Do you need a strong correspondence between the input text and lang_id

[a779159990]

Hello, I would like to ask about the use of text and language ID as input. If the text and ID do not match, there may be repeated audio conversations, such as
{
"language": "KO",
"text": "你好，你叫什么名字，很高兴认识你",
}
Can this situation be monitored or monitored through a parameter setting in the model, or does it require checking the correspondence between text and lang_id before inputting into the model?

[Flux9665]

There is no way to figure out the language of the text once you put it into the model, so you need to make sure that the text and the language match before putting it in.

Here are some more details: There are two language settings internally in the model. For simplicity, we always change both to be the same in the demos, but if you write the code to run the model yourself, you can modify them individually if you want.

The first one always has to be the same as the language of the text and it can be changed with the following method:

IMS-Toucan/InferenceInterfaces/ToucanTTSInterface.py

Line 113 in 6355e67

def set_phonemizer_language(self, lang_id):

The second one can be a different one than the text. If it matches, it sounds more like the speaker speaks without accent. If it is a different one, it will sound more like the speaker has an accent of that second language. You can change it with the following method:

IMS-Toucan/InferenceInterfaces/ToucanTTSInterface.py

Line 116 in 6355e67

def set_accent_language(self, lang_id):

A friend and I noticed that for many Asian languages, the choice of the speaker also has an impact on the accent, which is an accident and we are trying to improve it.