AdaDurIAN: Few-shot Adaptation for Neural Text-to-Speech with DurIAN


Paper

Submitted to InterSpeech 2020 [Here]

Authors

Zewang Zhang, Qiao Tian, Heng Lu, Ling-Hui Chen, Shan Liu

Abstract

This paper investigates how to leverage a DurIAN-based average model to enable a new speaker to have both accurate pronunciation and fluent cross-lingual speaking with very limited monolingual data. A weakness of the recently proposed end-to-end text-to-speech (TTS) systems is that robust alignment is hard to achieve, which hinders it to scale well with very limited data. To cope with this issue, we introduce AdaDurIAN by training an improved DurIAN-based average model and leverage it to few-shot learning with the shared speaker-independent content encoder across different speakers. Several few-shot learning tasks in our experiments show AdaDurIAN can outperform the baseline end-to-end system by a large margin. Subjective evaluations also show that AdaDurIAN yields higher mean opinion score (MOS) of naturalness and more preferences of speaker similarity. In addition, we also apply AdaDurIAN to emotion transfer tasks and demonstrate its promising performance.


Samples

Performances with different modules to be fixed during 1-minute few-shot speaker adaptation. All audios are converted by Griffin-Lim algorithm from mel spectrogtam.

text fix nothing fix phone embedding fix phone, tone (stress) and language embedding fix phone, tone (stress), language embedding and encoder
春眠不觉晓,处处闻啼鸟。夜来风雨声,花落知多少。
如今硝烟已散,对于普通旅行者而言,南高加索是一个陌生又熟悉的名词。
The future is scary but you can't just run to the past cause it's familiar.

Performances of speaker adaptation with different speakers and different amount of training data. All audios are converted by Griffin-Lim algorithm from mel spectrogtam.

Speaker adaptation of CN speaker with 1 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.

Speaker adaptation of EN speaker with 1 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.

Speaker adaptation of CN speaker with 3 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.

Speaker adaptation of EN speaker with 3 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.

Speaker adaptation of CN speaker with 20 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.

Speaker adaptation of EN speaker with 20 minute training data.

text SMA AdaDurIAN
倒计时十秒钟,10,9,8,7,6,5,4,3,2,1,0
If you find a path with no obstacles, it probably doesn’t lead anywhere.