read_fast.py audio quality compromise relative to read.py

[kebwi]

I'm quite pleased with the speed-up of read_fast.py over read.py. 25s audio generation took 27 minutes via read.py on preset fast and 17 minutes on preset ultra_fast (M1 Mac Studio) and took 6 minutes via read_fast.py. It doesn't appear that read_fast.py supports any preset options. However, I'm noticing a number of audio quality issues with the read_fast output. Most severely, the transcript consisted of two sentences and several words of the second sentence were dropped entirely. That won't do of course. What techniques can I employ to ensure there are no dropped words? Other than that, the quality is broadly worse in the read_fast.py case. For one thing, the prosody is much worse, especially the pauses (which I find problematic across the board with tortoise; I have yet to generate any audio I consider usable to produce an audiobook and I'm still only experimenting with short paragraphs, so this whole endeavor feels hopeless, but I'm hoping there are knobs I can fiddle with to improve it). So, what can be done to improve the overall audio quality of read_fast.py, especially its handling of pauses, and what can be done to improve prosody (especially pauses) of tortoise across all usages?

Thank you.