Skip to content

Release 4.4.1

Choose a tag to compare

@github-actions github-actions released this 21 Aug 14:11
· 169 commits to main since this release

📝 更新日志 / Changelog (v4.4.1)

中文

新增功能

  • 键盘识别历史面板:键盘扩展按键新增「识别历史」,可在输入法内查看近期记录并插入;支持滑动重新识别或重新后处理
  • 保留失败识别记录:识别失败、超时或取消时保留录音,并在历史中展示失败原因,便于重新识别或后处理
  • 豆包识别模型选择优化:火山引擎改为直接选择流式/文件 1.0/2.0 模型,不再依赖多个开关组合
  • Qwen3-ASR 1.7B Int8:本地 Qwen3-ASR 新增 1.7B Int8 模型
  • 输入完成后切换输入法:识别及后处理完成后可自动切换回指定输入法,便于二次编辑
  • 常驻通知跳转:开启后台保活后,可选择点击常驻通知打开的页面

改进优化

  • 录音重新识别:当选择流式识别模型时,重新识别会自动切换到文件识别引擎或模型,提高重新识别成功率(X-ASR 支持流式回放)
  • 供应商设置文案:统一各供应商的识别模型、语言与请求地址等标签
  • Pro 介绍:优化 Pro 版介绍页面

问题修复

  • 流式预览重复:增加更详细的流式预览过程日志,有相关问题可在导出详细日志后反馈排查
  • 滑块拖动生效:滑块仅在松手后写入设置,避免拖动本地模型线程数时反复加载
  • 自定义推理参数:修复 LLM 供应商自定义推理参数未正确保存的问题
  • 历史录音错绑:取消识别后不再误用上一段录音

English

New Features

  • Keyboard History Panel: New keyboard extension button to browse and insert recent transcripts; swipe to re-recognize or re-process
  • Keep Failed Recognition Records: Failed, timed-out, or cancelled sessions keep audio and appear in history with a reason, so they can be re-recognized or re-processed
  • Doubao Model Picker: Volcano Engine now uses a direct model picker for streaming/file 1.0/2.0 instead of stacked toggles
  • Qwen3-ASR 1.7B Int8: Added the Qwen3-ASR 1.7B Int8 local model
  • Switch IME After Input: After recognition (and post-processing) finishes, optionally switch back to a chosen IME for further editing
  • Keep-Alive Notification Tap: When background keep-alive is on, choose which page the persistent notification opens

Improvements

  • Recording Re-recognition: When a streaming model is selected, re-recognition automatically switches to a file recognition engine or model to improve success rate (X-ASR supports streaming replay)
  • Vendor Setting Labels: Unified recognition model, language, and endpoint labels across vendors
  • Pro Intro: Refined the Pro introduction page

Bug Fixes

  • Streaming Preview Logs: Added more detailed streaming-preview process logs; export detailed logs if you hit related issues
  • Slider Commit: Slider values apply only after release, avoiding repeated local-model reloads while dragging thread count
  • Custom Reasoning Params: Fixed LLM vendor custom reasoning parameters not being saved
  • History Audio Mix-up: Cancelled recognition no longer reuses the previous audio