-
https://git.ecker.tech/ aims to provide a place to share my efforts while maintaining true ownership of my code, as I do not trust GitHub.
XMR: 4B9TQdkAkBFYrbj5ztvTx89e5LpucPeTSPzemCihdDi9EBnx7btn8RDNZTBz2zihWsjMnDkzn5As1LU6gLv3KQy8BLsZ8SG
- Joined on
2022-10-10
sadly some of the models require a huggingface token to be functional
desu I don't think either VAD or diarization are necessary, as if your samples are that dirty, I don't think you should…
If you're ooming while generating try lowering your Sample Batch Size in settings.
It's OOMing when generating the latents. Low latent chunk counts with a very large dataset will OOM since…
It might have to do with sound file length. If there's even one really long sound file compared to the rest, it'll cause all other sounds to pad to the largest length to easily match. It could be…
I guess you have too large of a dataset. Bump it to 16 or 24, or however high it needs to be.
I just tested a model that had terrible output, and I manually set it to 24 chunks and it started…
For sure enable condition free, half-precision I'm still iffy on. I feel sometimes ever since BitsAndBytes was crammed in, it sometimes does boost quality a bit, but I feel it's placebo.
Which…
That is to also say, I might need to have it use the default behavior too, if it works better in some cases, since I believe the fast repo by default uses the old behavior (use the first 4 seconds…
I suppose so. I guess my fancy magic super duper accuracy boost is flawed in some cases.
If that's the case then, you can always manually set the voice chunk size to something to avoid it from…
Regarding latents and "Leveraging LJSpeech dataset for computing latents," does that mean that the latents are sensitive to the quality of the generated dataset by whisper/whisperx/whispercpp?
…
Yeah, it looks like the you can just copy ./voices/patrick/cond_latents_f90a07a1.pth file and just paste it into the fast repo's voice folder (I'd suggest a new voice folder, just to make sure…
Oh, there's one other difference between the tortoise-tts's: how the latents are calculated. I don't recall seeing the other tortoise saving latents, so there won't be an easy way to transplant…
Oh right, I can actually check the settings for the one generated through AIVC:
The only thing I can think of right now is that it's…
That's strange. Can you send over that model for me to test against?
I can't imagine I broke anything. The only thing I can think of would be it somehow not actually using the model for generating.
008a1f5f8f
simplified spawning the training process by having it spawn the distributed training processes in the train.py script, so it should work on Windows too
2feb6da0c0
cleanups and fixes, fix DLAS throwing errors from '''too short of sound files''' by just culling them during transcription
fc4e767676
cleanups and fixes, fix DLAS throwing errors from '''too short of sound files''' by just culling them during transcription
b78c882a95
cleanups and fixes, fix DLAS throwing errors from '''too short of sound files''' by just culling them during transcription
ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
Not an issue. It…