Tian2 Whisper
本地优先、跨设备语音输入工具
A local-first, cross-device voice input tool. A Mac Mini, NAS, or NVIDIA workstation runs Whisper speech-to-text offline; macOS and Windows clients handle the global hotkey, recording, real-time preview, and auto-paste into any active text field — with no monthly subscription and no audio leaving your LAN.
Dual-model realtime pipeline: a small model streams low-latency preview while Whisper Turbo Q4 delivers the final high-accuracy transcript. A local Qwen3 4B LLM handles punctuation, filler-word removal, and bilingual polishing — with safety validators that fall back to deterministic output if the LLM misbehaves.
Press-and-hold Fn
Hold to record, release to submit. Long pauses do not truncate. Custom hotkey combos supported.
Three modes
Faithful (light cleanup), Smart (filler words, self-corrections), Polish (full refinement) — all with LLM safety validators.
Cursor-aware
Reads preceding text from the active input field via accessibility APIs. Password fields excluded. Up to 1,500 characters of context.
LAN-only
Audio stays on your local network. Bearer tokens in macOS Keychain or Windows Credential Manager. Public-IP plain HTTP rejected.
Personal dictionary
User corrections feed back into ASR prompts and LLM polishing. Per-device JSON history with search — never uploaded.
Multi-backend
MLX Whisper on Apple Silicon, faster-whisper int8 on NAS, CUDA float16 on NVIDIA. Docker Compose or LaunchAgent.