抱歉最近在忙着跨国搬家，更新频率有所下降，欢迎提 issue/PR，紧急问题会尽快修。Sorry, I've been busy with an international move lately, so updates have slowed down — issues and PRs welcome, urgent ones will be fixed ASAP.

中文 | English

---

中文

update：官方词库管理Skill，帮你大幅提高识别准确率

• 语音识别：内置本地识别引擎、媲美云端引擎准确率；支持多家云端引擎厂商；支持流式识别、边说边出字，说完无需等待、快速输入；
• 文本处理：内置润色、Prompt优化、翻译功能，可自定义添加任意处理模版（比如改人设、改语气、小语种翻译等等）；
• 模型接入：支持主流厂商API接入；文本处理支持使用Ollama接本地模型；
• 词汇管理：支持热词、映射词，2种模式。热词用于校正语音识别引擎，映射词可作为兜底或个性化场景使用（如 Web coding -> Vibe Coding, "我的邮箱地址" -> xxx@gmail.com）；
• 历史记录：存储所有历史识别记录，包括原始文本和处理后文本，支持导出CSV；
• 配套Skill：真正做到100%准确率，打造只属于你的输入法，点这里安装Skill后跟你的agent说"Qwen3.5 不要识别成 Queen 3.5"，他就能自动帮你管理热词和映射词，同类错误不再犯

立即体验

方式一：直接下载DMG（推荐）

两个版本，共享配置文件，可随时替换安装：

系统要求：macOS 14+ (Sonoma)

界面预览

查看演示视频

为什么做Type4Me

市面上语音输入法，至少命中以下问题之一：贵（$30/月）、封闭（不可导出记录）、扩展性差（不能自定义Prompt）、慢（强制优化及网络延迟）

作为某最贵识别工具曾经的粉丝，心路历程就是：「它怎么可以这么好用，但又这么难用」 以及也不必所有的话都说的这么工工整整规规矩矩。

使用Tips

• 语音识别：
  • 推荐使用云端模型，成本极低（我高强度用说了5w字=5小时，对应5块人民币，豆包语音注册送40小时，配置指引）
  • 尽管本地模型效果还不错，但十分占用内存，内嵌Sense Voice用于流式识别（2GB内存占用）、Qwen3 ASR做校准（8GB内存占用），你也可以单独开其中一个，但体验不佳，Sense Voice中文不错、但英文单词十分拉垮。
• 文本处理（接入LLM）：
  • 依旧推荐使用云端模型，接入Coding Plan API，这类轻量文本处理Token消耗肉眼不可见；
  • LLM本地跑的内存占用比语音识别还高，而且效果相比云端模型相去甚远；
  • 不要使用思考模式，推荐轻量模型。作者自己用的是Seed-2.0-lite。例如Minimax M2.7无法关闭思考，处理时间会非常长。对于我们这种轻量文本处理完全没有必要，牺牲体验也换不到效果。
    • 如果你发现你的处理时间很长，请把你使用的厂商和模型告诉我，我看看代码里是否成功关闭思考（目前没有遍历测试所有API）
• 强烈建议搭配配套Skill使用：市面上所有的语音输入法，专有名词均无法做到很好的识别（例如：Qwen 3.5），搭配Skill使用1-2天，你将彻底迈入100%识别准确率

详细功能介绍

语音识别（略）

文本处理：需配置API Key，效果受模型影响，可自行调整/添加Prompt

每个模式可以绑定独立的全局快捷键，支持「按住说话」和「按一下开始/再按停止」两种方式。

Prompt 变量高级玩法

Prompt 模板支持三种变量，让语音输入从"听写"升级为"语音命令"：

用法示例：

词汇管理

• ASR 热词：添加专有名词（如 Claude、Kubernetes），提升识别准确率
• 片段替换：语音说「我的邮箱」，自动替换为实际邮箱地址

架构概览

ASR Provider 架构设计为可插拔：实现 ASRProviderConfig（定义凭证字段）和 SpeechRecognizer（实现识别逻辑），注册到 ASRProviderRegistry 即可添加新引擎。

参与贡献

欢迎提交 PR/Issue，这个项目是我全部自己用 Claude Code 写的。

对于 PR，即便有 bug/代码质量不好，我最常跟 Claude 说的一句话就是不要漏了人家的贡献。你大不了合完再改。

致谢

• SenseVoice - Alibaba FunAudioLLM
• streaming-sensevoice - @pengzhendong
• asr-decoder - @pengzhendong
• sherpa-onnx - k2-fsa
• Qwen3-ASR - Alibaba Qwen
• mlx-qwen3-asr - @moona3k

演示视频

demo

许可证

MIT License

---

English

Update: Official Vocabulary Management Skill to dramatically improve recognition accuracy

• Speech Recognition: Built-in local recognition engine with accuracy rivaling cloud engines; supports multiple cloud ASR providers; real-time streaming recognition with instant text output;
• Text Processing: Built-in voice polish, prompt optimization, and translation; add any custom processing templates (persona, tone, minority language translation, etc.);
• Model Integration: Supports mainstream provider APIs; text processing works with Ollama local models;
• Vocabulary Management: Two modes: hotwords and snippet replacements. Hotwords improve ASR accuracy for proper nouns; snippets enable personalized substitutions (e.g., "Web coding" -> "Vibe Coding", "my email" -> xxx@gmail.com);
• History: Stores all recognition records including raw and processed text, with CSV export;
• Companion Skill: Achieve 100% recognition accuracy. Install the Skill and tell your agent "Don't recognize Qwen3.5 as Queen 3.5" to automatically manage hotwords and snippets. Same mistakes won't happen again.

Get Started

Option 1: Download DMG (Recommended)

Two editions, sharing the same config files. You can switch between them at any time:

System requirements: macOS 14+ (Sonoma)

DMG shows "damaged" or app won't open?

Solution:

• Step 1: Open Terminal and run:

  xattr -d com.apple.quarantine /Applications/Type4Me.app

  spctl --master-disable
  

• Step 2: Go to System Settings > Privacy & Security > "Allow applications from", select "Anywhere"

• Step 3: Open the DMG and drag Type4Me to the Applications folder.

• Step 4 (Optional): Revert "Allow applications from" back to its previous setting.

Apple Developer certification is still pending. This is a normal macOS Gatekeeper prompt that can be resolved with the steps above.

Option 2: Give this repo link to your AI agent and let it deploy for you

Screenshots

Watch demo video

Why Type4Me

Every voice input tool on the market hits at least one of these: expensive ($30/month), walled garden (can't export history), inflexible (no custom prompts), slow (forced optimization + network latency).

As a former fan of the most expensive transcription tool out there, the journey was: "How can it be this good and this frustrating at the same time?" And honestly, not everything you say needs to sound perfectly polished.

Usage Tips

• Speech Recognition:
  • Cloud models recommended. Extremely affordable (50k characters = 5 hours of heavy use costs ~$0.70 USD. Volcano/Doubao gives 40 free hours on signup, setup guide)
  • Local models work well but are memory-intensive: SenseVoice for streaming (2GB RAM), Qwen3-ASR for calibration (8GB RAM). You can run just one, but the experience is compromised. SenseVoice handles Chinese well but struggles with English words.
• Text Processing (LLM):
  • Cloud models still recommended. Token cost for lightweight text processing is negligible;
  • Running LLM locally uses more memory than ASR and the quality gap vs. cloud models is significant;
  • Do NOT use reasoning/thinking mode. Use lightweight models. The author uses Seed-2.0-lite. Some models (e.g., Minimax M2.7) can't disable reasoning, resulting in very long processing times. For lightweight text processing, it's completely unnecessary.
    • If your processing time is long, tell me which provider and model you're using so I can check if reasoning is properly disabled in the code (not all APIs have been tested)
• Strongly recommended: Use with the companion Skill. No voice input tool handles proper nouns perfectly (e.g., "Qwen 3.5"). After 1-2 days with the Skill, you'll achieve 100% recognition accuracy.

Feature Details

Speech Recognition (See above)

Text Processing: Requires API Key. Quality depends on model choice. Prompts are fully customizable.

Each mode can have its own global hotkey. Supports both "hold to talk" and "press to start / press to stop".

Advanced: Prompt Variables

Prompt templates support three variables, upgrading voice input from "dictation" to "voice commands":

Example:

Vocabulary Management

• ASR Hotwords: Add proper nouns (e.g., Claude, Kubernetes) to improve recognition accuracy
• Snippet Replacement: Say "my email" and it auto-replaces with your actual email address

Architecture Overview

The ASR provider architecture is fully pluggable: implement ASRProviderConfig (define credential fields) and SpeechRecognizer (implement recognition logic), register with ASRProviderRegistry, and you have a new engine.

Contributing

PRs and Issues are welcome. This entire project was built by the author using Claude Code.

Regarding PRs: even if there are bugs or code quality issues, the number one rule is to never drop anyone's contribution. You can always fix things after merging.

Acknowledgments

• SenseVoice - Alibaba FunAudioLLM
• streaming-sensevoice - @pengzhendong
• asr-decoder - @pengzhendong
• sherpa-onnx - k2-fsa
• Qwen3-ASR - Alibaba Qwen
• mlx-qwen3-asr - @moona3k

For AI Agents

If you're an AI agent (Claude Code, Cursor, Copilot, etc.) tasked with building, deploying, or contributing to Type4Me, here's everything you need.

Read these files first

1. CLAUDE.md - full architecture guide, credential storage, key files, development patterns, and how to add new ASR/LLM providers
2. Package.swift - Swift Package Manager dependencies and build targets
3. scripts/deploy.sh - the build & deploy pipeline (calls scripts/package-app.sh)

Prerequisites

• macOS 14.0+ (Sonoma)
• Xcode Command Line Tools: xcode-select --install
• Python 3.12: brew install python@3.12 (for local ASR servers)
• CMake: brew install cmake (only if building SherpaOnnx punctuation engine)

Build & deploy

# 1. Clone
git clone https://github.com/joewongjc/type4me.git && cd type4me

# 2. (Optional) Build SherpaOnnx punctuation engine (~5 min, needs cmake)
bash scripts/build-sherpa.sh

# 3. (Optional) Setup Qwen3-ASR server (needs python3.12, Apple Silicon only)
cd qwen3-asr-server && python3.12 -m venv .venv && source .venv/bin/activate && pip install -r requirements.txt && cd ..

# 4. Deploy (builds, bundles .app, signs, installs to /Applications, launches)
bash scripts/deploy.sh

# Subsequent updates
git pull && bash scripts/deploy.sh

Steps 2-3 are optional. Skipping them disables local ASR, but cloud ASR works fine.

Code signing & permissions

deploy.sh handles code signing automatically to preserve macOS permissions across rebuilds:

• First deploy auto-creates a self-signed certificate ("Type4Me Local", valid 10 years) if no signing identity exists. This may trigger a Keychain password prompt that requires human interaction.
• Subsequent deploys reuse the same certificate. Accessibility/Microphone permissions persist, no re-grant needed.
• After first launch, the user must grant Accessibility permission once (System Settings > Privacy & Security > Accessibility > enable Type4Me).
• To override signing identity: CODESIGN_IDENTITY="Your Cert" bash scripts/deploy.sh
• Fallback to ad-hoc signing: CODESIGN_IDENTITY="-" bash scripts/deploy.sh (Accessibility permission will reset each build)

Key architecture points

• Swift Package Manager project, no .xcodeproj needed
• Local ASR: dual-engine design. SenseVoice (streaming partial results) + Qwen3-ASR (final calibration via MLX/Metal). Both run as Python WebSocket servers managed by SenseVoiceServerManager
• Cloud ASR: 7 providers implemented (Volcano, OpenAI, Deepgram, AssemblyAI, Soniox, Bailian, Baidu)
• Credentials: stored at ~/Library/Application Support/Type4Me/credentials.json (mode 0600), never in code or environment variables. GUI apps cannot read shell env vars from ~/.zshrc
• ASR provider architecture: plugin-based. To add a new provider: implement ASRProviderConfig + SpeechRecognizer protocol, register in ASRProviderRegistry.all. See CLAUDE.md for details
• Audio format: 16kHz mono PCM16-LE, 200ms chunks (6400 bytes)
• Text injection: clipboard-based Cmd+V paste with save/restore

Demo Video

demo

Star History

License

MIT License