Skip to content

VoxtMini 0.1.0-beta.3

Pre-release
Pre-release

Choose a tag to compare

@hehehai hehehai released this 15 Nov 01:32
· 10 commits to main since this release

[0.1.0-beta.3] - 2026-11-14

Added

  • Remote models, by your choice: set up remote speech and language services on the Models page (OpenAI, Doubao, Alibaba Cloud, Gemini Live, StepFun, Z.ai, Xiaomi MiMo; Anthropic, Gemini, DeepSeek, OpenRouter and others; Ollama, LM Studio and oMLX on your Mac) and choose one on a feature's page. Only that feature sends its audio or text, and only to that service; "Automatic" is always a local model. Keys stay in the Keychain and can be shared with the other ActNow apps. OpenAI-style services can use the Responses API, and Rewrite can search the web.
  • Apple Speech and Apple Intelligence as local models, with nothing to download (on this Mac only; Apple Intelligence needs macOS 26).
  • One Models page for speech and language models, local and remote, with filters; models can be hidden. More local models, marked experimental: Whisper large-v3-turbo, large-v3 and small (no live text), Qwen3-ASR 1.7B 6-bit, Gemma 4, Qwen3.5 9B, Ministral 3, LFM2 and others.
  • Dictionary: categories, how often each term was used, terms limited to an App enhancement group, and the corrections it made shown in the history.
  • History: play an entry's recording; ask follow-up questions about an entry, answered by Rewrite's model and kept with it.
  • Shortcuts: mouse buttons, and presets (fn included); Space starts a follow-up while Rewrite's answer card shows.
  • Output and recording: keep the result on the clipboard (on by default), Escape cancels a recording (on), show live text (on), and a voice end command (off): say "over", 完毕 or 以上 and pause to stop.
  • Show icon in Dock (off by default); closing the window keeps VoxtMini running with either icon.
  • Notifications when a download finishes or fails while the window is not in front; a Microphone submenu in the menu bar; reorder microphones by dragging them.
  • Announcements from actnow.dev, shown once as notifications.
  • One editable prompt per feature, in place of "Your rules": Transcribe's clean-up, Translate and Rewrite each show their prompt, to change as you like, with variables to insert and Reset to default. The input, the dictionary, App enhancement's scene and the output format stay VoxtMini's (see "Show the full prompt"). Rules you had written are kept, added to the end of the prompt.
  • Settings per model, from a model's ⋯ menu on the Models page: only the knobs that model has (Thinking, temperature, top-p, top-k, min-p, repetition penalty, maximum output; a main-language hint for speech models); empty keeps VoxtMini's tuned values. Compact model cards.
  • A self-check of the network, disk and Keychain under Settings → About, with a proxy connection test; a crash report is offered at the next start.

Changed

  • Languages: English, Chinese and Japanese. Chinese results are always in Simplified characters (any Traditional characters a model writes are converted); Japanese text is never altered. Models that do not know the three were removed.
  • Shorter text everywhere: explanations moved behind ⓘ; a smaller Settings window.
  • The Speech Recognition permission is asked for only when a feature uses Apple Speech.
  • VoxtMini now lives in the former Voxt repository on GitHub (hehehai/VoxtMini); Voxt's history ends at the tag voxt-final.
  • VoxtMini has its own page on actnow.dev (actnow.dev/voxtmini), linked from Settings → About with its download; updates come from its own feed there.
  • VoxtMini has its own identity on the Mac (bundle id dev.hehehai.voxtmini, folders ~/Library/…/VoxtMini, links voxtmini://), so it no longer shares one with Voxt 2. The first start moves your history, recordings, dictionary, settings and model claims over by itself. macOS asks once more for the microphone and Accessibility: grant them again.

Fixed

  • A dictation into a web page's text box (seen in Arc) sometimes did not go in, and the answer card said "The text did not show up where you type" although the cursor was in the box. VoxtMini now sends its keys as Voxt does, after every app's keyboard hooks, so a busy Mac cannot hold them up; and when the box still shows nothing half a second after ⌘V, with nothing else changed, it sends ⌘V once more before showing the card.
  • Clicks anywhere on the Mac could stall for seconds when VoxtMini's helper was swapped out: the helper now only listens to the mouse.
  • Permissions no longer flicker when VoxtMini comes to the front.
  • Live text no longer stops after the speech model was slow to start.

中文

新增

  • 远程模型,由你选择:在模型页设置远程语音识别和语言模型服务(OpenAI、豆包、阿里云百炼、Gemini Live、阶跃星辰、智谱、小米 MiMo;Anthropic、Gemini、DeepSeek、OpenRouter 等;本机的 Ollama、LM Studio、oMLX),再在功能页里选用。只有这个功能把它的音频或文字发出去,而且只发给这个服务;“自动”总是本地模型。密钥保存在钥匙串里,可以与其他 ActNow app 共用。OpenAI 类服务可以用 Responses API,转写可以联网搜索。
  • Apple 语音识别和 Apple 智能作为本地模型,无需下载(只在本机运行;Apple 智能需要 macOS 26)。
  • 统一的模型页:语音识别和语言模型、本地和远程都在一页,可以筛选,也可以隐藏模型。更多本地模型(标为实验性):Whisper large-v3-turbo、large-v3 和 small(没有实时文字)、Qwen3-ASR 1.7B 6-bit、Gemma 4、Qwen3.5 9B、Ministral 3、LFM2 等。
  • 词典:分类、每个词条的使用次数、只在某个应用增强分组里生效的词条,历史里标出词典做的更正。
  • 历史:播放一条记录的录音;针对一条记录继续提问,由转写的模型回答并随记录保存。
  • 快捷键:鼠标按键和预设(包括 fn);转写的答案卡片显示时,按空格继续追问。
  • 输出和录音:结果留在剪贴板(默认开)、Esc 取消录音(开)、显示实时文字(开),以及语音结束指令(关):说“over”“完毕”或“以上”再停顿一下就停止录音。
  • 在程序坞中显示图标(默认关);关掉窗口后,无论显示哪个图标,VoxtMini 都继续运行。
  • 窗口不在前台时,下载完成或失败会发系统通知;菜单栏里有麦克风子菜单;拖动即可调整麦克风顺序。
  • 来自 actnow.dev 的公告,以通知的形式各显示一次。
  • 每个功能一份可编辑的提示词,取代“你的规则”:转录的整理、翻译和转写各自显示提示词,可以随意修改,可以插入变量,也可以恢复默认。输入说明、词典、应用增强的场景和输出格式仍由 VoxtMini 负责(见“查看完整提示词”)。你写过的规则会保留,接在提示词末尾。
  • 按模型设置,在模型页模型的 ⋯ 菜单里:只显示这个模型支持的参数(思考、温度、Top-p、Top-k、Min-p、重复惩罚、最大输出;语音识别模型可提示主要语言);留空则使用 VoxtMini 调好的值。模型卡片更紧凑。
  • 设置 → 诊断与关于 里可以自检网络、磁盘和钥匙串,并测试代理连接;崩溃后下次启动会提示查看报告。

变更

  • 语言:英文、中文和日文。中文结果一律是简体字(模型写出的繁体字会被转换);日文不做任何改动。去掉了不支持这三种语言的模型。
  • 各处文字更短:说明放进 ⓘ;设置窗口更小。
  • 只有某个功能用 Apple 语音识别时才请求“语音识别”权限。
  • VoxtMini 现在放在 GitHub 上原来 Voxt 的仓库里(hehehai/VoxtMini);Voxt 的历史到标签 voxt-final 为止。
  • VoxtMini 在 actnow.dev 上有了自己的页面(actnow.dev/voxtmini),设置 → 诊断与关于 里链接它和下载页;更新也来自那里它自己的更新源。
  • VoxtMini 在 Mac 上有了自己的身份(bundle id dev.hehehai.voxtmini,文件夹 ~/Library/…/VoxtMini,链接 voxtmini://),不再和 Voxt 2 共用。第一次启动时自动搬过历史、录音、词典、设置和模型声明。macOS 会再请求一次麦克风和辅助功能权限:请重新授权。

修复

  • 往网页输入框(在 Arc 里遇到)里听写时,有时文字没进去,答案卡片说“文字没有出现在输入位置”,其实光标一直在框里。现在 VoxtMini 像 Voxt 一样,把按键发在所有 App 的键盘钩子之后,Mac 忙时也不会被拖住;⌘V 后半秒输入框还没变化、其他也都没变时,会再发一次 ⌘V,仍不行才显示卡片。
  • VoxtMini 的辅助进程被换出内存时,整台 Mac 的点击可能卡住几秒:现在辅助进程对鼠标只监听。
  • VoxtMini 回到前台时权限状态不再闪烁。
  • 语音识别模型启动慢时,实时文字不会再中断。