n0tls

experiments.n0tls.com


I spun up experiments.n0tls.com as a place to host small tools and one-off projects that don't quite fit as blog posts. Think interactive things, weird ideas, half-finished prototypes that actually work.

The first experiment is a voice evals tool — a way to compare output quality across open-weight TTS models by pulling levers on the underlying parameters.
It went from a vague conversation to a working tool in one sitting. The transcripts are up if you want to see how that unfolded:

The second experiment is a batch piper comparison — the same article synthesized two ways: one piper process per chunk (model reloads each time) vs. one process for all chunks (model loads once). A 29% speedup, and you can hear both to judge whether there's any quality difference.