Splitting a long audio file

[sbersier]

I wanted to share here a tool that you might feel helpful (or not...)

Consider the following file (librivox):
https://ia801401.us.archive.org/25/items/beckoningfairone_2211_librivox/beckoningfairone_08_onions_128kb.mp3

The file is 30 min and 46 sec.

In order to train a voice, the samples should be less than ~10 sec (cf Notes on https://github.com/voicepaw/so-vits-svc-fork) and typically more than one second.

SVC comes with a splitter: svc pre-split. So I put my mp3 file in dataset_raw_raw. Then, svc pre-split spits the output in dataset_raw.

The smallest file is 1.2 sec long an contain 0.4 sec of non silent audio, while the longest is 29 seconds with quiet a lot of files longer than 10 sec. Below, an histogram of the resulting lengths:

Certainly, fiddling with the parameters you can probably achieve a better result but the overall shape stays roughly the same (unless you put the threshold so high that you have a lot of very small files.)

So, I made my own audio splitter: split_audio.py in which you specify the desired average length (default is 5 sec). The default (recommended) usage is:

python audio_split.py --desired_duration <desired average duration [sec]> <path/to/your/long/audio/file>

The resulting distribution looks like this:

Does it help training or not? I don't know. I haven't compared yet split_audio.py and svc pre-split side by side.

You can download it from:

https://github.com/sbersier/split_audio

(PS: I hope split_audio is not too buggy... If you find bugs, don't hesitate to report.)

[hataori-p]

Is it splitting in pauses?
(I'm sorry I haven't tried it yet)

Edit: I'v just read the desc in your repo. Don't bother with a reply.