功能Features
每句话在说话停顿处断开，先出字，再出译文：以中文为主的句子译成英文，其余译成中文。结束录音，字幕直接成为文字稿。Each sentence is cut at a natural pause, shown right away, then translated: mostly-Chinese into English, everything else into Chinese. Stop recording and the captions become the transcript.
SenseVoice 先出字、判断语种；英文交给 Parakeet，中文由 FireRedASR2 精修。会后再把整段录音重新识别一遍。SenseVoice drafts and detects the language; English goes to Parakeet, Chinese is refined by FireRedASR2. After the meeting the whole recording is recognized again.
概要、要点、决定、待办（谁、什么时候），导出 Markdown 或纯文本。Summary, key points, decisions and to-dos with owners and dates. Export to Markdown or plain text.
智能校对结合前后文改错字；每一处改动都给出候选写法（包括原文，一键改回），也可以自己填，或者先听一下原声。Proofreading fixes misheard words from context. Every change comes with alternatives, including the original, or type your own, or replay the audio first.
这一步主要是做数据校验，确认没有漏掉的记录。
导入自己的术语表，开会前导入议程、PPT、名单提取本场术语；预装 22 万条术语，覆盖 27 个类别。Import your glossaries, or the agenda, slides and attendee list before a meeting. 220,000 built-in terms across 27 categories.
录音来源可选麦克风、电脑正在播放的声音，或两者一起；也能导入 MP3、M4A、MP4、MKV 等 13 种音视频格式。Record the microphone, whatever your PC is playing, or both. Or import 13 audio and video formats including MP3, M4A, MP4 and MKV.
边录边写入硬盘，字幕每两秒存一次；退出、崩溃或断电后，下次打开自动存成一条记录。Audio is written to disk as you record and captions are saved every two seconds. After a crash or power cut, the next launch saves everything automatically.
响岩 · 语音识别Phonolite · Speech recognition
第一遍用 SenseVoice 立刻出字、判断语种；英文交给 Parakeet，中文由 FireRedASR2 再精修一遍。说话停顿处断句，句首句尾留出余量，不吞字。The first pass with SenseVoice writes text instantly and detects the language. English goes to Parakeet, and Chinese gets a second pass from FireRedASR2. Sentences are cut at natural pauses with padding at both ends, so no words get clipped.
响岩 · 降噪Phonolite · Noise reduction
逐段估计信噪比，只在嘈杂的段落启用 DPDFNet 降噪；小声的段落先补音量再识别。清晰的录音保持原样，免得越处理越差。Signal-to-noise is measured segment by segment. DPDFNet noise reduction only kicks in where it's noisy, and quiet passages are boosted before recognition. Clean audio is left alone, so processing never makes it worse.
本机推理On-device inference
一个本机引擎开两条道：实时道逐句翻译，后台道校对与写纪要，各自排队、互不抢先。有显卡自动用显卡（独显、核显都行），整个模型放进显存；没有就用处理器，并给系统留出余量。One on-device engine runs two lanes: a live lane for sentence-by-sentence translation and a background lane for proofreading and notes, each with its own queue. With a GPU, dedicated or integrated, the whole model goes into video memory; without one it runs on the CPU and still leaves headroom for your system.
界面The app
左边录音库，中间带时间戳的文字稿，右边纪要。点时间戳从那里回放，点文字就能改，Ctrl+F 搜遍文字稿和纪要。Library on the left, a timestamped transcript in the middle, notes on the right. Click a timestamp to replay, click a line to edit, and Ctrl+F to search everything.
隐私Privacy
装好就能用，不注册、不登录。Install and go. No sign-up, no sign-in.
录音、文字稿和纪要都存在你的电脑上，导出到哪里由你决定。Recordings, transcripts and notes stay on your PC. You choose where exports go.
我们看不到你录了什么、说了什么。We can't see what you record or what was said.
安装包还在制作中，做好后会放在这里。The installer is still in the works and will be posted here.
Windows 10 / 11 · 安装后约 5.3 GB · 内存 8 GB 起Windows 10 / 11 · about 5.3 GB installed · 8 GB memory or more
手机上的响岩Phonolite on your phone
还没有代码签名证书时，Windows 可能提示“已保护你的电脑”：点“更多信息 → 仍要运行”。下载后可以核对 SHA-256：certutil -hashfile 文件名 SHA256。Until the installer is code-signed, Windows may show “Windows protected your PC”: choose “More info → Run anyway”. You can verify the SHA-256 with certutil -hashfile <file> SHA256.
感谢这些项目的作者。完整的许可证随安装包附带，也在 App 的“关于 → 开源许可”里。Thanks to the authors of these projects. Full license texts ship with the app and are listed under About → Open-source licenses.