@DanKornas: Building a person-specific talking head shouldn’t require piecing together preprocessing, training, and inference from …

X AI KOLs Timeline Tools
talking-head digital-human audio-driven python open-source onnx streaming

Summary

Ultralight Digital Human is an open-source Python project that lets you train a person-specific, audio-driven talking head from a 3–5 minute video, with support for HuBERT/WeNet audio features, ONNX export, and streaming inference.

Building a person-specific talking head shouldn’t require piecing together preprocessing, training, and inference from separate projects. Ultralight Digital Human is a Python project for training a person-specific, audio-driven digital human from a short source video. It helps you turn a 3–5 minute, full-face video into a trained talking-head workflow by extracting face landmarks and audio features before UNet training and inference. Key features: • Person-specific workflow – trains one dedicated model for the identity in the source video. • Two audio paths – supports HuBERT at 25 fps and WeNet at 20 fps. • End-to-end guidance – covers preprocessing, training, inference, and the final audio/video merge. • ONNX export checks – exports a checkpoint and compares ONNX Runtime output with PyTorch. • Streaming demo – processes 10 ms PCM frames with WeNet and ONNX Runtime. Free public GitHub repo. Link in the reply
Original Article
View Cached Full Text

Cached at: 07/25/26, 10:13 PM

Building a person-specific talking head shouldn’t require piecing together preprocessing, training, and inference from separate projects.

Ultralight Digital Human is a Python project for training a person-specific, audio-driven digital human from a short source video.

It helps you turn a 3–5 minute, full-face video into a trained talking-head workflow by extracting face landmarks and audio features before UNet training and inference.

Key features: • Person-specific workflow – trains one dedicated model for the identity in the source video. • Two audio paths – supports HuBERT at 25 fps and WeNet at 20 fps. • End-to-end guidance – covers preprocessing, training, inference, and the final audio/video merge. • ONNX export checks – exports a checkpoint and compares ONNX Runtime output with PyTorch. • Streaming demo – processes 10 ms PCM frames with WeNet and ONNX Runtime.

Free public GitHub repo.

Link in the reply

GitHub: https://github.com/anliyuan/Ultralight-Digital-Human…

If you’re into AI, ML, agents, and building real systems, join my newsletter (it’s free): http://dankornas.substack.com

If an AI coding assistant cannot see a project’s latest docs, it can guess the wrong API.

GitMCP is a remote Model Context Protocol server for builders who want AI assistants to work with public GitHub repositories and GitHub Pages documentation.

It helps you ground coding answers in current project material by exposing documentation fetching, documentation search, linked-page retrieval, and GitHub code search as MCP tools.

Key features: • Repository-specific endpoints – keep an assistant focused on one selected project. • Generic endpoint – switch between public repositories without configuring a separate server for each. • Documentation priority – checks llms.txt first, then AI-optimized docs, then the README or root page. • Targeted search – retrieves relevant documentation and code instead of loading an entire documentation set. • Hosted or self-hosted – use the cloud service with no signup, or deploy the open-source project yourself.

It’s open-source (Apache License 2.0), though package.json currently declares ISC.

Link in the reply

Similar Articles