Applications, Control-click it and choose Open. If macOS still blocks it, use System Settings → Privacy & Security → Open Anyway. Verify the DMG against SHA256SUMS-macos.txt before opening it.

SAID is a lightweight local voice keyboard for Windows 10/11, with a focused first macOS release. Tap your shortcut once, speak in Chinese, English, or both, and tap it again. Your words land at the caret without leaving your computer. Japanese and Korean recognition can be enabled from the tray or menu-bar settings without downloading another model.

On Windows, Right Alt is the default shortcut; F8 is available for AltGr keyboard layouts, and the settings window can record a custom key or modifier combination. The first macOS release uses Control–Option–Space. Recognition runs locally with the multilingual SenseVoice Small model through sherpa-onnx. Its recognition-language whitelist always includes Chinese and English; Japanese and Korean can be added as independent tags. A separate Chinese/English punctuation pass improves Chinese/English readability, while enabled Japanese/Korean output bypasses that pass and Chinese simplification so its writing is preserved.

Exact keeps the allowed-language recognizer transcript unchanged. Clean is the recommended default and uses bundled English/Chinese rules to fix high-confidence speech mistakes, fillers, punctuation, repetition, and explicit self-corrections—no language model or extra download. Windows also offers Adapt, which applies Clean and then uses an optional Qwen3 0.6B model through llama.cpp to organize the thought for the current application. SAID discloses the exact 639 MB download and about 850 MB required free space before downloading it. Adapt works offline after installation and can be removed from the tray menu. If the model is absent or its result fails preservation checks, SAID keeps the complete Clean text. Neither audio nor text is sent to a cloud service.

On Windows, Type while I speak is available from Setup & settings and the tray menu and is off by default for every installation. When enabled, SAID uses its existing local VAD and SenseVoice model to settle short phrases after a natural pause. Exact types each stable phrase as recognized. Clean repairs and replaces only SAID's current revisable tail while speech continues. Adapt runs on a separate coalescing worker and may revise that same tail; full-document organization waits until dictation ends. Streaming adds no second speech recognizer or cloud service.

User experience

The full Windows workflow is:

1. Put the caret in any editable field.
2. Tap your SAID shortcut. A compact status instrument appears without taking focus.
3. Speak naturally. The waveform follows the live microphone level. With streaming enabled, finalized phrases appear at the caret as you pause.
4. Tap the shortcut again. SAID settles the last phrase. Adapt may run one bounded whole-thought pass; while it runs, tap the shortcut once more to keep the complete Clean draft immediately.

On first launch, SAID runs a short microphone check, lets you choose Right Alt, F8, or a recorded custom shortcut, lets you rehearse it, and presents Type while I speak, the Chinese/English language whitelist with optional Japanese/Korean tags, Exact / Clean / Adapt, and startup choices before the first dictation. Double-click the SAID tray icon at any time to reopen Setup & settings. Its permanent section navigation jumps directly to Welcome, Microphone, Shortcut, or Dictation without stepping through the other pages; changes save automatically. Language changes take effect on the next dictation. The tray menu exposes the same output choices, Adapt download cancellation/removal, and per-application behavior overrides.

If focus changes before processing finishes, SAID copies the result instead of typing into the wrong app. During live typing, any manual typing, pointer click, focus change, injection failure, or uncertain caret state permanently revokes ownership for that dictation; already visible text is left untouched and the final result is copied. Text that was already in the field is never sent to the rewrite model and never becomes part of SAID's replaceable range. Clean is immediate and model-free. The optional Adapt model loads lazily and stays warm. Preservation checks reject prompt scaffolding, answers to dictated instructions, information loss, changed names/numbers/negation, and missing technical tokens; rejected output falls back to Clean.

Install on Windows

Run SAID-Setup-0.3.0.exe. The per-user installer does not need administrator access. It installs SAID, the local speech bundle, and Clean, adds Start Menu and uninstall entries, optionally creates a desktop shortcut, and can keep SAID ready after sign-in. The optional 639 MB Adapt model is not part of the installer and is downloaded only after explicit confirmation.

The portable SAID-windows-x64-0.3.0.zip is available for users who do not want an installed app.

Install on macOS

Open SAID-macos-universal-0.4.0.dmg, drag SAID to Applications, and Control-click it to choose Open on first launch. SAID appears in the menu bar rather than the Dock. Grant Microphone permission for recording and Accessibility permission so SAID can paste the local transcript into the app at the caret. Press Control–Option–Space once to start and once to stop. If Accessibility is not granted, the transcript is still copied to the clipboard.

The first macOS release deliberately covers the dependable one-shot path: local transcription, Exact and Clean output, Chinese/English by default, and optional Japanese/Korean. The Windows-only setup window, live phrase typing, Adapt model, application profiles, startup toggle, and custom shortcut recorder are not yet included in the macOS app.

Build on Windows

Prerequisites: Visual Studio 2022 Build Tools with Desktop C++, CMake 3.21+, Git, PowerShell 5+, and NSIS for the installer.

Set-ExecutionPolicy -Scope Process Bypass
./scripts/build-windows.ps1

The script builds said.exe, runs the contract, incremental-refinement, and performance tests, downloads and verifies the four-file speech bundle, creates the portable ZIP, and builds the monochrome NSIS installer. Pass -TestAdaptModel (the old -TestAdvancedGrammar name remains an alias) to additionally download the pinned Qwen3 model and run Clean, application-aware Adapt, bilingual content-loss, prompt-preservation, and warm-latency probes; that model remains in the build cache and is never copied into release packages. Pass -SkipInstaller only when NSIS is unavailable.

Run the unpacked app with:

./dist/SAID/said.exe

SAID is a tray app. Right-click its tray icon for status, Setup & settings, the model folder, or Quit SAID. It does not require administrator privileges; Windows prevents normal apps from injecting text into elevated windows.

Build the macOS DMG

Prerequisites: macOS 12 or later, Xcode Command Line Tools, CMake 3.21+, Git, curl, and roughly 3 GB of free build space.

./scripts/build-macos.sh

The script builds and tests a universal Apple Silicon/Intel SAID.app, downloads and SHA-256-verifies the same four-file speech bundle as Windows, ad-hoc signs local builds, and writes dist/SAID-macos-universal-0.4.0.dmg plus dist/SHA256SUMS-macos.txt. Set SAID_MACOS_ARCHITECTURES=arm64 for an Apple-Silicon-only local build.

For a public Gatekeeper-ready build, set APPLE_SIGNING_IDENTITY to a Developer ID Application identity. Notarization uses either an existing APPLE_NOTARY_PROFILE, or APPLE_ID, APPLE_TEAM_ID, and APPLE_APP_PASSWORD. The macOS GitHub workflow supports the corresponding repository secrets and imports the certificate from APPLE_CERTIFICATE_BASE64 using APPLE_CERTIFICATE_PASSWORD and APPLE_KEYCHAIN_PASSWORD. It attaches a DMG to a tagged release only after a signed, notarized build succeeds.

Manual CMake build

cmake -S . -B build -A x64 -DCMAKE_BUILD_TYPE=Release
cmake --build build --config Release --parallel

For an offline source build with an existing sherpa-onnx checkout:

cmake -S . -B build -A x64 `
  -DSAID_SHERPA_ONNX_SOURCE_DIR=C:/src/sherpa-onnx

The model is searched in this order:

1. --model C:\path\to\sense-voice-small.int8.onnx
2. %SAID_MODEL%
3. The legacy %VOICEKEY_MODEL% override for transition installs
4. models\sense-voice-small.int8.onnx beside said.exe
5. %LOCALAPPDATA%\SAID\models\sense-voice-small.int8.onnx
6. Legacy local and installed model folders

The selected model directory must also contain sense-voice-small.tokens.txt, ct-transformer-punctuation.int8.onnx, and silero-vad.onnx. SAID rejects partial bundles at startup instead of failing during dictation.

The optional Adapt model is searched through the legacy-compatible --grammar-model and %SAID_GRAMMAR_MODEL% overrides, beside the selected speech model, beside said.exe, and in %LOCALAPPDATA%\SAID\models. Its filename is Qwen3-0.6B-Q8_0.gguf. SAID downloads it with Windows Background Intelligent Transfer Service, verifies SHA-256 9465e63a22add5354d9bb4b99e90117043c7124007664907259bd16d043bb031, then activates it atomically. Downloads can be cancelled and resume after an app or Windows restart. A missing or rejected Adapt result never blocks dictation: SAID keeps Clean.

The Visual Studio build remains the supported release build. Public artifacts should be Authenticode-signed; unsigned local builds can trigger Microsoft Defender SmartScreen reputation warnings.

Existing installations

The transition release reads current SAID settings first and then falls back to the pre-rebrand registry key, model environment override, and model folders. It replaces the old startup entry with SAID, and the compatible single-instance guard prevents the old and new executables from running together. Old Whisper .bin files are not compatible with the new recognizer and are left untouched. The uninstaller deliberately leaves user model files in %LOCALAPPDATA%\SAID\models.

Hardware requirements and benchmark

SAID runs recognition entirely on the CPU; it does not require a dedicated GPU or NPU. These recommendations include enough headroom to keep an editor, browser, and normal development tools open while dictating.

	Practical minimum	Recommended for coding
Operating system	Windows 10 or 11, 64-bit	Windows 11, 64-bit
Processor	4-core x64 CPU	6 or more modern CPU cores / 12 or more logical processors
Memory	8 GB RAM	16 GB RAM; 32 GB for containers, virtual machines, or large local builds
Storage	SSD with 2 GB free during setup	SSD with 3 GB free for SAID, updates, and the development toolchain
Graphics	Integrated graphics are sufficient	No dedicated GPU is needed

The base installation occupies about 360 MB. Adapt adds a 639 MB download and requires about 850 MB free for download verification and atomic activation. Before Adapt is first used, the reference run used about 564 MB of memory while idle and reached a 605 MB high-water mark during startup. Qwen3 is lazy-loaded only when Adapt runs; Exact and Clean never load it.

Once Adapt is warm, its local rewrite engine uses about 1.10 GiB of resident memory: approximately 612 MiB of model-backed pages and 515 MiB of private working memory. Combined with the speech stack, expect SAID to use about 1.65 GiB while Adapt remains warm. Active Adapt inference averaged 5.6–6.7 logical CPU cores on the 32-thread reference machine, then returned to an inactive worker between requests. 16 GB RAM is recommended for normal Adapt use; use 32 GB when running containers, virtual machines, or large local builds beside SAID. See the Adapt resource profile for the measurement method, environment, and update procedure.

Reference performance

The current Release x64 build was measured on an Intel Core i9-14900HX with 32 logical processors available to the benchmark, 16 GB RAM, and SSD storage. Each result is the median of three new-process runs against the same 30.42-second bilingual recording; the model and file data were already in the operating-system cache.

Recognition threads	Model load	Recognition	Speed vs. recording	Peak memory
1	1.34 s	2.18 s	14x real time	601 MB
2	1.24 s	1.29 s	24x real time	602 MB
4	1.26 s	0.86 s	35x real time	609 MB
8 (automatic on this CPU)	1.29 s	0.95 s	32x real time	610 MB

For shorter samples, 5.59 seconds of Mandarin took 0.17 seconds and 7.15 seconds of English took 0.19 seconds after model loading. SAID keeps the model loaded, so the model-load cost normally occurs once when the app starts rather than after every dictation.

The automated development environment cannot provide native Windows timing, so these results are directional measurements of the Windows x64 executable under a compatibility layer, not a hardware certification. The practical minimum is deliberately conservative; native Windows results may vary with CPU generation, power mode, microphone driver, and other work running at the same time.

Runtime profile

• Native Win32 UI on Windows and a native AppKit menu-bar app on macOS; no Electron or browser runtime.
• The CPU-only SenseVoice, punctuation, and VAD models remain loaded after startup. Clean has no model. Qwen3 loads only for Adapt and remains warm afterward.
• Live typing uses VAD to decode each completed phrase once. Recognition never waits for Adapt; pending model work is coalesced to the newest revision, and only the most recent two phrases or 320 Unicode code points remain mutable.
• Clean uses deterministic high-confidence rules. Adapt uses deterministic decoding plus fail-closed checks for technical tokens, names, numbers, negation, CJK content coverage, shell syntax, and model scaffolding. Neither mode guarantees perfect prose; rejected or ambiguous changes keep Clean.
• SenseVoice uses non-autoregressive decoding with automatic language selection. SAID keeps Chinese/English enabled for code switching, rejects a phrase positively tagged as disabled Japanese/Korean, and removes residual disabled Japanese/Korean script before punctuation.
• Dictations up to 25 seconds are recognized intact. Longer audio is split at speech boundaries into segments no longer than 20 seconds.
• Inference uses up to half the logical processors, capped at eight threads.
• Audio capture is capped at ten minutes per dictation.

Current limitations

• Some elevated or unusually sandboxed applications can reject Windows Unicode input; SAID copies the transcript in that case.
• Users with AltGr keyboard layouts should select F8 during setup so Right Alt does not intercept AltGr characters.
• Very short English insertions inside Mandarin—especially uncommon names and one-syllable words—remain the hardest recognition case.
• SenseVoice accepts one language prompt rather than a native multi-language set, so the whitelist is a fail-safe output boundary around automatic recognition, not decoder retraining. Chinese and Japanese share Han characters; the detected SenseVoice language tag provides the phrase-level distinction.
• The generic Qwen3 0.6B candidate meets the size and warm-latency target but is conservative and is not equivalent to a larger or task-distilled writing model; Chinese whole-thought rewrites commonly fall back to Clean. Use Clean for predictable correction or Exact for quotations, legal text, code, shell input, and any task that requires post-ASR wording unchanged.
• SenseVoice is an offline recognizer, so streaming updates arrive phrase by phrase after a natural pause or six-second speech boundary rather than as unstable token-by-token hypotheses.
• The first macOS release supports one-shot dictation only; live typing, Adapt, per-app behavior, login startup, and shortcut customization remain Windows-only.

See Mixed Chinese/English ASR investigation for the benchmark and design rationale.