[github-actions]

🤖 I have created a release beep boop

0.3.0 (2026-11-07)

中文

新功能

• 电脑：粘贴后检查文字是否真的写进了输入框，没写进去时会提示。
• 电脑：没写进去、处理期间切换了窗口或识别失败时，结果以卡片留在悬浮窗上，可以直接粘贴或复制。
• 电脑：新增「复制上一条结果」，托盘菜单、快捷键和命令行都能用。
• 电脑：可以导入音频和视频文件转成文字，多个文件排队处理，可选 AI 逐段润色，导出 SRT 字幕或 TXT 文本。
• 电脑：词典新增「学习建议」，从 AI 润色的改动中找出常被识别错的词，可以逐条或批量加入词典、忽略，并能撤销。
• 电脑：规则页会提示重复、互相改回或与词典冲突的规则。
• 电脑：新增自动备份：每周一次，安装更新和替换导入前也会先备份，备份位置可以放到同步盘。
• 电脑：候补模型可以在服务出错时也自动切换，候补列表可以加入本地模型；选用本地识别时，录音不会因为出错而上传。
• 电脑：历史详情显示实际完成识别和润色的服务商、各环节耗时的时间线，以及术语表是否发出。
• 电脑：新增「可以说话」提示音。
• 电脑：「语音模型」页可以用最近一条录音试识别。
• 电脑：用本地模型重新识别时，可以选择不裁剪静音。
• 电脑、手机：语音识别新增豆包语音、腾讯云、讯飞（听写与实时转写）、百度、阶跃星辰、智谱 GLM-ASR 和小米 MiMo-ASR；在电脑上，会边说边识别的服务在说话时实时显示文字。
• 电脑、手机：阿里云百炼的 Fun-ASR 识别时带上术语表。
• 电脑、手机：需要多项凭据的服务商（如 AppId 加密钥）可以分项填写，测试连接会指出缺了哪一项。
• 电脑、手机：服务商设置新增「配置方法」，列出开通步骤和直达链接；识别服务商标出是否边说边识别、热词怎么发送。
• 电脑：「实时预览」分成「服务商的实时文字」和「本地模型预览」两个开关，悬浮窗的预览标签写明文字来自哪里，例如「预览 · 豆包语音」。

问题修复

• 电脑：窗口多、有置顶窗口或全屏程序时，悬浮窗不再被盖住。
• 电脑：连续听写时，新的悬浮窗不再被上一次的隐藏一起藏掉。
• 电脑、手机：中继连上后不再一直显示「认证中」，窗口打开时发生的其他变化也不再漏掉。
• 电脑：更新后窗口如果显示的不是新版本的界面，会自动重新加载一次。
• 电脑：取消听写后，原本为空的剪贴板会重新清空；「仅复制到剪贴板」时不再误报「窗口已切换」。
• 电脑：单键触发被系统停用或断开后自动恢复，不用重启；收不到松开时，录音不再一直进行。
• 电脑：本地识别出错后不再卡住，显卡出错时自动改用 CPU。
• 电脑：取消听写后可以马上开始下一次，不用等被取消的识别结束。
• 电脑：实时预览模型恢复可以下载。
• 电脑：润色结果被截断时如实说明，并保留识别原文。
• 电脑：麦克风就绪的状态不再因为程序繁忙而丢失。
• 电脑：更正词典帮助里关于术语表发给哪些服务的说明。
• 电脑：用实时识别模型听写的内容计入用量。
• 电脑（Linux）：AppImage 的启动程序改为所有用户可读、可执行，在 firejail 这类以其他用户身份运行的环境里也能启动。
• 电脑：系统钥匙串不可用或被锁时，先说明需要哪个钥匙串（Linux 为 GNOME Keyring 或 KWallet，macOS 为钥匙串，Windows 为凭据管理器）再退出，不再没有任何提示就退出。
• 电脑、手机：AI 润色的内置预设不再遗漏口述里的要求、条件、例外、数字和名称。

改进

• 电脑：本地模型识别长录音时按停顿分段，结果更完整。
• 电脑：Qwen3-ASR 的本地识别组件升级，已下载的模型不用重新下载。
• 电脑：没有会改变输出方式的场景时，按下快捷键后立即开始录音。
• 电脑：新安装默认「松开后继续录音」150 毫秒，不再丢掉最后几个字；已安装的设置不变。
• 电脑：边说边识别没有得到文字、改为整段识别时，历史详情写明原因。

其他

• 文档站同步介绍了以上新功能。
• 文档站新增「服务商设置」页，逐个说明各服务商怎么开通、凭据填在哪里。
• 电脑、手机：官方安装包会向托管中继和内置服务证明自己的版本，为以后只允许官方安装包使用这两项服务做准备；目前只统计，不影响使用。
• 从 0.2.x 第一次更新到 0.3.0 时没有更新前的自动备份，因为旧版本还没有这项功能。
• Mac 上其他应用全屏时看不到悬浮窗，这是目前已知的限制。
• 请把电脑和手机都更新到 0.3.0：0.2.x 读不出「润色结果不完整」的记录，同步会停在这类记录之前。

English

Features

• Desktop: after a paste, Voltip checks that the text really reached the input field and tells you when it didn't.
• Desktop: when the text wasn't written, the window changed while processing or recognition failed, the result stays on the overlay as a card you can paste or copy from.
• Desktop: "Copy last result", from the tray menu, a shortcut or the command line.
• Desktop: import audio and video files and turn them into text, several files in a queue, with optional AI polishing piece by piece and export to SRT subtitles or TXT.
• Desktop: "Suggestions" in the dictionary finds the words recognition keeps getting wrong from what AI polishing changed; add them to the dictionary or ignore them one by one or in batches, and undo.
• Desktop: the Rules page points out rules that duplicate each other, undo each other or clash with the dictionary.
• Desktop: automatic backups: once a week, and also before an update installs or an import replaces data; the backup folder can be on a synced drive.
• Desktop: fallback models can also take over when a service fails, and local models can join the fallback list; with local recognition selected, a failure never uploads the recording.
• Desktop: a history entry's details show the providers that actually recognised and polished the text, a timeline of how long each step took, and whether the glossary was sent.
• Desktop: a "Ready to speak" sound.
• Desktop: test recognition with the newest recording on the Speech models page.
• Desktop: a retake with a local model can keep the silence instead of trimming it.
• Desktop and Phone: new speech recognition services: Doubao Speech, Tencent Cloud, iFlytek (dictation and real-time transcription), Baidu, StepFun, Zhipu GLM-ASR and Xiaomi MiMo-ASR; on the desktop, the ones that recognise while you speak show the text live.
• Desktop and Phone: Alibaba Cloud Model Studio's Fun-ASR receives the glossary.
• Desktop and Phone: services that need several credentials (an AppId and a key, for example) take each one in its own field, and the connection test names the one missing.
• Desktop and Phone: a service's settings show how to set it up, with the steps and direct links; recognition services say whether they recognise while you speak and how hot words are sent.
• Desktop: Live preview is split into two switches, "Provider's live text" and "Local model preview", and the overlay's preview tag says where the words come from, such as "Preview · Doubao Speech".

Fixes

• Desktop: the overlay is no longer hidden behind other windows when many windows are open, a window is pinned on top or a program runs full screen.
• Desktop: starting a dictation right after the last one no longer hides the new overlay.
• Desktop and Phone: the relay status no longer stays on "Authenticating" after the relay connects, and other changes made while a window opens are no longer missed.
• Desktop: after an update, a window that shows an interface other than the new version's reloads itself once.
• Desktop: cancelling a dictation empties the clipboard again if it was empty before; "Copy to clipboard only" no longer says the window changed.
• Desktop: the single-key trigger comes back by itself when the system stops or cuts it off, without a restart; a recording no longer runs on when its key release never arrives.
• Desktop: local recognition no longer gets stuck after an error, and moves to the CPU when the graphics card fails.
• Desktop: after cancelling a dictation you can start the next one straight away, without waiting for the cancelled recognition.
• Desktop: the live preview model can be downloaded again.
• Desktop: polished text that was cut off is reported as such, and the transcript is kept.
• Desktop: the state saying the microphone is ready is no longer lost when Voltip is busy.
• Desktop: the dictionary's help text now says correctly which services receive the glossary.
• Desktop: takes dictated with a real-time recognition model now count towards usage.
• Desktop (Linux): the AppImage's launcher is readable and executable by every user, so it starts under firejail and other sandboxes that run it as another user.
• Desktop: when the system keychain is unavailable or locked, Voltip says which keychain it needs (GNOME Keyring or KWallet on Linux, the Keychain on macOS, Credential Manager on Windows) before it quits, instead of quitting without a word.
• Desktop, Phone: the built-in AI polish presets no longer drop the requests, conditions, exceptions, numbers or names in what you said.

Improvements

• Desktop: local models split long recordings at pauses, so less is left out.
• Desktop: the local recognition component Qwen3-ASR runs on is updated; downloaded models don't need to be downloaded again.
• Desktop: when no scene changes the output mode, recording starts as soon as you press the shortcut.
• Desktop: new installations keep recording for 150 ms after release by default, so the last words are not cut off; existing settings are unchanged.
• Desktop: when recognition while you speak returns no text and the whole take is recognised instead, the history details say why.

Other

• The documentation site describes the new features.
• The documentation site has a new Provider setup page that explains, provider by provider, how to sign up and where each credential goes.
• Desktop and Phone: official packages now prove their version to the hosted relay and the built-in services, in preparation for keeping these two services to official packages; for now this is only counted and changes nothing.
• The first update from 0.2.x to 0.3.0 makes no backup beforehand, because the older version doesn't have the feature yet.
• On a Mac, the overlay doesn't show while another app is full screen; this is a known limitation for now.
• Please update both your computer and your phone to 0.3.0: 0.2.x can't read entries whose polished text was cut off, so syncing stops before such an entry.

---

This PR was generated with Release Please. See documentation.