# Attentron: Few-shot Text-to-Speech Exploiting Attention-based Variable Length Embedding

Seungwoo Choi *     Seungju Han *     Dongyoung Kim *     Sungjoo Ha

Hyperconnect
Seoul, Republic of Korea
Accepted to Interspeech 2020

* Equal contributions, listed in alphabetical order.      † Corresponding author.

## Abstract

On account of growing demands for personalization, the need for a so-called few-shot TTS system that clones speakers with only a few data is emerging. To address this issue, we propose Attentron, a few-shot TTS model that clones voices of speakers unseen during training. It introduces two special encoders, each serving different purposes. A fine-grained encoder extracts variable-length style information via an attention mechanism, and a coarse-grained encoder greatly stabilizes the speech synthesis, circumventing unintelligible gibberish even for synthesizing speech of unseen speakers. In addition, the model can scale out to an arbitrary number of reference audios to improve the quality of the synthesized speech. According to our experiments, including a human evaluation, the proposed model significantly outperforms state-of-the-art models when generating speech for unseen speakers in terms of speaker similarity and quality.

## Text-to-Speech Samples for Unseen Speakers During Training

These examples are sampled from the evaluation set for Table 1 in the paper. Each column corresponds to a single speaker, and each row corresponds to different models. Our proposed model at last row, Attentron(8-8), showed the best result from evalutation including MOS in terms of the speaker similarity and speech quality.

VCTK p304VCTK p311VCTK p316VCTK p305VCTK p306VCTK p312
Text It's totally double standards. His third goal was superb. He declined to give further details. We had a reunion last week. You'd think there was a match on today. I think it must be the uniforms.
Ground-truth
LDE(1)
LDE(8)
GMVAE(1)
GMVAE(8)
Attentron(1-1)
Attentron(8-8)

## Methods

In this paper, we propose Attentron, a novel architecture of few-shot TTS for unseen speakers. We adopted two main components for our model:
• Fine-grained encoder, which includes an attention mechaninsm to extract detailed styles from multiple reference audios. It allows the model to take any number of the reference audios, and the quality improves with more reference audios. The fine-grained encoder generates a variable-length embedding which maintains temporal information.
• Coarse-grained encoder, which extracts overall information of speech and helps to stabilize the outputs.

### Citation

To refer our work, please cite our paper as follows:  @misc{choi2020attentron,   title={Attentron: Few-Shot Text-to-Speech Utilizing Attention-Based Variable-Length Embedding},   author={Seungwoo Choi and Seungju Han and Dongyoung Kim and Sungjoo Ha},   year={2020},   eprint={2005.08484},   archivePrefix={arXiv},   primaryClass={eess.AS} } 
This page uses a template from the project page of Ha et al, "MarioNETte: Few-shot Face Reenactment Preserving Identity of Unseen Targets".