Performance of Training

[Sleepwalking]

Dear Maintainer,

I have been trying out sprocket for some standard voice conversion tasks. Although no thorough benchmark has been done, it came to my notice that sprocket trains a lot slower than a barebone SPTK-based GMMVC training script. This made hyperparameter tuning a somewhat time-consuming process.

A quick inspection of the source code shows that the training script is entirely numpy-based except for the analysis/synthesis part. I wonder if any profiling has been done to identify the bottleneck (the main suspects are probably EM and DTW) and if there is any plan to optimize the code?

Thank you for making this implementation open source.

Regards,
Kanru Hua

[Sleepwalking]

I just found that mgcep runs slow on high-dimensional inputs.
I did run cProfile on sprocket though, and it shows that besides feature extraction, scikit-learn's GMM implementation is taking a lot of time. Any plan to add an option to use SPTK's gmm?