significant performance regression in 0.0.2

[pzoltowski]

I'm seeing serious performance regression on version 0.0.2. I'm having macbook m2 max (with latest MacOS). Before I was having around ~12x realtime speedup factor (in some cases somtimes even 20x). Right now on 0.0.2 I'm getting ~4.0x speedup (sometimes 5x). I uninstalled 0.0.2 and reverted to 0.0.1 just to be double sure and again got good fast performance with older version. Running in python 3.12.8, macOS 15.3.2. Attached sample testing script

playground_kokoro-mlx.py.txt

[ivanfioravanti]

Can you please try to generate a longer audio with 0.0.1? Does it work? It was failing for me.

[pzoltowski]

It do work for me with longer text provided but noticed the pipeline will automatically split it into into max ~30seconds chunks if text is to long.