-
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
I'll go on a limb and say I forgot to document that you need to first transcribe your voice files under Training > Prepare Dataset, since Bark requires a transcription of the reference audio…
I'm answering in reverse order, since the answer to the BitsAndBytes question is quite lengthy.
Is the fast work really fast? tinkering with the number of samples and iterations in your UI…
What's soo fascinating about your tortoise implementation is that it only uses ~6 GB Vram while generating. That's not the case with the original tortoise and not even with tortoise-tts-fast…
All of the warnings are fine, they haven't caused me any issues months ago when I used it last.
Your main error is, because of the way DLAS is written, your gradient accumulation size being too…
a325496661
Merge pull request 'Freeze pydantic package to 1.10.11' (#301) from Jarod/ai-voice-cloning:master into master
Weekly evaluation time:
Naturally, reducing the amount of RVQ bins to attend to makes it much better for the model. In my head I've…
Seems more of a Gradio issue than a me thing: https://github.com/gradio-app/gradio/pull/4835
mmm, this isn't necessarily related to the model, but moreso EnCodec in general, but I probably should have done more tests on EnCodec (and Vocos), as I found a teensy little oversight: voices…
Just wanted to chime in any say this discussion has been a gold mine. I've spent the last hour pouring through all the the updates and discussions. But after reading your progress report…
I'm not too sure how there would be a discrepancy between the models, but the first error seems to be that you don't actually have any voice samples to feed into TorToiSe (or it's unable to…
mmm, I'm pretty sure TorToiSe would raise a warning message about the input/output being "too long" due to there not being any stop tokens generated, so it won't know when to "stop" after some…
If I'm reading DLAS's code right, this strictly governs the number of samples to poll for a similar voice clip during training. This is only related to the sampling rate insofar as to being…
6c3f48efba
uses gitmylo/bark-voice-cloning-HuBERT-quantizer for creating custom voices (it slightly works better over the base method, but still not very good desu)
God I really need a better place to keep track of "news" outside of this issues thread, but people are reading this somehow, so I suppose it's fine.
Bark
I fixed Bark support with the web…
547e1d1277
updated bark support, it'll also query for vocos, it actually works (I don't know what specifically was the issue)
47076c11df
fixed that pesky issue with the LR without a scheduler not being set, tweaked ZeRO settings to make it faster
Despite the accuracy being lower than model A, the audio quality of the validation output for the AR for model C seems remarkably clear (but inconsistently so). I wonder if this is because of…
b06fe38c1f
kludge to get Vocos decoding working with eval/validation, rdo not generate plain AR/NAR output if training both models because they are not at ALL good metrics to judge the model quality separately
Alrighty, I've eval'd with Vocos. I've labeled them to make them clear so I don't have to keep referring to them by their dataset size + model size + RVQ levels:
model A: the old weights on…