IMPROVING SPEECH-TO-SPEECH TRANSLATION THROUGH UNLABELED TEXT Xuan-Phi Nguyeny Sravya Popuri Changhan Wang Yun Tang Ilia Kulikovand Hongyu Gong Meta AI USA

2025-05-08 0 0 311.02KB 5 页 10玖币

侵权投诉

IMPROVING SPEECH-TO-SPEECH TRANSLATION THROUGH UNLABELED TEXT

Xuan-Phi Nguyen∗† , Sravya Popuri?, Changhan Wang?, Yun Tang?, Ilia Kulikov?and Hongyu Gong?

?Meta AI, USA

†Nanyang Technological University, Singapore

nguyenxu002@e.ntu.edu.sg

{spopuri,changhan,yuntang,kulikov,hygong}@meta.com

ABSTRACT

Direct speech-to-speech translation (S2ST) is among the

most challenging problems in the translation paradigm due

to the signiﬁcant scarcity of S2ST data. While effort has

been made to increase the data size from unlabeled speech

by cascading pretrained speech recognition (ASR), machine

translation (MT) and text-to-speech (TTS) models; unlabeled

text has remained relatively under-utilized to improve S2ST.

We propose an effective way to utilize the massive existing

unlabeled text from different languages to create a large

amount of S2ST data to improve S2ST performance by

applying various acoustic effects to the generated synthetic

data. Empirically our method outperforms the state of the art

in Spanish-English translation by up to

BLEU. Signiﬁcant

gains by the proposed method are demonstrated in extremely

low-resource settings for both Spanish-English and Russian-

English translations.

Index Terms—

Speech-to-speech translation, augmentation,

unlabeled text

1. INTRODUCTION

Translating speech of a language to speech of another can be

done by trivially cascading an automatic speech recognition

(ASR) [

], machine translation (MT) [

], or combinatory

speech-to-text translation (S2T) [

] and ﬁnally a text-to-speech

(TTS) systems [

]. But such process suffers from

signiﬁcant inference latency and is prone to error propagation

through each stage. Alternatively, there is a growing interest in

developing direct speech-to-speech translation systems (S2ST)

[

]. Not only do these systems have faster inference

but also allow translations between unwritten languages and

dialects [

]. Recent speech-to-speech model [

] employs a

self-supervised pretrained speech encoder [

] and a discrete

unit mBART model to train a speech-to-unit translation

(S2UT), where the target speech is converted into discrete

units [

]. Such units are trained with self-supervision to

group speech frames based on their linguistic and prosodic

information using k-means clustering [16].

∗Work done during an internship at Meta AI.

Despite promising capability, direct speech-to-speech

(S2ST) models suffer from signiﬁcant data scarcity due to

the challenge of collecting human annotated parallel speech.

To improve S2ST performance, apart from self-supervised

pretraining on unlabeled speech [

], [

] also tried to

generate extra synthetic S2ST data from speech recognition

(ASR) data by converting its texts to speech in the target

language using a cascaded MT-TTS system [

]. Nonetheless,

such work only makes use of the available audio data, while

not utilizing the existing unlabeled text data from numerous

languages, sources and domains [

]. Text data is

known to be much more massive and diverse than the current

speech data. However, such textual data may be difﬁcult

to deal with in the speech paradigm as they lack crucial

information about speakers, speed, pitch and emotions.

In this work, we present an effective strategy to generate

synthetic speech-to-speech training data from the abundant

unlabeled text data so that the resulting speech data is not only

diverse in terms of semantic content, but also randomly varying

in acoustic features such as speaker tones. Our approach

consists of two processes: (i) “Text-aug” data generation,

which is the synthetic S2ST data created from unlabeled text;

and (ii) “Effects-aug” process, which is an on-the-ﬂy speech

augmentation process that transforms toneless Text-aug speech

into a varying-tone and noisy version that tries to mimic

the distribution of real speech data. Our approach does not

introduce any extra model or data supervision besides those

used in the recent S2UT baseline [

], which use supervised

MT and TTS models. Instead, we utilize an unsupervised MT

model [

] to generate data. In the experiments, our method

achieves up to 35.2 BLEU on the CoVoST-2 Es-En task and

35.1 BLEU on Europarl-ST En-Es task, surpassing the state-

of-the-art approach [

] by up to 2 BLEU. Further analysis

shows that Effects-aug is a crucial step for the extra data

to improve the performance. We also demonstrate that our

method achieves signiﬁcant performance gain of up to 28

BLEU in low-resource speech-to-speech setups with only 10

hours (hr) S2ST data.

arXiv:2210.14514v1 [cs.CL] 26 Oct 2022

文档加载中……请稍候！
如果长时间未打开，您也可以点击刷新试试。

下载文档到电脑，查找使用更方便

10 玖币 0人已下载

立即下载

摘要：

IMPROVINGSPEECH-TO-SPEECHTRANSLATIONTHROUGHUNLABELEDTEXTXuan-PhiNguyeny,SravyaPopuri?,ChanghanWang?,YunTang?,IliaKulikov?andHongyuGong??MetaAI,USAyNanyangTechnologicalUniversity,Singaporenguyenxu002@e.ntu.edu.sg{spopuri,changhan,yuntang,kulikov,hygong}@meta.comABSTRACTDirectspeech-to-speechtranslat...

展开>> 收起<<

IMPROVING SPEECH-TO-SPEECH TRANSLATION THROUGH UNLABELED TEXT Xuan-Phi Nguyeny Sravya Popuri Changhan Wang Yun Tang Ilia Kulikovand Hongyu Gong Meta AI USA.pdf

共5页,预览1页

还剩页未读，继续阅读

声明：本站为文档C2C交易模式，即用户上传的文档直接被用户下载，本站只是中间服务平台，本站所有文档下载所得的收益归上传人(含作者)所有。玖贝云文库仅提供信息存储空间，仅对用户上传内容的表现方式做保护处理，对上载内容本身不做任何修改或编辑。若文档所含内容侵犯了您的版权或隐私，请立即通知玖贝云文库，我们立即给予删除！

IMPROVING SPEECH-TO-SPEECH TRANSLATION THROUGH UNLABELED TEXT Xuan-Phi Nguyeny Sravya Popuri Changhan Wang Yun Tang Ilia Kulikovand Hongyu Gong Meta AI USA

相关推荐

开通VIP享超值会员特权

作者详情

相关内容

热门标签

举报选择: