PrimeDictate

A locally hosted, global hotkey dictation utility for fast desktop workflows.

It captures the default microphone, shows live local transcription in a small overlay, and types the final transcript into the current application after a silence auto-commit or manual stop using SharpHook (no synthetic paste, no clipboard round-trip on the hot path).

Features

• Global hotkeys: Configurable global shortcuts for start/stop toggle (default Ctrl+Shift+Space), emergency stop/discard (default Ctrl+Shift+Enter), and transcript history (default Ctrl+Shift+H).
• Voice commands: Optional phrases can be detected while dictating. Defaults are thank you to commit the active capture, potato farmer to discard it, and show me the money to open history; command phrases are removed before final text injection. Custom computer commands can run local command prompt actions and either stop the commit or continue with the remaining dictated text.
• Tray workspace UI: Open Workspace from the tray icon to browse per-session dictation threads and global runtime logs in a clearer, column-based dashboard layout.
• AI Prompt Modes (Ollama Integration): Automatically rewrite and format your transcripts using a local Ollama instance. Includes dynamic prompt modes (like Bug, Update, and Blog) and injects the active application's context into the LLM for application-aware output.
• Log signal over noise: Repeated adjacent log entries are collapsed (for example (... x12)) and history is capped to keep memory usage predictable.
• Live preview overlay: While recording, the app periodically re-transcribes the growing buffer with the selected local backend and shows the current hypothesis in a non-activating overlay.
• Compact mic overlay: The default overlay mode keeps a small lower-right microphone visible as a ready/listening indicator. Clicking it temporarily expands the larger transcript panel without changing your saved default mode.

• Silence auto-commit: When speech has stopped for the configured delay (default 3 seconds), PrimeDictate stops capture, runs a final transcription pass, and sends the final text once. Set the delay to 0 to commit only with the start/stop hotkey.
• Transcript history: Every committed transcript is saved to local history so you can review past dictations, recover text sent to the wrong app, and copy transcript text (with or without metadata).
• Impact stats and achievements: Successful dictations update local-only productivity stats, including words typed, estimated net time saved, average speaking WPM, last-14-day bars, and milestone notifications.
• History filters and detail view: History includes a filter dropdown (All, Injected, NotInjected) plus an expanded detail pane for full transcript and target metadata.
• History entry points: Open history from the tray menu, the history hotkey, voice command, the settings window, or the workspace toolbar.
• Final-only target typing: The target editor is not mutated while recording. Live corrections stay in the overlay so code editors and IntelliSense are not fighting backspace/retype updates.
• Coding mode: Optional setting sends an Enter key immediately after a successful transcript commit.
• Foreground guard: The foreground window is captured when recording starts; by default PrimeDictate still skips injection if focus changes before the transcript is ready.
• Return to original target (optional): A dictation setting can deliver the final transcript back to the window that had focus when recording started, first trying a safe direct write to the captured edit control on Windows and otherwise reactivating that window before typing.
• Built-in pointer cue: If Windows Mouse Sonar is enabled, PrimeDictate pulses it on recording/processing transitions by tapping Ctrl. It does not draw a custom pointer overlay or change the user's Windows setting.
• Custom audio earcons: PrimeDictate can play its own short start/stop tones so you hear when recording begins and when capture hands off to transcription.
• Launch at login: Installers enable automatic startup by default with a Windows Startup-folder shortcut so PrimeDictate is ready after a reboot. Silent MSI and Chocolatey installs can opt out, and Settings can switch between off, current-user startup, and all-users startup.
• Built-in updates: The tray app can check GitHub Releases directly, prompt when a newer release is available, download the matching x64/ARM64 MSI, verify its .sha256 sidecar, and hand off to Windows Installer.
• Audio: Windows default capture device via NAudio WASAPI (WasapiCapture), resampled to 16 kHz, 16-bit, mono PCM for local transcription engines.
• Mic isolation mode (best effort): Optional exclusive-capture setting can block other apps from the mic on supported devices; if exclusive capture fails, PrimeDictate automatically falls back to shared mode and continues dictation.
• Inference: A shared transcription engine abstraction with Whisper, Parakeet, and Moonshine ONNX models through sherpa-onnx, Whisper.net GGML for GPU/NPU-capable local Whisper runs, and experimental Qualcomm AI Hub Whisper packages for Snapdragon X NPU runs.
• Injection: SharpHook EventSimulator for Unicode text entry (no synthetic paste, no clipboard round-trip on the hot path).

Requirements

• .NET 8 SDK (or a compatible SDK that can build net8.0 projects).
• Windows is the primary target (WASAPI capture path). Other platforms may require a different capture implementation.
• The transcription backend uses local model files only. sherpa-onnx backends require ONNX model folders with tokens; Whisper.net uses GGML .bin files and optional OpenVINO sidecars.

---

🛠️ Local transcription backends and model files

Local transcription backends and model files

PrimeDictate includes a curated backend + model picker during first-run setup and in Settings.

Whisper

The built-in Whisper catalog currently includes:

Model	Typical use
Tiny English ONNX	Recommended fastest setup and lowest CPU cost for English dictation
Base English ONNX	Recommended balanced English dictation on CPU
Distil Small English ONNX	Faster distilled English model for longer sessions
Small English ONNX	Higher English accuracy with more compute cost
Tiny/Base/Small Multilingual ONNX	Non-English dictation with the same ONNX folder layout

When you download a Whisper model inside the app, PrimeDictate stores it under %LocalAppData%\PrimeDictate\models\whisper and saves the exact folder path in settings.

PrimeDictate expects a Whisper ONNX model folder containing:

• *-encoder.int8.onnx or *-encoder.onnx
• *-decoder.int8.onnx or *-decoder.onnx
• *-tokens.txt

The runtime checks managed downloads, installed app-local models under models\whisper, and repo-local models\whisper folders during development.

Whisper.net (GGML)

PrimeDictate also supports Whisper.net GGML models for larger Whisper variants and hardware acceleration. The compute picker only shows locally supported choices:

• CPU is always available.
• GPU is shown when a supported Whisper.net GPU runtime is present and the machine exposes CUDA or Vulkan.
• NPU is shown for supported GGML models with OpenVINO encoder sidecars.

When an older saved setting asks for an unsupported hardware path, PrimeDictate normalizes it at startup. If an installed Whisper.net hardware configuration is available, it prefers that; otherwise it falls back to CPU rather than trying an unsupported provider.

Parakeet

PrimeDictate also supports Parakeet ONNX via sherpa-onnx. The current catalog starts with:

Model	Typical use
Parakeet TDT 0.6B v3	Try a newer local English STT backend without leaving the PrimeDictate workflow

Downloaded Parakeet models are stored under %LocalAppData%\PrimeDictate\models\parakeet. PrimeDictate expects a model folder containing:

• encoder.int8.onnx
• decoder.int8.onnx
• joiner.int8.onnx
• tokens.txt

You can either download the managed Parakeet model in-app or browse to an existing extracted model folder.

Moonshine

PrimeDictate also supports Moonshine ONNX via sherpa-onnx for another lightweight local English path. The current catalog includes:

Model	Typical use
Moonshine Base (English)	Fast local English dictation when you want a smaller non-Whisper backend than Parakeet

Downloaded Moonshine models are stored under %LocalAppData%\PrimeDictate\models\moonshine. PrimeDictate expects a model folder containing:

• preprocess.onnx
• encode.int8.onnx
• uncached_decode.int8.onnx
• cached_decode.int8.onnx
• tokens.txt

You can either download the managed Moonshine model in-app or browse to an existing extracted model folder.

Qualcomm QNN (Experimental)

PrimeDictate includes an experimental Qualcomm QNN backend for native Windows ARM64 builds. The primary path uses Qualcomm AI Hub Whisper packages compiled for Snapdragon X devices. PrimeDictate installs the matching precompiled_qnn_onnx package, because the raw qnn_context_binary ZIP contains only encoder.bin and decoder.bin; ONNX Runtime needs the small encoder.onnx and decoder.onnx EPContext wrapper files to load those context binaries.

The first catalog entry is:

Model	Target	Package used by PrimeDictate
Qualcomm AI Hub Whisper Small (Snapdragon X Elite)	Snapdragon X Elite / X Plus class NPU	whisper_small-precompiled_qnn_onnx-float-qualcomm_snapdragon_x_elite.zip

The backend can attempt QNN HTP / Qualcomm NPU execution when the following are all true:

• the process is running as Windows ARM64
• the build contains the QNN runtime assets (QnnHtp.dll, onnxruntime_providers_qnn.dll, QnnSystem.dll)
• the selected Qualcomm AI Hub Whisper folder is present under %LocalAppData%\PrimeDictate\models\qualcomm-aihub-whisper
• the folder contains encoder.onnx, decoder.onnx, encoder_qairt_context.bin, decoder_qairt_context.bin, metadata.json, and multilingual.tiktoken
• the EPContext sessions can be created with CPU fallback disabled

If you browse to the raw qnn_context_binary package from Qualcomm AI Hub, PrimeDictate detects it and asks you to use the catalog download instead. That raw package is useful source material, but it is not directly runnable by ONNX Runtime without wrappers.

Maintainer knobs for the experimental Qualcomm backend:

• PRIMEDICTATE_QNN_STRICT=1: require QNN HTP session creation and inference with CPU fallback disabled
• PRIMEDICTATE_QNN_CONTEXT_CACHE=0|1: disable or enable QNN context caching (default 1)
• PRIMEDICTATE_QNN_CONTEXT_EMBED=0|1: control QNN context embed mode (default 1)
• PRIMEDICTATE_QNN_PROFILE=basic|detailed|optrace: optional QNN profiling, off by default
• PRIMEDICTATE_QNN_PROFILE_DIR=<path>: where profiling CSV output should be written

Important: Qualcomm QNN support is validated per model package and per device. PrimeDictate can prove strict QNN HTP activation only when the selected EPContext sessions actually load on QNN with CPU fallback disabled.

Example (PowerShell, from repo root), downloading the default bundled model manually:

New-Item -ItemType Directory -Force -Path "models\whisper" | Out-Null
Invoke-WebRequest -Uri "https://github.com/k2-fsa/sherpa-onnx/releases/download/asr-models/sherpa-onnx-whisper-base.en.tar.bz2" -OutFile "models\whisper\sherpa-onnx-whisper-base.en.tar.bz2"
tar -xjf "models\whisper\sherpa-onnx-whisper-base.en.tar.bz2" -C "models\whisper"

Larger models take longer to download and load. The first transcription after launch may still do noticeable disk I/O while the ONNX runtime initializes.

📦 Public Windows release (installers)

Public Windows release (installers)

Public installers target 64-bit x64 Windows only. Maintainers build MSI packages with the WiX Toolset via NuGet (WixToolset.Sdk). Only the .NET 8 SDK is required—no separate Inno Setup or WiX install.

MSI	When to use
Online	Installs the app under Program Files with Start Menu and ARP branding, then downloads models during install through DownloadModel.cmd / RunDownloadModelElevated.cmd and an elevated WiX QuietExec step. curl progress is written to the MSI log (for example msiexec /i PrimeDictate-….msi /l*v install.log).

Build the online installer:

.\scripts\Build-Installers.ps1

Outputs: artifacts\installer\. Version comes from Directory.Build.props.

See installer/README.md for details. Redistribute ONNX model files only in compliance with their license/terms.

Maintainer release commands

Use this from a clean main worktree when publishing a new release. Replace 4.3.1 with the next version.

$Version = "4.3.1"

git status --short --branch
# Update Directory.Build.props to $Version before committing.
dotnet build -c Release

git add -A
git commit -m "release: v$Version"

git ls-remote --tags origin "refs/tags/v$Version"
git tag -a "v$Version" -m "v$Version"

git push origin main
git push origin "v$Version"

git status --short --branch

If git ls-remote prints an existing tag, do not recreate it; choose the correct next version or inspect the existing release first.

GitHub Releases and Webflow links

Tagged pushes that match vX.Y.Z keep the workflow artifact upload for CI debugging and also publish installer assets to the matching GitHub Release. If the Release does not exist yet, the workflow creates it first. Release downloads come from GitHub Releases, not the temporary workflow artifact ZIP. If Azure Key Vault signing secrets are unavailable, the release flow still publishes assets as unsigned builds instead of failing before release upload.

• Release page: https://github.com/CakeRepository/PrimeDictate/releases/tag/vX.Y.Z
• Direct x64 MSI asset: https://github.com/CakeRepository/PrimeDictate/releases/download/vX.Y.Z/PrimeDictate-Setup-vX.Y.Z-x64.msi
• Direct ARM64 MSI asset: https://github.com/CakeRepository/PrimeDictate/releases/download/vX.Y.Z/PrimeDictate-Setup-vX.Y.Z-arm64.msi
• Latest release page: https://github.com/CakeRepository/PrimeDictate/releases/latest
• Chocolatey package asset: https://github.com/CakeRepository/PrimeDictate/releases/download/vX.Y.Z/primedictate.X.Y.Z.nupkg

Because the MSI filenames include the tag and architecture, Webflow should either link to the release page/latest page or update both direct MSI URLs each time a new release tag is published.

Chocolatey release alignment

Chocolatey packaging is part of the same vX.Y.Z tag release flow so MSI, docs, and Chocolatey stay aligned. The package downloads the versioned GitHub Release MSI for the machine architecture and verifies it with a SHA256 checksum embedded during package build.

• Chocolatey package id: primedictate
• Chocolatey package source/repo: https://github.com/CakeRepository/PrimeDictate
• Product overview page: https://www.flowdevs.io/portfolio/project/primedictate-local-ai-dictation-app
• Official release downloads: https://github.com/CakeRepository/PrimeDictate/releases

The tag-triggered workflow pushes to Chocolatey only when CHOCO_API_KEY is present. Without that secret, the workflow still publishes release assets to GitHub and logs that Chocolatey push was skipped.

For moderation retries, maintainers can repack and push locally with the same version after fixing package metadata/files. See installer/README.md for the local repack checklist and required moderation files.

Silent install and update commands (Windows)

• MSI install x64 (silent): msiexec /i PrimeDictate-Setup-vX.Y.Z-x64.msi /qn /norestart
• MSI install ARM64 (silent): msiexec /i PrimeDictate-Setup-vX.Y.Z-arm64.msi /qn /norestart
• MSI install without launch at login: msiexec /i PrimeDictate-Setup-vX.Y.Z-<arch>.msi LAUNCHATLOGIN=0 /qn /norestart
• MSI upgrade (silent): msiexec /i PrimeDictate-Setup-vX.Y.Z-<arch>.msi /qn /norestart
• MSI uninstall (silent): msiexec /x PrimeDictate-Setup-vX.Y.Z-<arch>.msi /qn /norestart
• Chocolatey install (silent by default): choco install primedictate -y
• Chocolatey install without launch at login: choco install primedictate -y --params "'/NoLaunchAtLogin'"
• Chocolatey upgrade (silent): choco upgrade primedictate -y
• Chocolatey uninstall (silent): choco uninstall primedictate -y

Tray shell and first-run setup

PrimeDictate now runs as a WPF tray app (no console window in normal use):

• Tray shell: Notification-area icon with Open Workspace, Settings, and Exit menu items.
• Tray status colors: Ready = Blue, Recording = Red, Processing = Green, Error = Yellow. Tooltip text follows app state (Ready, Listening, Processing transcript, Error).
• First launch: If %LocalAppData%\PrimeDictate\settings.json is missing or incomplete, a guided setup window appears with Welcome, Model, Commands, Replacements, and Impact tabs.
• Configurable commands: Global toggle, emergency stop, and history shortcuts are loaded from saved settings and applied to GlobalHotkeyListener at startup. Voice phrases for commit, emergency stop, history, and command prompt actions, including chained type ... text and Stop/Continue behavior, are configurable from the Commands tab.
• Backend picker + download: Setup and Settings include curated Whisper, Parakeet, Moonshine, Whisper.net, and Qualcomm AI Hub model options, local download progress, and a manual browse fallback.
• Experimental ARM64 QNN path: Native win-arm64 builds can expose an experimental Qualcomm AI Hub Whisper backend that uses ONNX Runtime QNN EPContext wrappers with explicit QNN diagnostics.
• Runtime model switching: Changing the selected backend or model causes the next transcription session to reload the correct engine automatically.
• Preview settings: Setup window includes the overlay style, silence auto-commit delay, optional coding-mode Enter key, PrimeDictate audio cues, and mic capture behavior.
• Impact dashboard: Settings includes a local stats tab with productivity cards, a 14-day words chart, and milestone achievements.
• Built-in update checks: After first-run setup, PrimeDictate checks GitHub Releases at most once per day when automatic checks are enabled. Failed install attempts clear the check timestamp so the next launch can retry. The tray menu also has Check for updates for a manual check.
• Installer continuity: The online MSI keeps one product identity for clean upgrades.
• Installer finish launch: The online MSI exposes “Launch PrimeDictate when setup completes” (checked by default), which starts the app after install.
• Launch at login: MSI installs add an all-users Windows Startup shortcut by default. Use LAUNCHATLOGIN=0 for silent MSI installs or Chocolatey /NoLaunchAtLogin when you do not want PrimeDictate to start when users sign in. The Settings window can move startup to the current user’s Startup folder or disable it later.

Publish folder only (no installer):

.\scripts\Publish-Windows.ps1

💻 Build and run

Build and run

cd path\to\PrimeDictate
dotnet run

The app starts in the tray. On first launch, complete setup, then focus another application and use your configured hotkey to start dictation. A live transcript appears in the overlay while you speak. PrimeDictate commits after the configured silence delay or when you press the start/stop toggle again. Set silence delay to 0 if you want the toggle hotkey to be the only commit trigger. The emergency stop shortcut and stop phrase discard the active capture without typing text.

Note: Stopping a running dotnet run (or any running PrimeDictate.exe) may be required before dotnet build can replace bin\...\PrimeDictate.exe on Windows (file lock on the apphost).

Using dictation reliably

• Keep the caret in the field where you want text before starting. The app does not move focus for you.
• Do not click into another application while the tray says Processing transcript; if the foreground window changes, injection is skipped for safety.
• Use the overlay as the live feedback surface. The focused editor receives only the final committed transcript.
• For a built-in mouse cue, enable Windows' "show location of pointer when I press CTRL key" setting. PrimeDictate uses that OS feature when available.
• Long monologues are heavier than short phrases because Whisper preview reprocesses snapshots and the final pass processes the full recording. A faster or smaller model helps if this becomes limiting.

⚙️ Configuration surface & Architecture

Configuration surface

Mechanism	Purpose
--enable-launch-at-login / --disable-launch-at-login	Command line switches that add or remove PrimeDictate.lnk from the current-user or all-users Windows Startup folder. All-users changes request elevation when needed.
User settings + first-run	Stored at %LocalAppData%\PrimeDictate\settings.json with FirstRunCompleted, launch-at-login scope, dictation/stop/history hotkeys, optional voice command phrases, selected backend, selected model id, resolved model path, optional exclusive mic capture toggle, overlay style, silence auto-commit delay, return-to-original-target toggle, audio cue toggle, automatic update checks, overlay placement, coding-mode Enter toggle, and baseline typing speed for impact estimates.
PRIMEDICTATE_QNN_* env vars	Maintainer and validation controls for the experimental Qualcomm QNN backend, including strict no-CPU-fallback mode, context caching, and optional QNN profiling output.

| Transcript history | Stored at %LocalAppData%\PrimeDictate\history.json with timestamp, transcript text, thread id, delivery status, target display name, optional error, and audio duration metadata. | | Impact stats | Stored at %LocalAppData%\PrimeDictate\stats.json with local aggregate word counts, audio duration, daily buckets, and unlocked achievements. |

Architecture (high level)

Area	Technology
Hotkey	SharpHook SimpleGlobalHook, keyboard only; toggle, emergency stop, and history gestures are loaded from settings and matched on KeyPressed.
Capture	NAudio WasapiCapture + MediaFoundationResampler to 16 kHz mono PCM.
Live preview	DictationController.LivePreviewLoopAsync snapshots the growing PCM buffer, re-runs the selected local backend for the overlay, and watches recent RMS level for silence.
Transcription	TranscriptionEngineHost selects and owns the active engine. Whisper, Parakeet, and Moonshine use ONNX models through sherpa-onnx OfflineRecognizer. Whisper.net handles GGML GPU/OpenVINO paths. The experimental Qualcomm backend drives Qualcomm AI Hub Whisper EPContext wrappers directly with ONNX Runtime and can attempt strict QNN HTP execution on native Windows ARM64.
Overlay	TranscriptionOverlayWindow is topmost, non-activating, and click-through so the target editor keeps focus.
Typing	WhisperTextInjectionPipeline.TranscribeAsync builds the final transcript, then InjectTextToTarget sends it once; optional coding mode follows with VcEnter.
Target safety + pointer cue	WindowsInputHelpers.cs captures the foreground window and focused control at recording start, can optionally restore that target for final injection, and uses Windows Mouse Sonar if enabled.

Experimental Qualcomm QNN maintainer workflow

Build the app and native ARM64 publish output as usual:

dotnet build PrimeDictate.sln
.\scripts\Publish-Windows.ps1 -RuntimeIdentifier win-arm64

The app catalog downloads the Qualcomm AI Hub Whisper Small Snapdragon X Elite precompiled_qnn_onnx package into:

$qaihubWhisperDir = Join-Path $env:LOCALAPPDATA 'PrimeDictate\models\qualcomm-aihub-whisper\whisper_small-precompiled_qnn_onnx-float-qualcomm_snapdragon_x_elite'

To validate that the precompiled Qualcomm AI Hub Whisper wrapper sessions can be created on QNN HTP:

dotnet run -- --qnn-whisper-smoke $qaihubWhisperDir Npu
dotnet run -- --qnn-whisper-proof $qaihubWhisperDir true

If you want the JSON result persisted to a file instead of relying on stdout, add an optional trailing output path:

$output = Join-Path $env:TEMP 'primedictate-qnn-smoke.json'
dotnet run -- --qnn-whisper-smoke $qaihubWhisperDir Npu $output

The older Moonshine QNN harness remains available for maintainer experiments with hand-prepared QDQ artifacts:

$moonshineDir = Join-Path $env:LOCALAPPDATA 'PrimeDictate\models\moonshine\sherpa-onnx-moonshine-base-en-int8'
dotnet run -- --qnn-smoke $moonshineDir Cpu
dotnet run -- --qnn-proof $moonshineDir true

The Whisper-specific harness can also validate standard Whisper ONNX encoder/decoder session creation. For standard sherpa Whisper folders this is session-creation-only; for Qualcomm AI Hub Whisper folders it validates the precompiled EPContext wrappers:

$whisperDir = Join-Path $env:LOCALAPPDATA 'PrimeDictate\models\whisper\sherpa-onnx-whisper-small.en'
dotnet run -- --qnn-whisper-smoke $whisperDir Cpu
dotnet run -- --qnn-whisper-proof $whisperDir true

You can add an optional trailing output path to persist the JSON result to a file instead of relying on stdout.

If strict validation passes, the result reports that QNN HTP session creation and inference succeeded with CPU fallback disabled. If it fails, PrimeDictate should not be treated as NPU-validated on that model/device combination.

Why not clipboard + Ctrl+V?

An earlier design put the transcript on the clipboard, simulated Paste, then restored the previous clipboard. Many applications handle paste asynchronously, so the restore often ran before the app read the new clipboard, and users saw the old clipboard (for example, a recently copied URL). The current design avoids that class of race by not using the clipboard for injection in the first place.

---

License

This repository’s application code is provided as in-repo source; follow the licenses of the dependencies (sherpa-onnx, NAudio, SharpHook, and the ONNX model terms from their respective publishers) when redistributing.

About FlowDevs & The Author

PrimeDictate is built by Justin Trantham, Co-Founder and Prime Automator at FlowDevs.

Founded in 2023, FlowDevs bridges the gap between business goals and scalable tech. We stop you from running your company on spreadsheets by building custom internal tools, client portals, and Power Platform solutions that actually scale. We specialize in workflow automation, integrations, and internal apps for Minnesota businesses.

• Visit FlowDevs.io
• Meet the Team
• Book a Discovery Call