Cantonese awful quality
I've tried register 20sec .mp3 with complete typescript. The testing results is awful. The speech combining Mandarin with cantonese, full sentence mostly are"mandarin accent" with some "characters" is speaking as Mandarin. it's like expecting American English but output is like in movies hearing native Spanish/Italian/Russian speaking English.
Thank you for the detailed feedback. We’ve observed this issue as well. Cantonese and Mandarin share many of the same character tokens, while Mandarin currently has far more training data, so the model can strongly favor Mandarin pronunciation.
As a potential solution, we may introduce a configurable language selector as part of the context in a future version. In the meantime, using Cantonese reference audio can help mitigate the issue and improve the generation success rate.