Help with Audio / Speech with Ollama

[vectorfield]

Thanks, that helps. That matches what I saw on dual 3090. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.