douvo

Douvo is a lightweight macOS voice input application designed for Apple Silicon Macs running macOS 14.0 or newer. It leverages Doubao ASR (Automatic Speech Recognition) as its primary speech-to-text engine, with optional AI post-processing to clean up transcripts before insertion. The app operates as a menu bar utility with a floating recording overlay. Users press a configurable trigger key to start recording, speak, and then press again to stop and insert the recognized text at the current cursor position in any application. A hold-to-talk mode is also available. Douvo supports three ASR paths: Doubao Web, Doubao Android, and Youdao Bage Shuo. Multiple paths can be selected simultaneously, running in parallel with results merged through AI post-processing. The default path is Web. Key features include: - Dictate into any app with short-press or hold-to-talk recording - Optional AI post-processing to fix wording, punctuation, filler words, tone, and output style - Voice-driven editing of selected text (up to 500 characters) - On-demand translation with a configurable translation key - Custom vocabulary and context injection (project terms, paths, names, user identity, current time, frontmost app, window title) - Choice between local MLX models on device or remote LLM providers - Menu bar UI, floating recording overlay, appearance switching, clipboard-aware insertion, and local diagnostics For AI post-processing, Douvo offers local models including S1-mini by Superwhisper (English-only, 335 MB), Spark-X2.5 1.7B and 4B (Chinese correction and reasoning), or remote LLM providers. Output styles include Natural, Concise, Structured, and Custom with tunable style strength. The app requires Microphone and Accessibility permissions. It is distributed as a DMG from GitHub Releases or via Homebrew Cask. Current builds are not notarized, so macOS may require manual quarantine removal on first launch. Douvo is built for mixed Chinese/English input workflows where mainstream local speech models struggle with recognition quality. It wraps the strong Doubao ASR path as an app-agnostic voice input tool with post-processing, vocabulary, translation, selected-text editing, clipboard-aware insertion, and local diagnostics. The project depends on observed Doubao and Youdao Bage Shuo client behavior and is not an official API, SDK, or integration for either service. Users need a valid Doubao account and must log in themselves. Audio sent for recognition is processed by Doubao's service. The project is released under the MIT License.