VoxAI iPhone 版 · 隐私政策
最后更新:2026 年 11 月 13 日 · 适用:iPhone 版 1.0.1
一句话版本:VoxAI 没有服务器,没有账号,没有追踪。录音、转录和声纹都存在你的 iPhone 上,作者一份都收不到。只有一种情况会让内容离开手机:你自己填了模型服务商的 API key、并在弹窗里同意之后,推卡、会后整理(包括整理前的校对,以及你打开的「录完自动整理」)和追问时,相关的转录文字会用你的 key 直接发给你选的服务商。录音本身从不发送。
VoxAI 是什么
VoxAI 是 iPhone 上的面对面 AI 副驾。面试、谈判、会议时把手机放在桌上,它实时把语音转成文字、区分谁在说话,在屏幕上递一句提示,会后整理摘要和行动项。
VoxAI 不做什么
- 没有自己的服务器,你的音频、转录和声纹从不上传给作者。
- 不收集任何个人信息,不需要注册或登录。
- 不追踪你,不集成任何分析、广告或崩溃上报 SDK,不发送使用统计或遥测。
- 不读取你的通讯录、照片、位置、日历或其他 App 的数据。
你的数据存在哪
全部在这台 iPhone 上,在 VoxAI 的 App 沙盒里。卸载 VoxAI,下面这些会随之删除。
录音
每场录制的音频存成 16 kHz 单声道 WAV 文件。VoxAI 不会自动删除录音,它会一直留在手机上,直到你在 App 里删除那场会议,或卸载 App。
转录、推卡和摘要
逐句转录(时间、说话人标签、原文)、会中出现过的提示卡、会后摘要和按模板整理出的文档(会议纪要、面试复盘等)、校对时改过的字(可以撤销)、你改过的说话人名字,存在 App 里的会议文件中。
语音转文字
转录使用 iOS 自带的本地识别引擎(SpeechAnalyzer),在手机上完成,音频不出手机。识别需要的语言资源由 iOS 从 Apple 下载,这是系统行为,VoxAI 不经手。
区分说话人和「我的声纹」
区分说话人用的是手机本地的 CoreML 模型(FluidAudio),在手机上完成。
「我的声纹」是你自愿录的一段 10 秒语音算出的一组数字特征,用来认出哪句是你说的。存在手机上的 App 文件里,不是录音本身,也不会上传。可以在设置里重录,卸载即删除。会后重新区分说话人时算出的声音特征只在处理过程中使用,不另外保存,也不建立跨会议的声纹库。
API key
如果你填了模型服务商的 API key,它存在 iOS 钥匙串里,只用于从这台手机直接调用该服务商,不会发给作者。
偏好设置和运行日志
识别语言、推卡节奏、你的词表和自定义整理模板等偏好存在本机的 UserDefaults。App 会在本机写一份运行日志,只记录状态、时间和出错信息,不记录转录原文,不上传。
锁屏和灵动岛
录制时,锁屏的实时活动和灵动岛会显示录制状态和最新一张提示卡的标题。它显示在你自己的屏幕上,旁边的人可能看得到。
录到别人的声音
请先征得对方同意。VoxAI 会录下麦克风收到的所有声音,包括在场其他人说的话。许多国家和地区的法律要求录音前取得所有参与者的同意。是否告知、是否取得同意,责任在使用者。
VoxAI 什么时候联网
| 什么时候 | 连向谁 | 发出去什么 | 默认 |
|---|---|---|---|
| 第一次用到说话人功能(区分说话人、录「我的声纹」) | HuggingFace | 只下载说话人模型,一次性,不发送任何音频或文字。对方能看到的只有你的 IP 地址和下载请求。 | 用到时自动下载 |
| 语音转文字 | Apple | iOS 下载识别所需的语言资源。音频在本地识别,不发送。 | 系统行为 |
| 推卡、会后整理、追问 (仅当你填了 API key 并同意) | 你选的服务商:Anthropic(api.anthropic.com),或你自己填的 OpenAI 兼容服务地址 | 推卡时发最近一段转录文字、你填的会前材料和场景说明;会后整理和追问时发整场转录文字(最多约 6 万字),自定义模板会带上你写的要求;打开「录完自动整理」(默认关闭)后,每场录完会自动整理一次。整理前校对人名和术语时,分段发整场转录文字,并附上本场的说话人名字、你在其他会议里给说话人起过的名字(最多 60 个)、你的词表和本场已有的摘要要点。不发送录音。用你自己的 key 和额度直接发出,不经过作者,内容怎么处理适用该服务商的隐私政策。 | 关闭:不填 key 就不会发生;「录完自动整理」默认关闭,「整理前校对人名和术语」可以在设置 › 会后整理里关掉。演示模式完全不联网 |
除了上表,VoxAI 不发起其他网络连接。
演示模式
不想接任何服务商时,可以在设置里把推卡来源切到「演示」。演示模式下的提示卡和会后摘要是本地生成的示例,不联网、不调用任何模型。录音和转录照常是真的,仍然只在手机上。
删除你的数据
- 删除单场会议:在会议列表左滑删除,录音和转录一起删掉。
- 删除 API key:设置 › 云端模型 › 清除。
- 删除全部:卸载 VoxAI。作者那边没有任何副本需要另外删除。
儿童
VoxAI 不面向 13 岁以下儿童,也不向作者收集任何人的数据。
政策变更
如果以后的版本改变了这里描述的任何数据行为,比如新增联网用途或引入新的第三方服务,本政策会在该版本发布时同步更新,页首日期随之改变。
联系
有疑问请发邮件到 [email protected]。
VoxAI for iPhone · Privacy Policy
Last updated: November 13, 2026 · Applies to: VoxAI for iPhone 1.0.1
TL;DR: VoxAI has no servers, no accounts and no tracking. Your recordings, transcripts and voice profile stay on your iPhone; the developer never receives any of it. Content leaves your phone in exactly one case: after you enter your own API key for an AI provider and agree to the prompt, the relevant transcript text is sent directly to that provider, with your key, to generate live cards, meeting notes (including the proofreading pass before it, and the automatic one after each recording if you turn that on) and answers. Audio is never sent.
What VoxAI is
VoxAI is an in-person AI copilot for iPhone. Put the phone on the table during an interview, negotiation or meeting: it transcribes speech in real time, tells speakers apart, shows short hints on screen, and summarizes the conversation afterwards.
What VoxAI does not do
- No servers of its own. Your audio, transcripts and voice profile are never uploaded to the developer.
- No personal data collection. No sign-up, no login.
- No tracking. No analytics, advertising or crash-reporting SDKs; no usage statistics or telemetry.
- No access to your contacts, photos, location, calendar or other apps' data.
Where your data lives
Everything stays on this iPhone, inside VoxAI's app sandbox. Deleting the app deletes all of it.
Recordings
Each session's audio is saved as a 16 kHz mono WAV file. VoxAI never deletes recordings automatically; they stay until you delete the session in the app or delete the app.
Transcripts, cards and summaries
Sentence-by-sentence transcripts (time, speaker label, text), the hint cards shown during the session, summaries and template-based notes (meeting minutes, interview review and so on), the corrections made by proofreading (which you can undo), and any speaker names you edit are stored in the app's meetings file.
Speech recognition
Transcription uses iOS's built-in on-device engine (SpeechAnalyzer). It runs on the phone and audio does not leave the device. iOS downloads the language assets it needs from Apple; that is system behavior and VoxAI is not involved.
Speaker separation and "My voice"
Telling speakers apart uses local Core ML models (FluidAudio) and runs on the phone.
"My voice" is a set of numbers computed from a 10-second sample you choose to record, used to recognize which sentences are yours. It is stored in the app's files on the phone, is not the recording itself, and is never uploaded. You can re-record it in Settings; deleting the app removes it. Voice features computed while re-labeling speakers after a session are used only during processing and are not kept; VoxAI does not build a cross-session voice library.
API key
If you enter an API key, it is stored in the iOS Keychain and used only to call that provider directly from this phone. It is never sent to the developer.
Preferences and log
Preferences such as recognition language, card pace, your vocabulary list and custom note templates are stored in UserDefaults on the device. The app keeps a local log of states, timings and errors that does not contain transcript text and is never uploaded.
Lock Screen and Dynamic Island
While recording, the Live Activity on the Lock Screen and the Dynamic Island show the recording status and the title of the latest hint card. They appear on your own screen, where people nearby may see them.
Recording other people
Ask for consent first. VoxAI records everything the microphone picks up, including other people in the room. Many jurisdictions require consent from all participants before recording. Informing others and obtaining consent is the user's responsibility.
When VoxAI goes online
| When | To whom | What is sent | Default |
|---|---|---|---|
| First use of a speaker feature (separating speakers, recording "My voice") | HuggingFace | One-time download only of the speaker models. No audio or text is sent; the server sees only your IP address and the download request. | Downloaded when first needed |
| Speech recognition | Apple | iOS downloads language assets. Audio is recognized locally and not sent. | System behavior |
| Live cards, summaries, follow-up questions (only after you enter an API key and agree) | The provider you choose: Anthropic (api.anthropic.com), or an OpenAI-compatible endpoint you enter | For live cards: the most recent transcript text, the prep notes you entered, and a scenario description. For notes and questions: the whole transcript text (up to about 60,000 characters), plus your instructions if you use a custom template; if you turn on auto-organize (off by default), each meeting is organized once automatically after recording. For the proofreading pass before organizing: the whole transcript text in chunks, together with this meeting's speaker names, names you have given speakers in other meetings (up to 60), your vocabulary list and this meeting's existing summary points. Audio is never sent. Requests go directly with your own key and quota, never through the developer, and are handled under that provider's privacy policy. | Off: nothing is sent without a key. Auto-organize is off by default; proofreading can be turned off in Settings › 会后整理 (After the meeting). Demo mode never goes online |
VoxAI makes no other network connections.
Demo mode
If you don't want to use any provider, switch the card source to "Demo" in Settings. In demo mode, cards and summaries are local samples: nothing goes online and no model is called. Recording and transcription are real and stay on the phone.
Deleting your data
- One session: swipe left in the list to delete it; its audio and transcript go together.
- API key: Settings › Cloud model › Clear.
- Everything: delete the app. The developer holds no copies to delete.
Children
VoxAI is not directed to children under 13 and collects no one's data.
Changes
If a future version changes any data behavior described here, such as a new network use or a new third-party service, this policy will be updated when that version ships and the date above will change.
Contact
Questions: [email protected].