Curious about differences in onnx export versions
#1
by Likster - opened
Would you be able to share the differences in methodology between your version and the KevinAHM/pocket-tts-onnx version?
Thank you!!!!
They're virtually the same implementation, I just cleaned up the code for generating the ONNX models and added tokenizer.json, tokenizer_config.json and the different embeddings to the export.
I should also point out that the q4 exports are basically useless, since they degrade quality too much to be usable. The best results are gotten by using flow_lm_flow and flow_lm_main in int8 and the rest in fp32.