VibeVoice (Speech Generation/Voice Cloning) on Framework Desktop with Strix Halo (AMD AI Ryzen MAX+)

In this video, I show how to generate natural-sounding speech locally on the Framework Desktop with AMD Ryzen AI Max “Strix Halo” — including cloning a voice from a short sample and creating multi-speaker conversations. The intro you hear at the start was entirely generated by VibeVoice, cloned from my own voice.

VibeVoice is Microsoft’s open-weight model for long-form, multi-speaker speech (released late August 2025). I’ll walk you through setup on Strix Halo using a Fedora toolbox and the Gradio UI, then demo single-speaker and multi-speaker clips, plus zero-shot voice cloning. I’ll also cover stability fixes for ROCm crashes.

Timestamps:
00:00 — AI-Generated Intro (VibeVoice)
01:47 — Setup on Strix Halo (Toolbox + Gradio)
03:28 — First Demo: Single-Speaker
05:18 — Multi-Speaker Conversations
05:42 — Clone Your Own Voice (Zero-Shot)
06:23 — Stability Fixes (librosa / numba / LLVM / ROCm)
08:26 — Generating a Full Podcast
09:33 — AI-Generated Podcast: How VibeVoice Works

— — —

Links & Resources:

GitHub repo (toolboxes, scripts, stability fixes):
https://github.com/kyuz0/amd-strix-halo-voice-toolbox

Framework Desktop (Strix Halo):
https://frame.work/

Strix Halo Homelab guide + Discord (by deseven):
https://strixhalo-homelab.d7.wtf/

VibeVoice (project):
https://github.com/microsoft/VibeVoice
https://microsoft.github.io/VibeVoice/

VibeVoice models (Hugging Face):
https://huggingface.co/microsoft/VibeVoice-1.5B
(Community mirror example for large weights):
https://huggingface.co/aoi-ot/VibeVoice-Large

Gradio (UI framework):
https://github.com/gradio-app/gradio

Librosa (audio features):
https://github.com/librosa/librosa

Numba (JIT; disabled in this toolbox fix):
https://github.com/numba/numba

LLVM (compiler backend):
https://llvm.org/

Видео VibeVoice (Speech Generation/Voice Cloning) on Framework Desktop with Strix Halo (AMD AI Ryzen MAX+) канала Donato Capitella