Skip to content

File length and segmentation

Frederic Brodbeck requested to merge file-length-and-segmentation into master

1. file lengths

I double-checked, and it looks like 10% of the files are not exactly 1 second long. they are shorter:

image

I propose pre-processing the dataset, by zero-padding the files in question. see https://scm.medienhaus.udk-berlin.de/freder/pi-zero-asr-paper/-/blob/file-length-and-segmentation/code/utils.py#L16

@Paul did you already do padding somewhere? (I feel like I remember you having talked about it)

2. feature extraction: segments

re-reading https://github.com/jameslyons/python_speech_features#mfcc-features I think, we should not use the default value for winlen but set our own. in fact winlen and winstep should be equal – which would result in the exact number of frames/segments we specify, without any gaps or overlap.

Edited by Frederic Brodbeck

Merge request reports