Release 4.4.2

📝 更新日志 / Changelog (v4.4.2)

中文

新增功能

• Gemini Transcribe：Gemini 新增 Transcribe 调用模式
• 识别历史时间轴：历史详情展示录音、识别、润色与文字写入各阶段耗时；弹窗支持按高度自适应

改进优化

• LLM 连通性测试：测试结果分段展示连接、首字、输出耗时与连接复用，并标明是否支持流式
• 识别与键盘响应：异步提交历史、缓存键盘首帧 UI，并更广泛地复用网络连接池，识别延迟更低、上屏更少卡顿、打开更快
• AI 润色文案：界面将「后处理」统一为「润色」
• 打字机效果：关闭后改为润色完成后一次性上屏
• Qwen3-ASR 1.7B 结果清洗：去掉模型标记，避免内部标签写入正文
• 收起键盘中断润色：隐藏键盘时取消进行中的 AI 润色并提交原文，避免下次按麦无响应
• 润色超时与取消：为 AI 润色加上超时上限；取消后不再错误走非流式回退
• [Pro] 热词预加载：启动时预构建音素索引，降低首次识别的后处理延迟

问题修复

• Bridge 模块：优化与 Bridge 模块的通信管理
• 后台通知语言：保活与悬浮等通知将跟随应用语言
• 悬浮球性能：未录音时不再持续刷新悬浮球光环特效

---

English

New Features

• Gemini Transcribe: Gemini adds a Transcribe call mode
• History Timing Timeline: History details show time spent on recording, recognition, polish, and text commit; the dialog adapts to available height

Improvements

• LLM Connectivity Test: Test results show connect, first-token, and output latency, connection reuse, and whether streaming is supported
• Faster Recognition and Keyboard: Async history commit, cached first-frame UI, and broader HTTP connection-pool reuse for lower recognition latency, less commit jank, and faster keyboard open
• AI Polish Wording: UI now uses “polish” instead of “post-processing”
• Typewriter Effect: When off, the full result is applied at once after polish finishes
• Qwen3-ASR 1.7B Cleanup: Strips model markers so they are not committed as text
• Cancel Polish on Keyboard Hide: Hiding the IME cancels in-flight AI polish and commits the raw transcript, so the next mic tap is not ignored
• Polish Timeout and Cancel: AI polish now has a timeout cap; cancelled requests no longer fall back to non-streaming
• [Pro] Hotword Preload: Builds phoneme indexes at startup to reduce first-recognition post-processing delay

Bug Fixes

• Bridge Module: Improved communication management with the Bridge module
• Notification Language: Keep-alive, floating, and similar notifications now follow the app language
• Floating Ball Performance: Aura effects no longer keep refreshing when not recording