Question About Multi-Language and Multi-Speaker Performance

[anhuynh219]

I see that it's possible to train both the variance and acoustic models with multiple languages and multiple speakers.

I have a few questions:

1. In a multi-dictionary or multi-language setup, does the performance in one language affect the others?
2. From my experiments with multi-speaker models, it seems that speakers with shorter training data tend to be influenced by other voices with longer training data, especially when the generated pitch goes beyond their original training range. Is this expected behavior?

Also, is there any prior evaluation or research on multi-speaker and multi-language performance?

[yqzhishen]

1. Yes, they affect each other, model will learn some common characteristics in singing.
2. Yes, speakers with less data can learn from speakers with more data.