notesum.ai

Published at October 30

Lina-Speech: Gated Linear Attention is a Fast and Parameter-Efficient Learner for text-to-speech synthesis

eess.AS

cs.AI

cs.SD

Released Date: October 30, 2024

Authors: Théodor Lemerle, Harrison Vanderbyl¹, Vaibhav Srivastav², Nicolas Obin³, Axel Roebel³

Aff.: ¹RWKV Foundation, Recursal AI, Eleuther AI, Sydney, Australia; ²Hugging Face, Paris, France; ³Sound Analysis and Synthesis STMS Lab IRCAM, CNRS, Sorbonne Université, Paris, France

Arxiv: http://arxiv.org/abs/2410.23320v1

Model	Objective eval.			Subjective eval.
	CER $\downarrow$	WER $\downarrow$	SECS $\uparrow$	N-MOS $\uparrow$	S-MOS $\uparrow$	#Params.
Ground Truth	1.5%	4.5%	-	$4.26\pm 0.19$	$4.30\pm 0.22$	-
StyleTTS2 Li et al. (2024)	0.8%	3.2%	0.89	$3.93\pm 0.22$	$4.02\pm 0.24$	148M
VALLE-X Plachtaa	7.6%	14.1%	0.92	$3.21\pm 0.27$	$3.07\pm 0.29$	300M
VoiceCraft Peng et al. (2024b)	3.4%	6.6%	0.94	$3.62\pm 0.24$	$3.55\pm 0.25$	833M
Lina-Speech (ours)	2.9%	7.5%	0.93	$\boldsymbol{4.18\pm 0.19}$	$\boldsymbol{4.10\pm 0.20}$	169M