Open Typeless

A macOS voice input tool with realtime dictation and LLM-assisted integrated dictation.

一款 macOS 语音输入工具，支持实时听写与大模型整合转写。

---

Architecture

flowchart LR
    User["User"] --> Key["Right Option Key"]
    Key --> Keyboard["Keyboard Service<br/>short press / long press / Esc cancel"]

    subgraph ElectronMain["Electron Main Process"]
      Keyboard --> PTT["Push-to-Talk Service"]
      PTT --> Perm["Permissions Service"]
      PTT --> Sound["Sound Service"]
      PTT --> Floating["Floating Window Manager"]
      PTT --> Insert["Text Input Service"]
      PTT --> ASR["ASR Service"]
      ASR --> Volc["Volcengine WebSocket Client"]
      PTT --> LLM["Doubao Formatter Service"]
    end

    subgraph Renderer["Renderer Process"]
      Recorder["Audio Recorder<br/>getUserMedia + AudioContext"]
      UI["Floating UI<br/>status + transcript"]
    end

    PTT -- "IPC: start/stop status" --> Recorder
    Recorder -- "PCM audio chunks" --> ASR
    ASR -- "status / interim / final result" --> UI
    ASR --> VolcCloud["Volcengine Streaming ASR"]
    LLM --> Ark["Doubao / Ark API"]
    PTT --> Insert
    Insert --> Target["Focused App Input Field"]

Loading

中文说明

项目简介

Open Typeless 是一个面向 macOS 的语音输入工具，围绕“尽量不打断工作流”来设计。

它提供两种输入模式：

• 听写模式：长按右侧 Option 键，实时转写，松手即输入
• 整合模式：短按右侧 Option 键开始识别，再短按一次结束，随后使用豆包大模型进行整理、纠错和结构化输出

功能特性

• 双模式语音输入：同时支持实时听写与整合转写
• 流式 ASR：基于火山引擎流式语音识别大模型
• LLM 整理：去口癖、去重复、处理自我修正、优化表达
• 结构化输出：自动分段，多事项时自动编号
• 悬浮反馈：通过浮窗展示当前模式、状态和转写内容
• 不抢焦点输入：识别完成后直接写入当前光标位置
• 全局快捷键：后台运行，无需手动切换回应用

系统要求

• macOS 12.0 或以上
• Node.js 22
• pnpm

快速开始

1. 安装依赖

pnpm install

2. 配置环境变量

复制 .env.example 到 .env：

cp .env.example .env

填写以下配置：

# Volcengine ASR
VOLCENGINE_APP_ID=your_app_id
VOLCENGINE_ACCESS_TOKEN=your_access_token
VOLCENGINE_RESOURCE_ID=volc.bigasr.sauc.duration

# Doubao / Ark
DOUBAO_API_KEY=your_ark_api_key
DOUBAO_MODEL=your_doubao_model_or_endpoint_id

3. 启动应用

pnpm start

4. 授权系统权限

首次使用需要授权：

• 麦克风权限：用于录音
• 辅助功能权限：用于文字插入
• 输入监控权限：用于全局监听右侧 Option 键

你可以在「系统设置 -> 隐私与安全性」中完成授权。

使用方法

听写模式

1. 长按右侧 Option
2. 浮窗顶部显示 听写模式
3. 下方实时显示 ASR 转写内容
4. 松开按键后，结果自动输入到当前光标位置

整合模式

1. 短按右侧 Option
2. 浮窗顶部显示 整合模式
3. 下方实时显示 ASR 内容
4. 再短按一次右侧 Option
5. 浮窗进入 Thinking
6. 豆包模型整理完成后，输出最终文本并自动输入

取消当前听写

• 在听写或整合录音过程中按 Esc 可取消本次输入

项目结构

src/
├── main.ts
├── preload.ts
├── renderer.ts
├── main/
│   ├── ipc/
│   ├── services/
│   │   ├── asr/           # Volcengine ASR session and WebSocket client
│   │   ├── keyboard/      # Global right-Option gesture detection
│   │   ├── llm/           # Doubao transcript formatter
│   │   ├── permissions/   # macOS permission checks
│   │   ├── push-to-talk/  # Core dictation orchestration
│   │   ├── sound/         # Start/stop sound cues
│   │   └── text-input/    # Insert text into focused app
│   └── windows/           # Floating window lifecycle and sizing
├── renderer/
│   └── src/
│       ├── modules/asr/   # Audio recorder and floating UI
│       └── styles/        # Floating window styling
├── shared/                # Shared types and IPC channel constants
└── types/                 # Global preload API typing

开发命令

pnpm start
pnpm typecheck
pnpm lint
pnpm package
pnpm make

常见问题

为什么第一次启动时麦克风会慢一点？

首次使用时浏览器侧需要完成 getUserMedia 和 AudioContext 预热，因此第一次响应通常会略慢于后续使用。

为什么按下按键后的前一小段语音也能识别？

当前实现会在 connecting 阶段就启动本地录音并缓存音频，等 ASR WebSocket 真正 ready 后再补发，尽量减少前几秒丢字。

如何修改快捷键？

当前默认键位是右侧 Option。可以在 keyboard.service.ts 中调整 triggerKey。

English

Overview

Open Typeless is a macOS voice input app designed around one principle: keep dictation fast, global, and interruption-free.

It supports two input flows:

• Realtime Dictation: hold the right Option key, speak, release to insert
• Integrated Dictation: tap the right Option key to start, tap again to stop, then use Doubao to rewrite and structure the text before insertion

Features

• Dual-mode voice input for direct dictation and LLM-assisted rewrite
• Streaming ASR powered by Volcengine speech recognition
• Transcript cleanup for filler words, repetitions, and self-corrections
• Structured output with paragraphing and conditional numbering
• Floating feedback window for mode, status, and transcript visibility
• Focus-safe insertion into the currently focused text field
• Global shortcut workflow that works while the app stays in the background

Requirements

• macOS 12.0+
• Node.js 22
• pnpm

Setup

1. Install dependencies

pnpm install

2. Configure environment variables

cp .env.example .env

Then fill in:

VOLCENGINE_APP_ID=your_app_id
VOLCENGINE_ACCESS_TOKEN=your_access_token
VOLCENGINE_RESOURCE_ID=volc.bigasr.sauc.duration
DOUBAO_API_KEY=your_ark_api_key
DOUBAO_MODEL=your_doubao_model_or_endpoint_id

3. Start the app

pnpm start

4. Grant macOS permissions

The app needs:

• Microphone for audio capture
• Accessibility for text insertion
• Input Monitoring for listening to the global right Option shortcut

Grant them in System Settings -> Privacy & Security.

How It Works

Realtime Dictation

1. Hold the right Option key
2. The floating window shows 听写模式
3. Live ASR text appears below
4. Release the key to insert the final transcript

Integrated Dictation

1. Tap the right Option key
2. The floating window shows 整合模式
3. Live ASR transcript appears below
4. Tap the right Option key again
5. The floating window switches to Thinking
6. The Doubao model rewrites the transcript and inserts the final output

Cancel

• Press Esc during dictation to cancel the current session

Core Modules

• Push-to-Talk Service: coordinates keyboard gestures, ASR lifecycle, LLM rewrite, preview, and insertion
• ASR Service: manages connection state, buffered audio, interim/final transcript handling
• Audio Recorder: captures microphone input in the renderer and streams PCM chunks
• Floating Window: displays mode labels, transcript text, and processing feedback
• Text Input Service: writes the final text into the focused app

Development

pnpm start
pnpm typecheck
pnpm lint
pnpm package
pnpm make

License

MIT