WpfOCR

Windows desktop OCR tool: screenshot, annotate, recognize text (PP-OCR / RapidOCR packs), long screenshot, screen recording, PDF workbench, ASR/TTS, optional translation, and optional local HTTP API.

Current version: 1.0.2

Languages: English · 中文

Screenshots

Features

Area	Description
Screenshot recognition	Region capture → text OCR or barcode/QR recognition according to the current result tab; keeps the selected tab; multi-monitor DXGI capture
Screenshot annotate	WeChat-style tools: rect / ellipse / arrow / pen / text, color dots, undo / save / confirm
Long screenshot	Pick a window → auto-scroll stitch → open in viewer (no OCR)
Screen recording	Window or region → HUD (move/resize region, draggable bar) → MP4 (x264/x265/AV1 via FFmpeg only) + optional system/mic audio
GIF recording	Same region flow → capture 24 fps → preview (output FPS, scale, palette) → silent GIF
Clipboard	Paste image and run OCR; copy image / text; menu/tray can switch on-capture copy mode (image / file / path) and re-copy the last screenshot
Overlay text	Text layer on the image; drag-select and copy
PDF workbench	Open PDF → page OCR → edit lines → export searchable PDF (invisible text layer)
ASR / TTS	Offline speech recognition (sherpa-onnx) and TTS (Sherpa + SAPI / WinRT system voices); install voices in-app
SAPI x86 helper	Sidecar x86host.exe (32-bit SAPI web only) for classic voices visible only in x86 processes
Devices	CPU · NVIDIA CUDA (GPU) · Intel / DirectML (iGPU); missing accel → CPU
Install features	In-app download of models and runtimes (CN mirrors when locale is Chinese)
Hotkeys	Toggle main window · snap annotate · snap OCR · voice input (configurable)
HTTP API	Local JSON API (default 127.0.0.1:1224)
CLI	Batch OCR, list models / SAPI voices, probe CUDA, multi-monitor snap test

Requirements

• Windows 10/11 (x64)
• .NET Framework 4.8 (runtime on end-user PCs)
• Build: Visual Studio / MSBuild with .NET Framework 4.8 targeting pack (or SDK that can build net48 WPF)
• Optional: NVIDIA GPU + CUDA stack matching the ORT GPU package (onnxgpu64)
• Optional: DirectML-capable GPU for the “核显” device option (onnxdml64)
• Optional (screen record): FFmpeg 4.4 shared libraries under ffmpeg64/ next to the exe

Project layout

OCR/
├── WpfOCR/                 # Application source (WPF, net48, x64)
│   ├── Assets/
│   └── bin/Release/
│       ├── net48/          # Dev output (models / runtimes live here)
│       └── WpfOCR/         # Slim package: WpfOCR.exe + x86host.exe + managed deps
├── x86host/                # Standalone 32-bit SAPI web helper (x86host.exe only)
├── docs/                   # README screenshots
├── scripts/publish-release.mjs
├── README.md
├── README.zh.md
├── CHANGELOG.md
└── AGENTS.md

Model packs and large native runtimes are not stored in source. Place or install them next to the executable:

WpfOCR/bin/Release/net48/
├── WpfOCR.exe
├── config.toml              # created/updated at runtime
├── ocrmodels/               # OCR packs (rapid-ch, rapid-i18n, …)
├── asrmodels/               # ASR packs (optional)
├── ttsmodels/               # TTS voices (optional)
├── translatemodels/         # Translation ONNX (optional)
├── onnxcpu64/               # ONNX Runtime for CPU EP (on-demand install)
├── onnxgpu64/               # CUDA ORT + CUDA libs (optional)
├── onnxdml64/               # DirectML ORT (optional)
└── ffmpeg64/                # FFmpeg shared DLLs for record (optional)

Each OCR pack needs ONNX models + configs.txt (and dict/keys as required by the pack).

Build & run

cd WpfOCR
dotnet build -c Release

Run:

./WpfOCR/bin/Release/net48/WpfOCR.exe

A plain build does not ship models, onnxcpu64, full CUDA, or FFmpeg. Use Tools → Install features (or first-run wizard) inside the app.

Slim release package (bin\Release\WpfOCR\)

Release builds also produce a small redistributable under WpfOCR\bin\Release\WpfOCR\:

• Includes: WpfOCR.exe, x86host.exe (32-bit SAPI web), managed dependencies, wetext/ (ITN), Assets, LICENSE.
• Does not include: OCR/ASR/TTS models, onnxcpu64 / onnxgpu64 / onnxdml64, OpenCV/Skia/PDFium/Sherpa natives, ffmpeg64.
• End users install those via Install features (downloads from mirrors / NuGet CDN).
• Translation is not covered by the installer: place Opus-MT ONNX under translatemodels/ yourself if needed.

For local development with models and GPU already present, run bin\Release\net48\ instead.

Release archive (release/wpfocr_x.x.x.7z)

node scripts/publish-release.mjs

Runs Release build, then packs WpfOCR/bin/Release/WpfOCR/ into release/wpfocr_<version>.7z (requires 7-Zip on PATH). The release/ folder is gitignored.

In-app install (recommended)

1. First launch may open the install wizard (defaults: OpenCV, ORT CPU, OCR rapid-ch, first two ASR packs, FFmpeg; GPU/iGPU off).
2. Later: Tools → Install features
   • Components: OpenCV, Skia, PDFium, Sherpa, ORT CPU (onnxcpu64), OCR/ASR packs, CUDA, DirectML, FFmpeg.
   • Voices: TTS models with language filter; progress shows total batch size and downloaded bytes.
3. Using a feature that needs a missing package prompts to open the installer (e.g. OCR without any ORT → install onnxcpu64).

Runtime	Role	Typical size
onnxcpu64	CPU ONNX Runtime (required for OCR if no GPU/iGPU ORT)	~16 MB
onnxgpu64	NVIDIA CUDA EP + CUDA/cuDNN (optional)	large
onnxdml64	DirectML EP for iGPU (optional)	~18 MB
OpenCV	Capture / image pipeline	~61 MB
ffmpeg64	Screen record encode/mux	~72 MB

Download prefers CN mirrors when UI or system locale is Chinese.

Optional env vars for local full libraries (do not commit secrets/paths into docs meant for others):

Variable	Meaning
WPF_OCR_CUDA_LIB	Folder with full CUDA / onnxgpu64 DLLs
WPF_OCR_FFMPEG_LIB	Folder with FFmpeg 4.4 shared DLLs

Configuration

Settings are stored in config.toml beside the exe (also editable via Tools → Settings / Record options). The Settings window groups options into tabs: General, OCR, Hotkeys, Speech, Capture, API.

[ocr]
model_pack = "rapid-ch"
model_variant = "简体中文 mobile"
device = "Cpu"          # Cpu | Gpu | IntelGpu
det_limit = 960
det_thresh = 0.3
det_box_thresh = 0.5
use_cls = true

[ui]
hotkey = "Ctrl+Alt+O"           # show / hide main window
hotkey_snap = "Ctrl+Alt+Q"      # screenshot annotate
hotkey_snap_ocr = "Ctrl+Alt+W"  # screenshot + OCR
minimize_to_tray = true
capture_log = false             # true → log/capture.log
ui_lang = "zh"                  # zh | en

[http]
http_enabled = true
http_host = "127.0.0.1"
http_port = 1224
service_mode = false            # keep engine warm

[pdf]
pdf_invisible_text = true
pdf_dpi = 150                   # internal raster DPI; page size follows original PDF

[record]
record_codec = "x264"           # x264 | x265 | av1
record_fps = 24
record_crf = 28                 # x264/x265 only, 0–51, higher = smaller
record_av1_crf = 56             # AV1 only, 0–63 (different scale; 56 ≈ half of x265 CRF28 size)
record_audio = true
record_audio_src = "Speakers"   # Speakers | Mic | MicAndSpeakers
record_audio_kbps = 96
record_max_size = false
record_max_w = 1920
record_max_h = 1080
record_lock_aspect = true      # lock aspect when resizing HUD region after Start (free before Start)

[asr]
asr_voice_mode = "stream"       # stream = live; offline = record until hotkey stop, then one-shot ASR
asr_voice_polish = true         # LLM polish for voice input (needs asr_llm_url + asr_llm_model)
asr_voice_split = true
asr_voice_split_sec = 5         # split only after this many seconds of silence (1–30); do not cut continuous speech
asr_live_mode = "stream"        # stream | offline (offline splits on silence)
asr_live_polish = false         # LLM polish each live-caption sentence
asr_live_split = true           # auto-split after polish / completed sentences
# asr_llm_url / asr_llm_model required when a polish checkbox is on (OpenAI-compatible)
# asr_llm_token = ""            # do not commit secrets
# asr_llm_prompt = "..."
# Polish sends prior output in the same session as context (homophones / names).
# Voice HUD: line 1 stays “listening…”; line 2 is “识别中” / “润色中” plus transcript.
# After polish+inject, line 2 is cleared.
# Esc during recognize/polish stops the session with no inject.

[gif_record]
gif_fps = 8                     # default output FPS in preview (1–24); capture is 24 fps
gif_max_size = true
gif_max_w = 1280
gif_max_h = 720
gif_colors = 128                # palette colors in preview (32/64/128/256)
gif_scale = 100                 # default scale % in preview

Leave a hotkey string empty to disable that hotkey.

Do not commit real config.toml if it encodes machine-specific paths or preferences you want private.

Screen recording

1. Capture → Screen record (or the toolbar button): click a window or drag a region.
2. HUD (drawn outside the capture area):
   • Red frame; drag the 5px strip outside the red line to move, or 8 grips to resize. Before Start, aspect ratio is free; after Start, resizing locks aspect ratio unless disabled in record options (record_lock_aspect = false). Encoded resolution is fixed when you press Start.
   • Floating control bar: drag via the left grip; collapse to mini bar; Options before Start (record/GIF settings); start/pause share one slot.
   • Bar auto-positions above/below the region and stays within the current monitor (multi-monitor safe).
3. Stop → confirm save → MP4 is written; Explorer opens and selects the file.
4. Capture → Record options: codec (x264 / x265 / AV1), FPS, CRF (x264/x265) and AV1 CRF (AV1 only, separate scale 0–63, default 56), audio source, max output size. AV1 needs an encoder in ffmpeg64 (libsvtav1 / libaom-av1); a missing encoder fails with a message and does not fall back to x264.

GIF recording

1. Capture → GIF record: same window/region pick.
2. Same HUD; capture at 24 fps; after Stop, the preview window lets you set output FPS (1–24), scale, and palette colors, then save a silent GIF.
3. Capture → GIF record options: default output FPS, max width/height, default colors.

Notes

• MP4 / GIF recording requires FFmpeg shared under ffmpeg64/ (install in-app or place manually). OpenCV is not used for video encode.
• System-loopback audio pads silence on wall-clock gaps so late sound is not shifted to the start of the file.
• Temporary files live under the app tmp/ folder and are cleaned up after a successful save.
• GIF size grows quickly with resolution and duration — use preview scale/FPS and the max-size limit.

Default hotkeys

Hotkey	Action
Ctrl+Alt+O	Toggle main window show / hide
Ctrl+Alt+Q	Screenshot annotate
Ctrl+Alt+W	Screenshot and OCR
Ctrl+Alt+V	Voice input (press again to stop)
Ctrl+Alt+B	Live caption

Tray icon: left-click toggles the window; context menu includes voice input, clipboard OCR, on-capture copy mode (image / file / path), and exit. Switching copy mode re-copies the last screenshot to the clipboard in the new form. Closing the main window typically hides to tray rather than exiting. Capture → Voice input in the main menu is the same toggle.

UI language

• Tools → Language → 中文 / English (applies immediately)
• Or set UI language in Settings → General
• Persisted in config.toml as ui_lang = "zh" or "en"

CLI

WpfOCR --image <path> [options]
WpfOCR --snap [--out <dir>]
WpfOCR --list-models
WpfOCR --list-sapi              # local SAPI + (x64) x86host voices
WpfOCR --probe-cuda
WpfOCR --help

Useful options: -d gpu|cpu, -p rapid-ch, -v <variant>, -m <models-dir>, --det-limit, --no-cls.

x86host (32-bit SAPI only)

Some classic SAPI voices register only for 32-bit processes. Ship x86host.exe next to WpfOCR.exe (built from x86host/; a Release build of WpfOCR also builds and copies it).

Item	Detail
Role	HTTP helper: list SAPI voices + synth WAV; no GUI, no OCR/ASR
Start	On demand by the x64 app, or run x86host.exe manually
Bind	127.0.0.1 only, default port 17886
Idle	Exit after 60s without requests (--idle-ms to override)
API	GET /api/sapi/status · GET /api/sapi/voices · POST /api/sapi/synth · POST /api/sapi/shutdown

dotnet build x86host/x86host.csproj -c Release
x86host.exe --port 17886 --idle-ms 60000
x86host.exe --list-sapi

In the UI, choose engine SAPI: local voices plus x86-only entries (display name ends with · x86); speak/export for those voices goes through x86host.

HTTP API (overview)

When enabled, a local server listens on http_host:http_port (default loopback only).

• GET /api · /api/status — capabilities
• POST /api/ocr — image (JSON base64 or multipart)
• GET /api/ocr/get_options — OCR options snapshot
• GET /api/asr/models · POST /api/asr — speech recognition
• GET /api/tts/models · POST /api/tts — TTS (wav base64)
• POST /api/itn — inverse text normalization

Bind to 127.0.0.1 unless you intentionally expose the service on a trusted network.

Full field reference: HTTP-API.md · HTTP接口文档.md (中文).

Capture diagnostics

Set capture_log = true in config.toml to write log/capture.log (multi-monitor / DPI troubleshooting). Keep it off for normal use.

CLI: WpfOCR --snap dumps full-monitor bitmaps under log/snap/ (or --out).

License

WpfOCR application source code (this repository’s WpfOCR/ sources, scripts, and docs authored for the project) is released under the MIT License. See LICENSE.

Copyright (c) 2026 WpfOCR Contributors

You may use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Software, subject to including the copyright and permission notice. THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND.

Third-party components

Bundled or optional dependencies are not all MIT. Models, FFmpeg builds, CUDA/cuDNN, and some native libraries keep their own terms. See THIRD_PARTY_NOTICES.md.

Do not commit large ONNX weights, CUDA redistributables, or FFmpeg shared binaries into git without a redistribution plan that matches their licenses.

See also

• LICENSE — MIT (application source)
• THIRD_PARTY_NOTICES.md — dependency license notes
• CHANGELOG.md
• HTTP-API.md · HTTP接口文档.md — HTTP API
• README.zh.md — Chinese documentation