Suggestion: Add Q8_0 quant
Just learnt about CrispASR, and wanted to say that you guys are AWESOME. ❤
I'm playing with this VoxCPM2 model for some cross-language voice cloning (Japanese reference voice to English/Chinese speech). It works miles better than Qwen-TTS, which only spits out gibberish for me. I have discovered that, while Q4_K works reasonably well, the full-precision F16 model actually produces voices with much better emotion and pacing. Out of curiosity, I made a Q8_0 quant by myself with crispasr-quantize, it works just like F16, and is 56.8% in size (2.63GB vs 4.63GB). It makes Q4_K looks like "not doing the model justice" by comparsion. So I thought it is maybe a good idea to also have a Q8_0 quant available for download.
P.S. Also tried Q6_K. It doesn't really save any disk space (only ~500MB smaller than Q8_0), and voice quality is noticeably lower than Q8_0, thus not recommended.