BlueMagpie-TTS improves Taiwanese code-switching speech synthesis

The authors introduce BlueMagpie-TTS, a system designed to address the poor adaptation of off-the-shelf text-to-speech models to Taiwanese Mandarin accents and code-switching contexts. The system integrates PangolinTokenizer, Barbet language model, and VoxCPM2 acoustic stack to reduce tokenization errors and improve pronunciation accuracy.

• PangolinTokenizer achieves the lowest token rate (0.485 tokens/character) among nine tested tokenizers.
• Barbet, a billion-parameter Traditional-Chinese language model, ranks first on a 14-task evaluation.
• BlueMagpie-TTS lowers Character Error Rate from 11.45% to 4.81% and Word Error Rate from 14.83% to 5.36%.
• In blind listening tests, 65.6% of majority votes preferred BlueMagpie-TTS over other systems.

The system provides significant improvements in accuracy and naturalness for Taiwanese-accented code-switching speech synthesis.