Could be temperature or often its just something missed in the training samples.
Would love to know how far prompt engineering can go and even the effect punctuation has in the text such as : and other ways of expressing dialogue and context. Seems to be little out there. I…
A smart trick I ran into is running generated audio through RVC with its own Harvest trained model, helps smooth out the audio and make it more consistent
Have you played around with this?…
I wouldn't be surprised if those artifacts and weirdness are because of so many short samples used to train it on.
I would also increase the temperature to over .5, and increase diffusion temp…
I started editting it but it was so tedious I decided to redo the sample set again in audacity and actually validate this time to cull the very short segments.
How do you go about manually changing the whisper.json to re-slice without messing with the metrics that the whisper json creates? Will that not cause some sort of conflict or is that extra data…
Overall, it's better to be testing the checkpoints along the way and until it no longer changes or begins to get worse. I will disagree with gforce on the autoslicing bit, I feel it's better…
Its just the method the original repo by neonjib used, I havent made any thorough comparison but I gather the 'new' method only uses the first 4 secs of the sample clip iirc, whereas the old one…
Honestly I wing it using the recommended settings more or less and get good results. I've found no considerable difference between 1000 steps and 2500, seems to be an efficient training method. I…
Yes as usual I have found the quality of the training data is very important, a test model I made had no considerable differences in the output quality whether it was using 12 samples or 50,…
So using that commit above the yaml can be generated and then I just needed a duplicate of each in /training and in subject dataset file, for some reason it changes where it wants the file to be…
Shit, I forgot to do this. Gomen. I'm a bit tied up for the rest of the week, so...
You might get lucky with just copying an existing
do_tts.py(such as from 152334H's fork, I…
Oh right I see so its the updated google packages that may be stuffing it up, I did revert back to 31da215c5f0f77a249dd5131bf8c3a4f149299a3 and the file pathways worked again enough to actually…