They didn't say it's due to tokenization. > This is likely because they’re trained on a lot of data generated synthetically with text-to-speech and/or because

They didn't say it's due to tokenization. > This is likely because they’re trained on a lot of data generated synthetically with text-to-speech and/or because understanding the tone of the voice (apparently) doesn’t help the models make more accurate predictions.