I wasted some time over the weekend setting up speaches.ai and Ollama to try doing some audio transcription locally.
It went poorly, so I ended up just using gemini-flash-latest and paying for the damn tokens.
It was just using the Tesla P4 card in my server which is quite limited at 8GB and Pascal-era... perfect for transcoding Plex and Immich machine learning, but not quite up to the task of decent AI models for contextualising audio transcripts.