whisper.cpp setup on Fedora with AMD GPU support

This English summary uses the full English captions supplied for Adrien’s French-titled tutorial. Adrien walks through compiling whisper.cpp for local AI speech transcription on Fedora Linux with an AMD Radeon RX 7600. He uses Vulkan for GPU support and explains the alternative build options for CPU use and Nvidia CUDA. The tutorial covers dependencies, a dedicated service user, installation under /opt/whispercpp, shell environment variables, and SELinux contexts.

The demonstration uses the large-v3 model. Adrien reports roughly 3 GB of VRAM use after loading it and suggests small Q5 for less capable hardware. He avoids the English-specific .en models because his sample is in French. These are observations and recommendations from his setup, rather than performance guarantees for other machines.

He launches whisper-server on port 8081 with French selected explicitly. The --convert option uses FFmpeg installed on the machine to convert compressed audio. For the browser demonstration, he extracts an MP3 from a 2-minute-20-second video with Audacity, submits it, and compares the resulting text with the recording. He points out a transcription error in the channel name.

The final section configures a self-hosted systemd service, startup behavior, logs, and a firewall rule if needed. Adrien also proposes passing transcripts to llama.cpp for meeting summaries; this follow-up workflow is discussed rather than demonstrated.