Closed SamuelLarkin closed 3 months ago
This is breaking for a similar reason described in https://github.com/roedoejet/EveryVoice/issues/452#issuecomment-2166684559 except it's probably also the fact that you have the pitch_target and energy_target in the config set to 'phones' instead of 'frames', and if learn_alignment is set to False
, it will look for the durations instead of calculating the prior that is used by the jointly-learned alignment. This also appears to be working as-expected, but we could consider adding an error message. Honestly, turning this to False is really only intended to be done by advanced users of EveryVoice though, so I'm hesitant to provide too much direction here.
How to reproduce
everyvoice wizard
learn_alignment
everyvoice preprocess config/everyvoice-text-to-spec.yaml -c model.learn_alignment=false
Log