MeloTTS icon indicating copy to clipboard operation
MeloTTS copied to clipboard

Wrong pronunciation in english(all voices)

Open loscrossos opened this issue 11 months ago • 4 comments

I have been using Melo a bit and just noticed it has several problems with pronunciation.

He's chokin', how? Everybody's jokin' now. we installed the plugin and all other plugins available.

this is a sentence from Eminems "lose yourself" and a random sentence.

The word "chokin" will be pronounced "cokin" and the word "plugin" will be randomly pronounced correctly and wrong ("ploogin") depending on the sentence. Other TTS do not have this problem..

There are other small errors but this are the ones i can safely reproduce.

i could reproduce these on my local installation and on this online demo.

https://huggingface.co/spaces/mrfakename/MeloTTS

loscrossos avatar Jan 22 '25 13:01 loscrossos

Any fix?

yukiarimo avatar Feb 09 '25 08:02 yukiarimo

have you tried the "EN_NEWEST" model? it's not ~~very well~~ documented but if you specify "EN_NEWEST" as language instead of "EN", the "v3" model that it loads is much higher quality. Not available through any of the web demos. See the download utils module for more info on what's available. There's only one speaker for this latest version of the model.

aaronjbecker avatar Feb 13 '25 21:02 aaronjbecker

Is it using a different BERT model? I'm using custom voice, so if it is not, then I am not interested.

yukiarimo avatar Feb 14 '25 06:02 yukiarimo

EN_NEWEST has the same problem

loscrossos avatar Feb 28 '25 08:02 loscrossos