notesum.ai

Published at December 5

Representation Purification for End-to-End Speech Translation

cs.CL

cs.SD

eess.AS

Released Date: December 5, 2024

Authors: Chengwei Zhang¹, Yue Zhou¹, Rui Zhao¹, Yidong Chen¹, Xiaodong Shi¹

Aff.: ¹Xiamen University

Arxiv: http://arxiv.org/pdf/2412.04266v1

Refer to caption

Training in transcript-free setting
Models	En-De	En-Fr	En-Ru	En-Es	En-It	En-Nl	En-Pt	En-Ro	Avg.
Fairseq ST Wang et al. (2020a)	22.7	32.9	15.3	27.2	22.7	27.3	28.1	21.9	24.8
Revisit ST Zhang et al. (2022)	23.0	33.5	15.6	28.0	23.5	-	-	-	-
W2V2-TransformerFang et al. (2022)	24.1	35.0	16.3	29.4	24.8	28.9	30.0	23.1	26.5
CCSRD Zhao et al. (2023)	25.4	35.8	16.8	30.2	25.8	-	-	-	-
DUB (Large) Zhang et al. (2023a) ${\dagger}$	26.2	35.3	-	30.4	-	-	-	-	-
BT4ST Fang and Feng (2023) ${\dagger}$	26.6	36.9	-	31.2	-	-	-	-	-
SRPSE	26.2*	36.5*	17.6*	31.2*	26.1*	30.4*	31.9*	24.6*	28.0*
Training in multi-task setting
Memory-ST Yuan et al. (2024)	23.2	33.5	-	28.6	23.9	27.6	28.7	-	-
XSTNet Ye et al. (2021b)	25.5	36.0	16.9	29.6	25.5	30.0	31.3	25.1	27.5
STEMM Fang et al. (2022)	25.6	36.1	17.1	30.3	25.6	30.1	31.0	24.3	27.5
ConST Ye et al. (2022)	25.7	36.8	17.3	30.4	26.3	30.6	32.0	24.8	28.0
MCTN Zhou et al. (2024)	25.9	36.1	17.1	30.3	25.7	-	-	-	-
Siamese-PT Le et al. (2023)	26.2	36.9	16.8	29.8	25.9	29.8	32.1	24.8	27.8
CCSRD Zhao et al. (2023)	26.1	37.1	17.8	31.0	26.4	-	-	-	-
M³ST Cheng et al. (2023)	26.4	37.2	18.3	31.0	26.6	30.9	32.8	25.4	28.6
CMOT Zhou et al. (2023)	27.0	37.3	17.9	31.1	26.9	31.2	32.7	25.3	28.7
SRPSE	26.9*	37.4*	18.3*	31.4*	27.0*	31.4*	32.8*	25.5*	28.8*