Create State Align Labels for Synthesis

Hello there, I am creating a TTS on a language which is not supported by festival.

I’ve trained a model based on state-aligned labels. I created the labels by myself, then aligned them using “run_state_aligner.sh”. The training went well, but when I’ve reached the synthesis phase, there was a problem with my non-aligned labels, which I’ve tried to synthesis.

Is there any way to use non-aligned labels as an input for the synthesis? It worked well on my phone-align labels, but it throws the following error with the state-align labels.

state_index = int(state_index) - 1
ValueError: invalid literal for int() with base 10: '-'

[ronanki]

@chazo1994 You are right.

We’ll soon provide support for full-contextual labels without time-stamps at synthesis time in order to get rid of this confusion. In which case:

1. The Input to duration model is full-contextual labels without time-stamps and outputs labels with time-stamps (by adding predicted durations).
2. The acoustic model always takes labels with time-stamps in order to synthesize speech.

[rhajit1001]

@chazo1994 You were right, thanks!