Table 2.

Performance of various automatic speech recognition models on CMaiSpeech^a test sets (code-switching) in terms of the word error rate.

Methods	WER-CSb, %	WER-ENc, %	WER-ZHd, %
Character-based transfer learning	27.22	40.15	26.75
Syllable-based transfer learning	29.54	38.13	29.22
Azure cloud application programming interface	59.20	95.57	54.23
Character-based meta–transfer learning (proposed)	24.78	41.23	24.28
Syllable-based meta–transfer learning (proposed)	22.20	36.13	21.54

^aCMaiSpeech: China Medical University Hospital Artificial Intelligence Speech.

^bWER-CS: word error rate–code-switching.

^cWER-EN: word error rate–English.

^dWER-ZH: word error rate–Zhōngwén.