(Question) Max Length and datasets about styletts2 HOT 1 CLOSED

The dataset was prepared by someone else and it has become a benchmark dataset in TTS research: https://keithito.com/LJ-Speech-Dataset/. I honestly don't know any logic behind this splitting, but this seems to be a standard already and everyone is testing their models on this dataset.

As for when max_len is reached, it will just clip the samples in a batch to max_len. For example, if the length of your batch is [400, 300, 200, 100], then your batch length will be [4, 100] (which is the minimum of your batch size, while any sample longer than 100 will be clipped to 100 randomly. However, if you set max_len = 80, then you will get [4, 80] instead.

Yes, you can set max_len = 1200 if you have enough RAM. The heuristics is that the larger the value the better, so the model will learn better context. This is the same as the context window in LLM training, and people have found that longer context window helps with learning.

from styletts2.