표 6. | Table 6. 각 데이터셋에 대한 BPE 토큰 개수별 음절 오류율(%) | Character error rate (%) by number of BPE tokens for each dataset

Num. BPE	Fleurs-ko	Kmsav	Evalclean	Evalother
Char	4.30 (±0.04)	10.21 (±0.35)	6.54 (±0.06)	6.95 (±0.05)
5 k	4.51 (±0.13)	10.20 (±0.13)	6.63 (±0.08)	7.01 (±0.09)
10 k	4.29 (±0.05)	10.08 (±0.11)	6.52 (±0.02)	7.05 (±0.08)
15 k	4.29 (±0.10)	9.63 (±0.21)	6.48 (±0.07)	6.93 (±0.03)
20 k	4.29 (±0.14)	9.70 (±0.38)	6.45 (±0.04)	7.04 (±0.06)

The mean and standard deviation of the recognition results are shown for two Conformer models of the same structure trained with different initial values.