I cannot speak for the authors, but my impression is that the vocoder (WaveRNN) part of current TTS systems takes much more computing than the part that this research addresses, so this may not help all that much.