I am planning to extend TTS to TF using eager mode training. If you are interested please let me know so we can work together. I already put something out which works fine without tf.function tracing but I need more TF expertise to go beyond.
Do I get it right that if this works out right we probably can get a TfLite/TensorRT model out of it that may allow inference on lower end hardware?
Yes, I’d say it looks more promising with TF. We can also pass models between torch and tensorflow. Maybe we can initially target only inference so we can pass our pretrained models and use them right away. Then we can step forward to training and the rest