Generating Furina-Style Speech With CosyVoice3 On Arc B50 And Quadro M4000

Fri, 05 Jun 2026 00:00:00 +0000

This started as a simple automation idea: collect LLVM and HPC news, summarize it, and have a Furina-style voice read the result aloud.

The text side was not the interesting part. The hard part was making CosyVoice3 generate that audio on the right GPU.

There were two very different targets:

Intel Arc Pro B50, where the goal was to use the current upstream PyTorch XPU path instead of Intel Extension for PyTorch or an older Intel-specific fork.
NVIDIA Quadro M4000, an old Maxwell GPU, where the goal was to find the newest CUDA stack that still includes kernels for compute capability 5.2.

Both runs had to end with a real WAV/OGG file, and both had to prove that the core PyTorch model inference happened on the GPU rather than silently falling back to CPU.

Intel on ./Code

Generating Furina-Style Speech With CosyVoice3 On Arc B50 And Quadro M4000