VoxFlow

A macOS asset workbench for voice, screenshots, screen recordings, clipboard history, and coding-agent instructions.

Press ⌥Space to open the launcher and recover recent voice, screenshot, screen recording, and clipboard assets. Dictation, captures, recordings, and copied content become searchable, copyable, reusable local history.

🌐 Website  ·  ⬇️ Download  ·  English  ·  中文  ·  繁體中文  ·  日本語  ·  한국어

At A Glance

VoxFlow is an asset workbench and fast launcher for the app you are already using. It is not a voice assistant: it does not take over the window, submit messages, or move you into another input box. It turns voice, screenshots, screen recordings, clipboard items, and agent commands into searchable, previewable, reusable local assets that return to your current workspace.

What You Want To Do	Trigger	Output	Boundary
Open the launcher	⌥Space	Raycast-style launcher	Recent Assets is selected by default; keyboard navigation first
Recover recent assets	Launcher -> Recent Assets	Second-level asset browser	Voice, screenshots, and clipboard share search and filters
Dictate text	Hold the shortcut, speak, release	Current cursor position	No focus stealing, no auto-submit
Manage clipboard assets	Copy text, images, files, links, or colors	Asset history	Noise filters still skip content that should not be saved
Fix misrecognized terms	Runs after ASR final output and optional LLM correction	Text before insertion	Local deterministic rules; learned candidates stay user-controlled
OCR a clipboard image	Copy an image, press ⌘⇧V	Current cursor position	Image-only workflow; does not start normal dictation
Capture and process a screenshot or screen recording	Press ⌘⇧A, select a region	OCR result panel	Translation, summary, and speech playback are optional
Run selection actions	Select text, press ⌘⇧F/J/K/L/P	Action HUD or result panel	F opens the action card; J translates, K summarizes, L sends to Task Assistant, P sends to Ask AI
Ask AI from the launcher	Type a question in the launcher, choose "Ask AI"	Ask AI chat HUD	Reuses your configured LLM provider; multi-turn, streaming, Markdown
Search the web from the launcher	Type a keyword, choose a Quicklink	Default browser	Built-in Google, Bing, Perplexity, GitHub, StackOverflow, YouTube, Bilibili, X, Xiaohongshu, Taobao, JD
Open a URL from the launcher	Type a URL or bare domain	Default browser	Auto-detects http/https/bare domain/localhost/IP+port; first result is "Open URL"
Review screenshot and recording records	Open Workbench → Screenshot	Local screenshot and recording history and OCR text	Stored locally; records can be searched, favorited, copied, and deleted
Compose an AI prompt	Combine current-window context with spoken intent	Copyable prompt	Copy only, no injection, no auto-submit
Command local coding agents	Speak a task assistant name and task	Codex / Claude / CodeBuddy / terminal agent session	Dispatches only to registered sessions

Who It Is For

• You often talk to ChatGPT, Claude, Codex, Cursor, or other AI tools and need to describe intent, context, or revision requests quickly.
• You run Codex, Claude, CodeBuddy, or other terminal agents and want to dispatch spoken instructions to the right local session.
• You write code and frequently explain bugs, add notes, draft commit messages, or document investigation steps.
• You extract text from screenshots, web pages, error dialogs, or images, then translate or summarize it.
• You speak mixed Chinese and English, where technical terms and product names are easy to misrecognize.

Reading Map

If You Want To...	Start Here
Install and try it	Quick Start
Understand the launcher and assets	Corrections, OCR, And Agent Workflows
Understand speech models	Dictation And Speech Models
Understand OCR, translation, summary, and agents	Corrections, OCR, And Agent Workflows
Check where data goes	Privacy
Understand the stack and open-source dependencies	Tech Stack And Open-Source Dependencies
Build from source	Run From Source

Dictation And Speech Models

Hold To Speak, Release To Insert

VoxFlow works like a keyboard layer. Hold your dictation shortcut, speak, and release. A small transcription overlay appears while you are speaking, then the final text is inserted into the current cursor position.

There is no need to switch apps or manually copy text back.

Live Transcription

While you speak, VoxFlow shows recognized text in real time so you can stay oriented. It works for short commands, long explanations, Chinese, English, and mixed Chinese-English speech.

VoxFlow includes the system speech recognizer plus local and cloud ASR providers. Apple Speech works out of the box; Qwen3-ASR, Whisper, FunASR, SenseVoice, NVIDIA Nemotron, Parakeet, and Omnilingual cover local workflows, while Groq, Tencent Cloud, and Alibaba Cloud provide online recognition. The Models page labels local versus online, streaming capability, and language coverage explicitly.

Supported Speech Models

VoxFlow does not force every local model into the same runtime. Each provider follows the route that best matches its upstream model format and latency target:

Provider / Model	Current Runtime Route	Recommended Use
Apple Speech	Apple Speech / SFSpeechRecognizer	Out-of-the-box dictation without downloading a model
Qwen3-ASR 0.6B	speech-swift Qwen3ASR MLX 4bit	Default local route using the unified speech-swift runtime
Qwen3-ASR 1.7B	speech-swift Qwen3ASR MLX 8bit	Higher-accuracy local route sharing the same speech-swift loading and session path as 0.6B
Whisper Turbo / Large V3	WhisperKit .mlmodelc	High-quality full-recording transcription after capture ends
FunASR	Sherpa-ONNX	Local Chinese fallback path; not CoreML
SenseVoice	FluidAudio / CoreML	Local multilingual and short-utterance transcription
Paraformer	FluidAudio / CoreML int8	Local Chinese transcription
NVIDIA Nemotron 0.6B	speech-swift NemotronStreamingASR / CoreML	Local multilingual streaming transcription
Parakeet Streaming	speech-swift ParakeetStreamingASR / CoreML	Low-latency local streaming dictation for English and European languages
Omnilingual ASR	speech-swift OmnilingualASR / CoreML	Broad-language offline transcription and experimental workflows

Cloud providers send recorded audio to the selected service. Groq returns a final transcript after recording; Tencent Cloud and Alibaba Cloud support real-time WebSocket transcription.

Cloud Provider	Status	Streaming	Default Model / API	Configuration
Groq (Free)	Supported	No	whisper-large-v3-turbo audio transcription	API Key, model
Tencent Cloud	Supported	Yes	Realtime Speech Recognition WebSocket, 16k_zh	AppID, SecretId, SecretKey
Alibaba Cloud	Supported	Yes	DashScope WebSocket, fun-asr-realtime	Bailian API Key
Volcengine Cloud	Planned	Planned	Doubao streaming ASR	To be determined
Mistral Voxtral, AssemblyAI, ElevenLabs Scribe	Not yet supported	To be determined	Reserved providers	None

Corrections, OCR, And Agent Workflows

Personal Corrections And Optional LLM Correction

Speech recognition can struggle with technical terms such as Python, JSON, TypeScript, framework names, or product names. VoxFlow can run a conservative correction pass through your own OpenAI-compatible provider after dictation finishes.

The new Personal Corrections page runs deterministic local fixes after ASR final output and optional LLM correction. It can also learn candidate rules from edits you make after insertion. The LLM pass remains intentionally restrained: it fixes obvious recognition mistakes instead of rewriting your tone or polishing your content.

Clipboard OCR, Screenshot Capture, Translation, And Summary

Copy a screenshot and press ⌘⇧V to OCR the clipboard image and paste the recognized text into the current cursor position. Press ⌘⇧A to select a screen region and open a result panel with Original Image, OCR, Translation, and Summary tabs.

This is useful for web pages, error dialogs, screenshots, design mockups, and chat history. OCR text can be copied, spoken, translated, or summarized, but it does not feed the permanent Personal Corrections learning loop.

Agent Compose And AI Coding 助手 Command Center

Agent Compose combines visible window context, OCR text, and your spoken intent into a prompt you can paste into an AI tool. It only copies the result; it does not inject, submit, or press Enter for you.

AI Coding 助手 Command Center is for local coding-agent terminals. After you enable it, speak a task assistant name and instruction, and VoxFlow resolves the target agent, shows confirmation state, and dispatches the instruction to the matching Codex, Claude, CodeBuddy, or other registered terminal session.

Workbench

VoxFlow also includes a full asset workbench:

Page	What You Can Do
Home	Review asset history, today's additions, source breakdown, and reusable content; search, copy, or delete voice, screenshot, and clipboard assets
Personal Corrections	Manage deterministic correction rules, learned candidates, enablement, and recent events
Styles	Choose output styles such as original, formal, email, or coding notes
File Transcription	Import audio or video files, transcribe them, export txt/md/srt, or save as notes
Notes	Record voice notes, edit Markdown, search, and review recent notes
Screenshot	Browse captured screenshots and screen recordings with OCR text, favorites, search, and paging
AI Coding 助手	Review registered agents, aliases, working directories, branches, and dispatch logs
Settings	Manage input devices, shortcuts, models, translation models, permissions, privacy, and data
Help	Find permission guidance, version information, and project links

Highlights

• VoxFlow Palette launcher: Press ⌥Space for a Raycast-style launcher with Recent Assets selected by default, arrow-key navigation, Enter, and ⌘K actions.
• Asset history workbench: Successful ASR text, screenshots, and clipboard text/images/files/links/colors share one asset system; Home shows asset counts, source breakdown, and reusable content.
• Global dictation: Works in any editable text field, not only inside VoxFlow.
• Non-intrusive overlay: Shows live text and voice activity without taking focus.
• Multiple ASR providers: Start with the built-in system recognizer; local Qwen3-ASR, Whisper, FunASR, SenseVoice, NVIDIA Nemotron, Parakeet, and Omnilingual providers are being unified under the same runtime model; providers without real-time streaming are marked as Non-streaming in Models.
• Stable text insertion: Temporarily switches input source before paste, then restores both input source and clipboard to reduce CJK input-method interference.
• Input device selection: Choose your microphone; long device names are handled gracefully.
• Shortcut recording: Record the key you want to use and configure short-press behavior.
• Clipboard image OCR: Copy a screenshot or image, press ⌘⇧V, and VoxFlow recognizes the image text and pastes it into the current field.
• Screenshot OCR: Press ⌘⇧A, select a screen region, then review the original image, OCR text, translation, and summary in a result panel.
• Screenshot and recording library: Captured screenshots and screen recordings are kept in the Screenshot page with OCR text, favorites, search, and one-click copy/delete actions.
• Inline screenshot annotation: Region capture supports pen/shape/text/mosaic/scroll tools, undo/redo, and quick translate/summary flow before final insert/output.
• AI Coding 助手 Command Center: Dispatch spoken instructions to Codex, Claude, CodeBuddy, or other registered local terminal agents.
• Agent Compose: Turn current-window OCR context plus spoken intent into a prompt; it only copies the result and never auto-submits.
• OpenAI-compatible providers: Add, test, edit, and delete providers; LLM API keys are stored in macOS Keychain.
• Personal corrections and context hotwords: Fix repeated misrecognitions with local rules, and use current-window OCR to extract temporary context terms.
• History and notes: Search, copy, edit, and reuse previous input, screenshots, and copied content.
• File transcription: Turn recordings, videos, or meeting audio into text.
• Local-first data: History, personal corrections, settings, notes, and jobs live locally; LLM correction is opt-in.

Quick Start

Download & Install

Download the latest version from GitHub Releases:

1. Open VoxFlow-1.9.0-macOS.dmg
2. Drag VoxFlow into the Applications folder
3. On first launch, if macOS cannot verify the app, Control-click the app and choose Open

After installation, open Workbench -> Screenshot to verify your screenshot and recording records and OCR history at first use.

To try the latest main-branch implementations of Personal Corrections, AI Coding 助手, or Screenshot OCR, run from source; these capabilities may be newer than the latest stable Release.

Requirements

• macOS 15 Sequoia or later
• A Mac with a microphone

First Permissions

VoxFlow needs a few macOS permissions:

Permission	Why It Is Needed	Where
Accessibility	Listen for the global shortcut and insert text into the current app	System Settings -> Privacy & Security -> Accessibility
Microphone	Record your voice	System Settings -> Privacy & Security -> Microphone
Speech Recognition	Use the system speech recognizer	System Settings -> Privacy & Security -> Speech Recognition
Screen Recording	OCR the current window for Agent Compose, screenshot OCR, and screen recording	System Settings -> Privacy & Security -> Screen Recording

If you use a local Qwen3-ASR model, Speech Recognition permission is not required. Microphone permission is still required.

If the shortcut does not respond after granting permissions, quit and reopen VoxFlow.

Default Shortcuts

Shortcut	Action
⌥Space	Open the VoxFlow Palette launcher
Dictation shortcut	Hold to speak, release to insert at the current cursor; configurable in Settings
⌘⇧V	OCR the clipboard image and paste recognized text
⌘⇧A	Capture a screen region and open the OCR result panel
⌘⇧F	Open the selection action HUD for the selected text (Translate / Summarize / Task Assistant / Ask AI)
⌘⇧J	Translate the selected text directly
⌘⇧K	Summarize the selected text directly
⌘⇧L	Send the selected text directly to Task Assistant
⌘⇧P	Send the selected text directly to the Ask AI chat HUD

Selection-action shortcuts can be changed or cleared individually in Settings → Selection Actions → Activation.

How To Use

Dictation

1. Place your cursor in any text field.
2. Hold the dictation shortcut.
3. Speak. The overlay shows live recognition.
4. Release the shortcut. The final text is inserted at the cursor.

Voice Notes

Open the workbench and go to Notes. Click the record button to start a quick note. VoxFlow transcribes as you speak, then lets you edit and review the note afterward.

File Transcription

Open File Transcription, select an audio or video file, and let VoxFlow process it. Completed jobs can be copied, exported, or saved as notes.

Clipboard Image OCR

Copy a screenshot or image, then press ⌘⇧V. VoxFlow reads the image from your clipboard, runs OCR, and pastes the recognized text into the current cursor position.

If the clipboard does not contain an image, this shortcut does not start normal dictation; it is reserved for the clipboard image OCR workflow.

Screenshot OCR, Translation, And Summary

Press ⌘⇧A, then select a region of the screen. VoxFlow captures that region, runs OCR, and opens a result panel with Original Image, OCR, Translation, and Summary tabs. You can copy or speak the available text from the panel.

Translation can use Apple system translation, a configured LLM, or a local translation model. Summary can use a configured LLM or a local summarizer. If no translation or summary model is available, the OCR text still remains usable.

Screenshot Record Library

Every screenshot captured with ⌘⇧A is saved as a local screenshot record so you can review it later in Workbench → Screenshot. You can search, filter by favorites, switch page size, copy recognized text, and delete entries. Image previews are loaded from local files and are not synced or uploaded.

Agent Compose

Agent Compose reads visible text and optional OCR context from the current window, combines it with your spoken intent, and produces a prompt for AI tools such as ChatGPT, Claude, Codex, or Cursor. It preserves the safety boundary: copy only, no injection, no auto-submit.

AI Coding 助手 Command Center

Enable AI Coding 助手 Command Center in Settings, then use the existing voice shortcut to enter the command HUD. Say an agent name and task, such as “frontend, check the button state,” and VoxFlow resolves the target, asks for confirmation when needed, and dispatches the instruction to that terminal agent session.

Launcher: Ask AI, Quicklinks, and Open URL

Press ⌥Space to open the launcher. In addition to searching apps, commands, and assets, you can also:

• Ask AI: Type any question, select "Ask AI", and press Enter. The launcher closes and the right-side HUD enters Ask AI chat mode. It reuses your configured LLM provider and supports multi-turn conversation, streaming replies, and Markdown rendering. The session stays in memory, so reopening Ask AI lets you continue asking follow-ups. When no provider is configured, the HUD shows a configuration hint instead of sending a request.
• Quicklinks: Built-in sites include Google, Bing, Perplexity, GitHub, StackOverflow, YouTube, Bilibili, X, Xiaohongshu, Taobao, and JD. Typing a site name, Chinese name, or alias (such as gh, tb, or b站) prioritizes that site; pressing Enter opens the search results in your default browser.
• Open URL: When you type a full URL, a bare domain (such as github.com/openai/codex), localhost:3000, or 127.0.0.1:8080, the first result is automatically selected as "Open URL" and Enter opens it in the default browser. Bare domains are normalized to https://.

The selection action panel (⌘⇧F) and the direct selection Ask AI shortcut (⌘⇧P) both send the selected text into the same Ask AI chat HUD, so you don't need to open the launcher first.

Improve Names And Terms

Use Personal Corrections for deterministic fixes, or enable current-window OCR context boost so project names, people names, product names, and technical terms can become temporary hotwords for the current task.

Enable LLM Correction

Open Settings -> Models, add an OpenAI-compatible provider, fill in Base URL, Model, and API Key, then test the connection. Once it works, enable LLM Correction in the same settings page.

LLM API keys are stored in macOS Keychain. Cloud ASR credentials for Groq, Tencent Cloud, and Alibaba Cloud are stored in the local SQLite settings database and can be revealed, hidden, or removed from Models.

Privacy

VoxFlow is local-first by default.

• Asset history, personal correction rules, notes, transcription jobs, and non-secret settings are stored locally.
• LLM API keys are stored in macOS Keychain; cloud ASR credentials are stored in the local SQLite settings database.
• Apple Speech may process audio according to macOS system behavior.
• Local Qwen3-ASR runs on-device after the model is downloaded.
• LLM correction is disabled by default. When enabled, only recognized text is sent to your configured API provider.
• When you select a cloud ASR provider, recorded audio is sent to that provider. Local models keep audio on-device. VoxFlow does not proactively upload notes, asset history, or clipboard content.
• Clipboard assets are saved locally for launcher and Home review; noise filters skip meaningless high-frequency changes.
• Clipboard image OCR can still be used as a one-off OCR entry.
• Screenshot and recording records (OCR text + screenshot files captured via ⌘⇧A) are stored locally and are never uploaded.

See Privacy for more details.

FAQ

Question	Answer
The shortcut does nothing	Check Accessibility permission, then quit and reopen VoxFlow
The overlay appears but no text shows up	Check Microphone, Speech Recognition, or the selected model state
Screenshot and recording records are missing	Go to Settings → Data & Privacy → Data Management, then check storage health and open the data folder to confirm Application Support/VoxFlow/Screenshots/ has image records. Also verify Screen Recording permission.
How do I disable a default screenshot annotation tool?	The current version does not expose a persistent "default annotation tool" setting; switch to the Select/Cursor tool in each capture panel to avoid entering annotation mode by default.
LLM correction does not run	Make sure it is enabled in Settings and the default provider passes the connection test
Why is my API key hidden?	That is expected. Use the reveal button while editing if you need to inspect it
Can I use it offline?	Download and select a local Qwen3-ASR model
Can deleted history or notes be restored?	Deletion is local and immediate, so please confirm before deleting

Run From Source

If you want to build the app yourself:

git clone https://github.com/xingbofeng/VoxFlow.git
cd VoxFlow
make run-dev

Common commands:

make run-dev      # Daily development: Debug + native arch, package and launch .app
make run-native   # Native Release for local checks close to shipped behavior
make build        # arm64 Release, used for release/DMG
make install      # Install into /Applications
swift test        # Run tests

Tech Stack And Open-Source Dependencies

VoxFlow is a native macOS app, not an Electron wrapper. The codebase is split into SwiftPM targets, keeps local-first paths local by default, and only uses cloud providers that the user explicitly configures.

Area	Stack / Open-Source Dependency	Used For
App shell	Swift 6, SwiftUI, AppKit, SwiftPM	Menu-bar app, Workbench, Settings, HUD, and macOS window lifecycle
System APIs	AVFoundation, Speech, Vision, Accessibility, Pasteboard	Recording, Apple Speech, screenshot/clipboard OCR, text insertion, and current-window context
Screenshot capture & annotation	VoxFlowScreenshotKit, ScreenCaptureKit, CoreGraphics, Vision	Region capture, annotation tools, scroll capture, and screenshot rendering
Local ASR	speech-swift Qwen3ASR / Nemotron, WhisperKit, FluidAudio, Sherpa-ONNX vendor runtime	Qwen3-ASR, NVIDIA Nemotron, Whisper, SenseVoice, Paraformer, and FunASR routes
Cloud ASR / LLM	OpenAI-compatible HTTP, Groq, Tencent Cloud realtime ASR, Alibaba DashScope	Online transcription, LLM correction, translation fallback, summary, and Agent Compose
Personal Corrections	Packages/VoxFlowVoiceCorrectionKit, inspired by TypeWhisper deterministic post-processing and focused text observation	Local rule matching, conflict resolution, learned candidates, and benchmark fixtures
Context hotwords	Packages/VoxFlowContextBoostKit, Vision OCR, NaturalLanguage	Extract temporary Top-K hotwords from current-window OCR text for the current prompt only
AI Coding 助手	Rust agent-cli/ helper/router, JSON IPC, MCP self-reporting	Dispatch spoken instructions to local Codex, Claude, CodeBuddy, or terminal agents
Verification	XCTest, Makefile, GitHub Actions, JiWER cross-check scripts	Unit tests, release builds, ASR/correction benchmarks, and metric validation

Attribution and licensing notes live next to the relevant modules: Packages/VoxFlowVoiceCorrectionKit/NOTICE.md, SOURCE_ATTRIBUTION.md, and MODIFICATIONS.md document TypeWhisper references and adaptation boundaries; Vendor/ contains packaged local runtime/vendor assets; AI Coding 助手 keeps only the Rust helper and no longer ships the old Python CLI.

Source Layout

Sources/                         # Swift app code, domain modules, ASR providers, text insertion, and other SwiftPM targets
Packages/VoxFlowVoiceCorrectionKit/ # Personal Corrections engine, benchmark fixtures, and package tests
agent-cli/                       # Rust helper/router source for AI Coding 助手; builds the bundled `voxflow` binary and `vox` shim
Tests/                           # Swift unit tests plus Python tests for ASR benchmark tooling
Resources/                       # App icon and bundled resources
Vendor/                          # Local runtime/vendor assets required by packaged builds
docs/                            # GitHub Pages site, privacy docs, design notes, and implementation plans
scripts/                         # Build, ASR benchmark, and architecture-check helper scripts
tools/                           # Auxiliary verification tools; currently JiWER cross-check only, not an agent CLI
.github/                         # CI, Pages, Release workflows, and release notes

AI Coding 助手 has a single maintained CLI implementation: the Rust source in root-level agent-cli/. The old Python vf-agent / agent-cli reference helper has been removed. Remaining Python files are for benchmarks, architecture checks, or Personal Corrections metric cross-checks; they are not part of the app runtime and are not distributed as the user-facing CLI.

Third-Party Modules And Open-Source Licenses

License

VoxFlow is distributed under GPL-3.0-or-later. Third-party components keep their original license notices and attribution. See docs/third-party-licenses.md.

Third-Party Modules

Unified Modules and References

Type	Module / Source	Link	What It Is Used For
Third-party dependency	speech-swift (Qwen3ASR, NemotronStreamingASR, ParakeetStreamingASR, OmnilingualASR, Qwen3TTS, Qwen3Chat, KokoroTTS, MADLADTranslation)	GitHub	Local ASR/TTS/translation/chat runtime
Third-party dependency	WhisperKit	GitHub	Local Whisper transcription
Third-party dependency	FluidAudio	GitHub	Local ASR pipeline for Paraformer/SenseVoice
Third-party dependency	Sherpa-ONNX	GitHub	FunASR local inference runtime
Third-party dependency	onnxruntime (Vendor/CSherpaOnnx)	GitHub	Inference runtime bundled with Sherpa-ONNX
In-repo module	VoxFlowContextBoostKit	Repo path	OCR context hotword extraction
In-repo module	VoxFlowVoiceCorrectionKit	Repo path	Deterministic correction engine and benchmarks
In-repo module	agent-cli (Rust)	Repo path	Local terminal AI agent dispatching helper
Reference source	TypeWhisper	GitHub	Deterministic correction flow + focused observation learning (conceptual only; no source copy)
Reference source	FlashText	GitHub	Matching/replacement approach inspiration (no runtime reuse)
Reference source	JiWER	GitHub	Evaluation and benchmark cross-check reference
Reference source	OpenAI Evals	GitHub	Benchmark/test-case organization style reference
Reference source	LanguageTool	GitHub	Error-correction fixture and testing style reference

License and Attribution References

Path	What It Covers
LICENSE	Project-level license
SOURCE_ATTRIBUTION.md	Third-party source references and adaptation scope
MODIFICATIONS.md	Upstream adaptation notes
Packages/VoxFlowVoiceCorrectionKit/NOTICE.md	TypeWhisper-derived source licensing
Vendor/	Vendored runtime license declarations
Package.swift + NOTICE/LICENSE in Sources/ and Packages/	Component dependency and license declarations

Connect

Follow me on X: @Counterxing

WeChat

Scan the QR code below to add the author on WeChat and share feedback or usage notes.