Preprocess metadata and wavs in finetuning

[nickonlyone]

How should I preprocess metadata and wavs? I want to use vits2 to finetune and it says in readme that You can prepare metadata.csv with python script if your input data has different format. But what script should I use? And how to make my wav files in right format? I see .mel.pt in db-finetune folder, when i downloaded it. But I can not understand how to add this files to my own wav files

[nshmyrev]

mel.pt created automatically, you don't need them originally.

For metadata.csv you can create your own script

We recommend to tune stabletts checkpoints instead of vits2, it is much higher quality

[nickonlyone]

s

Thanks for the answer.
Now I have a question how to prepare metadata for training in stableTTS? I see no readme that can helps me. I want to train in russian

[nshmyrev]

It is exactly same format as for VITS

[nickonlyone]

I have an error: IndexError: index 3 is out of bounds for dimension 0 with size 3 when I try to train with stabletts.
It happens in this line: phone_embeddings.append(embeddings[word_index]) in vosk-tts/training/stabletts/matcha/text/init.py file. Any solution for this? Many thanks for your answers

This is how my lines in train.txt looks: wavs/027_part600.wav|0|Только потому он и остался жив.|_

[nshmyrev]

Вот тут

https://github.com/alphacep/vosk-tts/blob/master/training/stabletts/matcha/data/text_mel_datamodule.py#L219

поменять text_to_sequence_aligned на text_to_sequence_g2p.

[nickonlyone]

Поменял, но text_to_sequence_g2p принимает и возвращает только одно значение. Что в таком случае делать с перменной bert и text_norm? Просто еще сама функция get text возвращает все эти значения.
В text_to_sequence_g2p я стал передавать просто text без aligned.
И правильно я понял что функция text_to_sequence_g2p вызывается из папки vits2?