SayIt v0.2.0

SayIt v0.2.0

Added

• Push-to-talk and hands-free now accept F13–F24. Windows recognises these keys, but
  keyboards rarely have them, so they almost never clash with another program's shortcuts. A
  programmable keyboard or an external device with buttons can emit one of them purely to
  trigger recording, and a remapping tool such as PowerToys Keyboard Manager can point a key
  you never use at one of them.

Improved

• Recordings are written to disk the moment you release the key, rather than after
  recognition finishes. Speak for at least half a second and the audio is already a file;
  from then on a recognition timeout, a dropped connection or even force-quitting the app
  leaves it in History, where you can play it back or run recognition again. Previously the
  recording was saved only once the whole pipeline had finished, so anything that went wrong
  in between took the audio with it. (Provided you have not turned off history or recording
  retention in Settings.)
• Entries with no text now say why. History distinguishes between speech recognition
  returning no text, the connection dropping so the audio never reached the server, and timing
  out while waiting for a result — three situations that call for looking in completely
  different places.
• Text insertion no longer leaves the overlay spinning indefinitely. Some programs,
  particularly those running as administrator, make the insert call hang. The overlay now
  releases the interface after a dozen seconds or so, and the recognised text is already in
  History to copy from.
• AI cleanup switches off the model's thinking mode, so text arrives sooner. Many models
  reason before answering by default, which a short cleanup task does not need but still costs
  you several seconds. This applied to DeepSeek, Qwen, MiMo and Zhipu only; services added
  through a custom endpoint — aggregating gateways, relays, self-hosted endpoints — kept
  running with thinking on. All OpenAI-compatible services are now covered, with nothing to
  configure; the few that reject these parameters (OpenAI proper and Azure OpenAI) are skipped
  automatically and keep working as before.

Fixed

• The overlay no longer sits on "Processing" and ends with nothing — no text, and no entry in
  History. If the connection to the server dropped mid-recording (network hiccups and waking
  from sleep can both do it), SayIt carried on recording, then waited a minute after you
  released the key before quietly returning to idle: no warning, and that recording was gone.
  It now tells you the moment the connection drops and ends the recording there, instead of
  letting you talk into a connection that is no longer live. If the request cannot be sent at
  all when you finish, it says so immediately rather than waiting it out.
• The top of the overlay pill is no longer cut off when the Windows text size is enlarged.
  With Settings → Accessibility → Text size at anything other than 100%, the overlay's content
  was taller than the window itself and the excess was clipped — the top of the pill looked as
  if something was pressing on it, and the microphone name above it could be half visible. The
  tray icon's right-click menu was affected the same way, with "Quit" at the bottom possibly
  cut off. Both windows now scale with that setting, so everything stays fully visible wherever
  the slider is.

---

简体中文

新增

• 可以把 F13–F24 设为录音快捷键：这些按键 Windows 认得，但键盘上通常没有，因此几乎不会和其他程序的快捷键冲突。可编程键盘或带按键的外接设备可以把某个键输出成 F13–F24，专门用来触发录音；用改键工具（如 PowerToys 键盘管理器）把一个不常用的按键映射过来也可以。「按住说话」和「免提模式」都支持。

优化

• 录音在你松开按键的那一刻就写入本地，不再等识别结果：只要说满半秒，这段录音就已经存成文件，此后无论识别超时、连接断开还是应用被强制关闭，它都在历史记录里，可以回放，也可以点「重新识别」重跑一遍。此前录音要等整个识别流程走完才保存，中途出岔子就跟着一起没了。（前提是没有关掉设置里的历史记录或录音保留。）
• 没出文字的记录会写明原因：历史里那些空记录现在会区分「语音识别没有返回文本」「连接已断开，这段语音没有送达服务端」和「等待识别结果超时」，不再一律显示成同一句话——这三种情况该查的地方完全不同。
• 文字插入卡住时不会再永久停在「处理中」：个别程序（尤其是以管理员权限运行的）会让插字动作迟迟不返回，此前悬浮窗就一直转下去；现在最多再等十几秒便把界面放开，识别出的文字已经保存在历史记录里，可以从那里复制。
• AI 润色不再让模型「先想一想」，出字更快：很多模型默认开启思考模式，润色这种短文本任务并不需要它先推理一轮，却要为此多等好几秒。此前只有 DeepSeek、通义千问、MiMo、智谱这几家会被自动关闭，用自定义地址接入的其他服务（各类聚合网关、中转站、自建端点）仍然带着思考在跑。现在所有 OpenAI 兼容服务都会自动关闭，不需要任何设置；个别不接受这类参数的服务（OpenAI 官方、Azure OpenAI）会自动跳过，照常可用。

修复

• 说完话悬浮窗一直停在「处理中」，最后既没有文字、历史记录里也找不到这段录音：录音期间与服务器的连接如果断了（网络波动、休眠唤醒之后都可能发生），此前 SayIt 察觉不到——它会继续录下去，松开按键后再空等一分钟，然后悄悄回到待机，你收不到任何提示，那段话也就没了。现在连接一断当场提示并结束这次录音，不再让你对着已经断开的连接继续说；结束时如果请求根本发不出去，也会立刻说明原因，不再空等。
• 把系统的「文本大小」调大之后，悬浮窗胶囊顶部被切掉一条：Windows 设置 → 辅助功能 → 文本大小只要不是 100%，悬浮窗里的内容就会比窗口本身高，超出的部分被直接裁掉——胶囊顶部看起来像被什么东西压住了，胶囊上方的麦克风名字也可能只露出一半。托盘图标的右键菜单同样受影响，最下面的「退出」可能显示不全。现在这两个窗口会跟着这项设置一起放大，滑块拉到多少都能完整显示。