notesum.ai

Published at November 26

Learning Monotonic Attention in Transducer for Streaming Generation

cs.CL

cs.AI

Released Date: November 26, 2024

Authors: Zhengrui Ma¹, Yang Feng², Min Zhang³

Aff.: ¹Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences; University of Chinese Academy of Sciences; ²Key Laboratory of Intelligent Information Processing, Institute of Computing Technology, Chinese Academy of Sciences; Key Laboratory of AI Safety, Chinese Academy of Sciences; University of Chinese Academy of Sciences; ³School of Future Science and Engineering, Soochow University

Arxiv: http://arxiv.org/abs/2411.17170v1

		En-Es					En-De
	Chunk Size ( $ms$ )	320	640	960	1280	$\infty$	320	640	960	1280	$\infty$
Transducer	AL ( $ms,\downarrow$ )	886	1193	1591	1997	-	1126	1434	1830	2215	-
	BLEU ( $\uparrow$ )	24.33	25.82	26.36	26.40	26.75	19.99	22.10	22.20	22.96	23.10
	COMET ( $\uparrow$ )	67.94	69.92	70.48	70.65	71.14	62.81	65.01	65.75	66.26	67.03
	AL ( $ms,\downarrow$ )	997	1239	1606	1991	-	1215	1470	1860	2215	-
MonoAttn-Transducer	BLEU ( $\uparrow$ )	24.72	26.74	27.05	27.41	27.48	20.22	22.47	22.94	23.74	24.42
(Posterior)	COMET ( $\uparrow$ )	68.98	70.71	71.21	71.90	72.24	64.24	67.06	68.22	68.54	69.82
	AL ( $ms,\downarrow$ )	932	1182	1599	1967	-	1138	1413	1826	2191	-
MonoAttn-Transducer	BLEU ( $\uparrow$ )	23.00	26.46	27.07	27.42	27.48	19.26	22.62	23.51	24.01	24.42
(Prior)	COMET ( $\uparrow$ )	68.24	70.45	71.33	71.99	72.24	63.85	67.63	68.65	69.27	69.82