🏗️ This project's AI agent harness is built with Harness Engineering.

A macOS voice input tool — hold a hotkey to record, release to get polished text auto-pasted into any input field.

Local AI transcription → LLM text optimization → auto-paste. Private, fast, effortless.

Features

• Press & Speak — Hold Option+Space (customizable) to record, release to transcribe + optimize + paste. No window switching needed.
• Privacy-First Local STT — Six AI engines (Whisper, SenseVoice, Paraformer, Moonshine, FireRedASR, Zipformer CTC) run entirely on your Mac via Metal GPU. Audio never leaves your device.
• AI-Powered Polish — LLM auto-corrects grammar, removes filler words, and structures your text. Built-in technical term correction (e.g. phonetic Chinese → "React"), with custom vocabulary support.
• Auto-Paste Anywhere — Optimized text is automatically pasted into your active input field — Slack, WeChat, VS Code, browsers, any app.
• 99+ Languages — 6 engines, 12 models, covering 99+ languages. The system recommends the best model based on your language.
• On-Demand Models — Lightweight app, download only the STT models you need. One-click switch, progress display, smart recommendations.
• ESC to Cancel — Cancel at any stage (recording, transcribing, optimizing) by pressing ESC.
• History — Review past transcriptions with original and AI-optimized text side by side.
• Custom Vocabulary — Add professional terms, names, and product names. Auto-learning detects corrections and validates via LLM.
• Dark & Light Themes — Dual theme support to match your preference.

System Requirements

• macOS 11.0+
• Apple Silicon processor recommended (for Metal GPU acceleration)

Getting Started

1. Install

Download the latest .dmg from GitHub Releases, open it, and drag Input 0 into your Applications folder.

Using Homebrew:

brew install --cask input0

On first launch, macOS may show a security warning. Go to System Settings → Privacy & Security and click Open Anyway.

2. Grant Permissions

Input 0 needs these macOS permissions to work properly. The app will prompt you on first launch:

Permission	Why
Microphone	Records your voice
Accessibility	Simulates Cmd+V to paste text into other apps

Go to System Settings → Privacy & Security to grant each permission, then restart the app.

3. Configure LLM API Key ⚡

This step is required. Without an API key, voice transcription still works, but the AI text optimization (grammar fix, filler removal, structuring) will be skipped.

1. Open Input 0 → Click the ⚙️ Settings icon in the sidebar
2. Find the LLM API section
3. Enter your settings:

Field	Description	Default
API Key	Your OpenAI (or compatible) API key	(empty — must fill in)
API Base URL	Service endpoint	https://api.openai.com/v1
Model	LLM model name	gpt-4o-mini

Using a third-party provider? (e.g. Azure OpenAI, Groq, local Ollama, or any OpenAI-compatible API)

• Change the API Base URL to your provider's endpoint
• Set the Model to your provider's model name
• Enter the corresponding API Key

Click Test Connection to verify your API key works.

4. Download an STT Model ⚡

This step is required. You need at least one model to transcribe voice to text.

1. Open Input 0 → Go to Settings → Model section
2. The app recommends the best model based on your language setting — follow the recommendation or pick your own
3. Click Download next to the model you want
4. Wait for the download to complete (progress bar shown)

Which model should I choose?

If you speak...	Recommended model	Size
Chinese	SenseVoice Small	~228 MB
English	Whisper Large v3 Turbo	~1.5 GB
Japanese / Korean	SenseVoice Small	~228 MB
Cantonese	Paraformer Trilingual (zh/en/yue)	~234 MB
Multiple languages	Whisper Large v3	~2.9 GB
Chinese SOTA (highest accuracy)	FireRedASR Large v1	~1.74 GB
Want the fastest	Moonshine Base (EN-only)	~274 MB
Want smallest download	Whisper Base	~142 MB

5. Start Using

1. Press and hold Option+Space (or your custom hotkey)
2. Speak naturally
3. Release the key — your text is transcribed, polished, and auto-pasted

That's it! 🎉

Troubleshooting

Model Download Issues

If a model download fails or gets stuck:

1. Check your network — Models are downloaded from Hugging Face. If you're in a region where Hugging Face is slow or blocked, consider using a VPN or proxy.
2. Retry — Close and reopen the app, then try the download again. Partial downloads will resume from where they left off.
3. Disk space — Make sure you have enough free space. The largest model (Whisper Large v3) needs ~2.9 GB.
4. Manual download — As a last resort, you can manually download model files and place them in:

   ~/Library/Application Support/com.input0/models/<model-id>/
   

   See the model registry for exact file names and download URLs.

No Text Output After Speaking

• Make sure at least one STT model is downloaded and selected
• Check that your Microphone permission is granted
• Check that Accessibility permission is granted (required for auto-paste)

AI Optimization Not Working

• Verify your API Key is set correctly in Settings
• Click Test Connection to check the API is reachable
• If using a third-party provider, double-check the API Base URL and Model name

Hotkey Not Responding

• Make sure no other app is using the same hotkey
• Try changing the hotkey in Settings
• Restart the app after changing the hotkey

All STT Models

Model	Size	Best For
Whisper Base	~142 MB	Fast & lightweight, good for daily use
Whisper Small	~466 MB	Balanced accuracy and speed
Whisper Medium	~1.4 GB	Excellent multilingual accuracy
Whisper Large v3	~2.9 GB	Highest accuracy, 99 languages
Whisper Large v3 Turbo	~1.5 GB	Top accuracy for English & multilingual
Whisper Large v3 Turbo Q5	~547 MB	Quantized high-accuracy, balanced size
SenseVoice Small	~228 MB	Best for Chinese / Japanese / Korean
Paraformer Chinese	~217 MB	Chinese-optimized, ultra-fast inference
Paraformer Trilingual	~234 MB	Chinese + English + Cantonese (only Cantonese-capable model)
Moonshine Base (EN)	~274 MB	English-only, ~5x faster than Whisper
FireRedASR Large v1	~1.74 GB	Chinese ASR SOTA (CER ≈ 2%), for maximum accuracy
Zipformer CTC (Chinese)	~350 MB	Offline Chinese CTC, next-gen Kaldi, lightweight

License

This project is licensed under CC BY-NC 4.0. You are free to share and adapt, but not for commercial use.