Extract duration from tacotron2 model

[linhld0811]

I don't send any issue to this comment. Following to the tutorial in training model fastspeech2, we have to extract the duration from alignment of tacotron2 model( on function get_duration_from_alignment on file extract_duration.py).
I just want to know what exactly of this term "duration". Anyone help me to figure out this definition?!

[trfnhle]

Character and Mel-spectrogram have different lengths, len(character)< len(mel). In fast speech, we mapping from characters to mel so each character will represent for a chunk mel and length of this chunk, we call duration.

[linhld0811]

I have a question: Can i use time alignment( duration in ms of each phoneme which is extracted by kaldi) for transform to duration format that you use for training fastspeech2?!

[azraelkuan]

@linhld0811 yes, but you should use the same phoneme in tts as you use in kalid.
check this: https://montreal-forced-aligner.readthedocs.io/
and this example file for ljspeech: https://github.com/ivanvovk/DurIAN/blob/master/filelists/train_filelist.txt