AV2Text

Local transcription for WindowsWindows 本地转写工具

Turn audio and video into text you can use. 把音视频，变成能用的文字。

Drop in a file, paste a video link, or talk into the mic. AV2Text transcribes on your own machine — the audio never leaves the computer. 拖入文件、粘贴视频链接，或者直接对着麦克风说。转写在本机完成 —— 音频不出这台电脑。

Download for Windows下载 Windows 版 Read the guide看使用说明

0.1.3 · 1217 MB · Windows 10 / 11 (64-bit)184 MB · Windows 10 / 11（64 位）

Runs fully offline完全离线运行 No account不需要账号 Free for personal use个人使用免费

One recording, three views一条录音，三种视图

Original · Corrected · Translated原文 · 校正 · 译文

They are not three features — they are three faces of the same transcript. Same timecodes, same blocks. Fix a line in the original and the corrected pane follows; the translation re-runs from the corrected text. 三页不是三个功能，而是同一份内容的三张脸。时间码对齐、分块一致：在原文里改一处，校正页跟着改，译文按改后的原文重新生成。

Original原文 ASR
[00:12]We'll hook up the on-site tooling data today.我们今天把现场的工据接一下。
[00:24]This data stays local — don't push it to the cloud.这个数据得先落到本地，别往云上传。
Raw model output模型原始输出 editable可直接改
Corrected校正 restore
[00:12]We'll hook up the on-site tooling data today.我们今天把现场的工具数据接一下。
[00:24]This data stays local — don't push it to the cloud.这个数据得先落到本地，别往云上传。
Fixes homophones and names修同音字与专名 in place就地改字
Translated译文 optional可选
[00:12]We'll hook up the on-site tooling data today.
[00:24]This data stays local — don't push it to the cloud.
Runs from the corrected text按校正后的原文生成 text only只有文字出网

What it does能做什么

Four ways in, one way out四种输入，一个出口

Sound in声音进来 Text out — with timecodes, ready to export. 文字出去 —— 带时间码，可以直接导出。

Drop in a file文件转写

mp3, mp4, wav and friends. Drop it on the window and it starts.mp3、mp4、wav 等常见格式，拖进窗口就开始。

Paste a link链接转写

Bilibili, YouTube, Douyin, Xiaohongshu and 46 sites in total. The audio track is pulled first, then transcribed.B站、YouTube、抖音、小红书等 46 个站点，自动取音轨再转写。

Live from the mic麦克风实时

Meetings and lectures land on screen as they are spoken, under a second behind.会议、讲座边听边出字，上屏延迟不到 1 秒。

Who said what说话人分离

Speakers are separated into blocks, and you can rename them in one place.把人分成块，改名一次，三页同步。

Text out of images图片与 PDF 取字

Screenshots, scans and PDFs are read locally too, layout kept roughly in place.截图、扫描件和 PDF 也走本机取字，版面尽量按原样保留。

Seven ways out七种导出

TXT, MD, SRT, DOCX, PDF, JSON and LRC — subtitle timecodes stay aligned.TXT、MD、SRT、DOCX、PDF、JSON、LRC —— 字幕时间码对齐。

TXTMDSRTDOCXPDFJSONLRC

Privacy隐私

Your audio never leaves this computer音频不出这台电脑

Transcription is done by a speech model running on your own CPU. There is no account, no upload, and no server holding your recordings. 转写由本机 CPU 上的语音模型完成。没有账号，没有上传，也没有哪台服务器存着你的录音。

  • Audio and video are processed locally — never uploaded.音频与视频在本机处理 —— 不上传。
  • Models run on your CPU — they ship inside the installer, so nothing is downloaded on first launch.模型在本机 CPU 上跑 —— 模型都打包在安装包里，首次启动不下载任何东西。
  • It only goes online when you ask it to — pasting a link fetches that page's audio track.只在你主动用的时候才联网 —— 粘贴链接是去那个站点取音轨。
  • Cloud translation is optional — when used, only the text goes out, with a random device id and no account.云端翻译是可选功能 —— 用了才把文字发出去，只带一个随机设备号，没有账号。
  • Diagnostics are redacted by default — links, keys and device ids are masked before you share a file.诊断包默认脱敏 —— 分享前链接、密钥、设备号一律打码。
  • Details in the privacy policy.细节见隐私政策。

Download下载

Get AV2Text for Windows下载 AV2Text

Version 0.1.3 · installer 1217 MB · Windows 10 / 11 (64-bit). 版本 0.1.3 · 安装包 1217 MB · Windows 10 / 11（64 位）。

Download the installer (.exe)下载安装包（.exe） Release notes更新日志

The installer is large on purpose: all speech models ship inside it, so there is nothing to download on first launch — install and it works offline. If the download stalls, the direct mirror link also works. 安装包大是有意的：全部语音模型都打包在里面，首次启动无需下载任何东西，装完即可离线用。若下载中断，也可以直接用镜像直链。

  • System系统Windows 10 / 11 · 64-bit
  • Memory内存8 GB (16 GB recommended)（推荐 16 GB）
  • Disk磁盘5 GB free可用
  • GPU显卡Not needed不需要
  • Network网络Only to fetch models仅首次下载模型

FAQ

Questions people ask first最常被问到的几件事

Do I need a graphics card?需要显卡吗？

No. It runs on the CPU in real time. On this machine a GPU was measured to be slower, so it is not used.不需要，CPU 就能实时跑。实测在这类机器上上显卡反而更慢，所以没有用它。

Which languages are supported?支持哪些语言？

Chinese and English with the current speech model. Multilingual recognition was dropped on purpose rather than shipped half-working.当前的语音模型支持中英。多语言识别是有意砍掉的 —— 与其半能用，不如不做。

Is my recording uploaded anywhere?录音会被上传吗？

No. Everything is transcribed on your machine. The only optional network feature is cloud translation, and that sends text, not audio.不会。全部在本机转写。唯一可选的联网功能是云端翻译，发出去的是文字，不是音频。

Is it free?免费吗？

Free for personal use. See the terms.个人使用免费，见服务条款。

Do I have to cut long recordings into pieces?长录音要自己切吗？

No — audio is processed in blocks. Keep about 4 GB of memory free and close heavy apps while a long file runs.不用，音频是按块处理的。转长文件时留出大约 4 GB 可用内存，把吃内存的程序关掉。

What comes out at the end?最后能导出什么？

TXT or MD for reading, SRT and LRC for subtitles, DOCX and PDF for handing over, JSON when another program has to read it.自己读用 TXT 或 MD，做字幕用 SRT 和 LRC，交出去用 DOCX 和 PDF，要给别的程序读就用 JSON。