Are you using a phone language model or word language model. I noticed that he predicted phone, which comes from arg max is able to identify silence, but for
- real_phone = alignment_phoneme[i]
It seems to put the non silence phone, so there is a mismatch for phones for example each word starts with a silence, so predicted phones will have silence phones but real phone from alignment will be non-silence, do you add a forced silence into the text?
Are you using a phone language model or word language model. I noticed that he predicted phone, which comes from arg max is able to identify silence, but for
It seems to put the non silence phone, so there is a mismatch for phones for example each word starts with a silence, so predicted phones will have silence phones but real phone from alignment will be non-silence, do you add a forced silence into the text?