Motivation for calculating spectrograms during training about speech HOT 1 CLOSED

It depends what you mean by more efficient. Right now the data loading and preprocessing takes up a small fraction or the runtime (at least on my machine), so precomputing the spectrograms won't make the models train any faster. However, in terms of wasting less computation and hence energy, then yes it would be more efficient.

The main motivation for not pre-computing the features is that I wanted to design the data loading to make things like on the fly data augmentation easy to implement (for example adding in clips of noise to the audio). These types of transformations are typically done in the raw time-domain. However for TIMIT they aren't very common..

from speech.