Pipeline challenges & fixes (part 3)
Image fetching was unreliable -> random, irrelevant images + strict rate limiting made it worse.
Also ran into issues with Ollama not starting consistently.
Solution:
Moved the pipeline to WSL -> both Ollama and the voice generator run significantly faster there.
Now using Ollama to:
โข generate better keywords for image fetching
โข create intro/outro
โข act as a bridge between sections
To avoid conflicts, I explicitly shut down Ollama before running the voice generator, wait ~5 seconds, then start Kokoro.
Still the best voice output so far.
Slowly stabilizing the system.