Supported audio/index length?

[ghost]

Hello,

I am building a dataset to train with and need to ask a few questions before proceeding.

What is the max supported/suggested audio length? is several minutes alright or should the audio be limited to about ~20 seconds or so? Likewise, is there a reasonable limit to the length of the index?

Thank you.

[Sleepwalking]

Hello, this is a bit case-dependent. My experience tells that long files are tricky to handle. For speech data it works the best with audio under 20 seconds. Actually the memory usage is O(N^2) where N is the file length so it'll quickly run out of memory for anything lasting longer than a minute. Quality also degrades when there are long silences.

There are tricks to tame those long audio files. For example, you can do this in two steps. First drop frames from the acoustic features and do a rough alignment. With this initial alignment, you can then cut the file into smaller pieces and do full scaled alignment on each of them.

[ghost]

I see. I should have clarified that I was asking about training and not just alignment.
The audio was way too long, several minutes each, and while I wasn't running out of memory the training wasn't progressing. I roughly segmented the audio with another tool and it worked as expected.

Thank you very much for the advice. Especially with the suggested tricks.