qwen-audio

VRCS is a subtitle, translation, language learning, and AI tool for VRChat.

The code for AAAI 2025 “Large Language Models Are Read/Write Policy-Makers for Simultaneous Generation”

Realtime full-duplex voice Agent plugin for DeepSeek Harness WebUI and future WeChat Mini Program clients

MV -> karaoke subtitles: DashScope ASR word-level timings (JSON) -> .ass with smooth per-character colouring, plus a Tkinter timeline editor and ffmpeg burn-in.

给 DSH 里的 agent 装一对耳朵：整段波形直接交给会听音频的模型（本地 Qwen2-Audio 7B，4bit 约 4.5G 显存），不走先转写再读文字那条路。模型听完描述语气、情绪、停顿和背景，插件再算基频、能量、频谱重心和停顿分布核对，两边对得上才下结论。超过一分钟先问要不要分段；能盯 QQ 语音，静默 9 秒凑齐才算说完；八个本地模型可下可切，也能填 Key 走云端。Windows 10/11，MIT，v0.1 测试版