RepoPicks Radar

Updated every day around 23:30 UTC

Guides

LiveTalking

lipku/LiveTalkingReviewed by AI

Checked by RepoPicks Radar on

What is LiveTalking?

LiveTalking shows a digital person who talks in real time: you send text or audio, it moves the mouth in a video of a face you provide, and sends the result to a browser over WebRTC, to a live platform over RTMP, or out as a virtual webcam. The README points it at unmanned shopping streams, avatar customer service, courses and batch short videos; it can copy a voice, be cut off mid-sentence and play a chosen video while quiet. You need Linux or Windows with an NVIDIA card, Python, and model files fetched from file-sharing folders.

How easy is it to install LiveTalking?

D · terminal, several tools or a specific graphics card · Windows

  • Windows D
  • Linux D

Linux is the tested system and still a D: create a conda environment with Python 3.12, install a PyTorch build that matches your CUDA version, pip install the requirements, download wav2lip.pth and an avatar folder from a Quark or Google Drive folder and move them into the right directories by hand, then run app.py and open the ports. Windows takes the same route; the README also mentions an all-in-one Windows package on a file-sharing site, with no instructions for it. Both need an NVIDIA card: the frame-rate table starts at an RTX 3060.

Coming soon: install in 1 click with the RepoPicks app.

LiveTalking on GitHub

Stars
10,356on Nov 10, 2026
Downloads
—all releases
Growth
+84stars in 6 days

Stars over time

10,25010,27510,30010,32510,35010,375 Nov 4Nov 6Nov 8Nov 10 10,25010,27510,30010,32510,35010,375 Nov 4Nov 7Nov 10

Facts

Runs on
Windows, Linux
You need to install
Python, NVIDIA graphics card, FFmpeg
To build from source
Python, Docker, Graphics card, FFmpeg
Disk space
510 MB–1.3 GB · estimate
Accounts or keys
Qwen through DashScope, Doubao or another model service, only for the chat mode (the echo mode just repeats your text), Tencent Cloud or Azure Speech, only if you want their voices instead of the free EdgeTTS
Cost
Freemium · This version is free under Apache-2.0. The README has a table comparing it with a paid "商用版" (commercial version) that adds the wav2lipls model and a set of extra abilities, and it also offers a hosted trial at livetalking.top. Speech and model services you plug in (Tencent Cloud, Azure, Qwen) bill separately; EdgeTTS costs nothing.
License
Apache-2.0 (Permissive)
Can replace
HeyGen, Synthesia, D-ID
Last change on GitHub
Oct 20, 2026

LiveTalking risk signals (6)

  • AttentionDownload links to file-sharing sites

    README links to https://drive.google.com/drive/folders/1FOC_MD6wdogyyX_7V1d4NDIO7P9NlSAJ?usp=sharing

    automatic check

  • Attentionmodels and ready packages come from file sharing links

    The model files are offered on Quark Cloud ("夸克云盘 https://pan.quark.cn/s/83a750323ef0"), and so are the "Windows整合包 https://pan.quark.cn/s/a040bf5cb065" and the "桌面客户端 https://pan.quark.cn/s/d7192d8ac19b". Files on a personal netdisk can be replaced at any time and carry no checksum in the repository.

    AI review

  • Attentioncopies a real persons face and voice

    The Avatar page "上传视频自动生成数字人形象" builds the digital person from a video you upload, feature 2 is "支持声音克隆" (voice cloning), and the suggested use is unmanned selling streams. Together that makes a video of a real person saying words they never said, which needs their permission and is restricted in several countries.

    AI review

  • Attentioninstructions ask you to open every udp port

    The start-up note says "服务端需开放端口 TCP:8010, UDP:1-65536" — open the whole UDP range on the server. The pages it then serves include "/admin.html 管理后台" for "实时监控会话状态与全局配置", and no password or token for it appears anywhere in the README.

    AI review

  • Infovideos you publish must carry the projects watermark

    The 声明 (declaration) section: "基于本项目开发并发布在B站、视频号、抖音等平台上的视频需带上 LiveTalking 水印和标识" — videos made with it and posted on Bilibili, WeChat Channels or Douyin must carry the LiveTalking watermark. That condition is not part of the Apache-2.0 licence the repository carries.

    AI review

  • Infothe dockerfile in the repository cannot build as is

    The Dockerfile ends with "Copy ../nerfstream /nerfstream", a path outside the build folder, and runs "conda activate" in a RUN step, which does not carry over. The Docker section of the README does not build it either: it points to ready images hosted by AutoDL and UCloud.

    AI review

Evidence is shown in its original language.

What Radar could not check about LiveTalking

  • Whether it runs on macOS, which the badge at the top of the README claims while every instruction and figure assumes an NVIDIA card
  • What the files on the netdisk and Google Drive folders actually contain
  • The frame rate on your own card, and whether it stays above the 25 the README calls real time
  • Whether the admin page and the API ask for any password once the ports are open
  • How many people it really serves at once, and how the mouth and the interruption look in practice

Similar projects

The RepoPicks app

Coming soon: install in 1 click with the RepoPicks app.

Before you download

See the release page on GitHub