tensorflow-wavenet icon indicating copy to clipboard operation
tensorflow-wavenet copied to clipboard

Published 20 hours ago • verified publisher icon ibab

Better silence trimming

Open lemonzi opened this issue 9 years ago • 3 comments

The current implementation computes the short-time RMSE amplitude, applies a given threshold on its value (the units are raw amplitude), and trims the audio from the beginning until a frame above the threshold is found, and from the last frame above the threshold until the end. If the result is empty, that file is discarded.

At some point, we might want to implement a better algorithm, such as using a threshold relative to the maximum amplitude in the example or applying a smoothing filter. The algorithm should also be configurable.

lemonzi avatar Sep 20 '16 23:09 lemonzi

Related: #59.

lemonzi avatar Sep 21 '16 01:09 lemonzi

Related: https://github.com/librosa/librosa/issues/360

fehiepsi avatar Oct 26 '16 10:10 fehiepsi

Following up on this, any updated functionality that you recommend using?

PetrochukM avatar Jun 01 '18 19:06 PetrochukM