Should I use mean-var normalization for single user if stats_path is defined

[utunga]

Hi!

Just getting started training a model with Coqui TTS. I'm starting with just training tacotron model to start - using train_tacotron.py script. For the record I forked the repo - to add my own data preprocess setps at the v0.0.14 tag.

The pre-processing step correctly creates the stats and its a single speaker audio source so I feel like its approprirate to set the stats_path and set the signal_norm bool to true.

config.json settings around normalization are like this:

However I ran into a problem when I had signal normalization enabled. So I had to set signal_norm to false in order to actually run a train. (Which works but is far from ideal)

Specifically this part of audio.py normalize() function would fail every time:

Basically for some reason the 0th axis of the sample data was 513 when the size_fft/2 was 512.

I mention this just in case someone else has seen something similar??? Bit of a long shot, to be honest, but I figure it can't hurt to ask.

Or maybe I need upgrade to another version or something?

[utunga]

Thanks in advance for any comments - much appreciated!