Generative Spoken Dialogue Language Modeling

Tu Anh Nguyen; Eugene Kharitonov; Jade Copet; Yossi Adi; Wei Ning Hsu; Ali Elkahky; Paden Tomasello; Robin Algayres; Benoît Sagot; Abdelrahman Mohamed; Emmanuel Dupoux

doi:10.1162/tacl_a_00545

Generative Spoken Dialogue Language Modeling

Tu Anh Nguyen, Eugene Kharitonov, Jade Copet, Yossi Adi, Wei Ning Hsu, Ali Elkahky, Paden Tomasello, Robin Algayres, Benoît Sagot, Abdelrahman Mohamed, Emmanuel Dupoux

Research output: Contribution to journal › Article › peer-review

4 Scopus citations

Abstract

We introduce dGSLM, the first ‘‘textless’’ model able to generate audio samples of naturalistic spoken dialogues. It uses recent work on unsupervised spoken unit discovery coupled with a dual-tower transformer architecture with cross-attention trained on 2000 hours of two-channel raw conversational audio (Fisher dataset) without any text or labels. We show that our model is able to generate speech, laughter, and other paralinguistic signals in the two channels simultaneously and reproduces more naturalistic and fluid turn taking compared to a text-based cascaded model.^1,2.

Original language	American English
Pages (from-to)	250-266
Number of pages	17
Journal	Transactions of the Association for Computational Linguistics
Volume	11
DOIs	https://doi.org/10.1162/tacl_a_00545
State	Published - 14 Mar 2023
Externally published	Yes

Bibliographical note

Publisher Copyright:
© 2023 Association for Computational Linguistics.

Access to Document

10.1162/tacl_a_00545

Cite this

@article{fb2de5ac7919423ca5c249fef2e482b3,

title = "Generative Spoken Dialogue Language Modeling",

abstract = "We introduce dGSLM, the first {\textquoteleft}{\textquoteleft}textless{\textquoteright}{\textquoteright} model able to generate audio samples of naturalistic spoken dialogues. It uses recent work on unsupervised spoken unit discovery coupled with a dual-tower transformer architecture with cross-attention trained on 2000 hours of two-channel raw conversational audio (Fisher dataset) without any text or labels. We show that our model is able to generate speech, laughter, and other paralinguistic signals in the two channels simultaneously and reproduces more naturalistic and fluid turn taking compared to a text-based cascaded model.1,2.",

author = "Nguyen, {Tu Anh} and Eugene Kharitonov and Jade Copet and Yossi Adi and Hsu, {Wei Ning} and Ali Elkahky and Paden Tomasello and Robin Algayres and Beno{\^i}t Sagot and Abdelrahman Mohamed and Emmanuel Dupoux",

note = "Publisher Copyright: {\textcopyright} 2023 Association for Computational Linguistics.",

year = "2023",

month = mar,

day = "14",

doi = "10.1162/tacl_a_00545",

language = "American English",

volume = "11",

pages = "250--266",

journal = "Transactions of the Association for Computational Linguistics",

issn = "2307-387X",

publisher = "MIT Press Journals",

}

TY - JOUR

T1 - Generative Spoken Dialogue Language Modeling

AU - Nguyen, Tu Anh

AU - Kharitonov, Eugene

AU - Copet, Jade

AU - Adi, Yossi

AU - Hsu, Wei Ning

AU - Elkahky, Ali

AU - Tomasello, Paden

AU - Algayres, Robin

AU - Sagot, Benoît

AU - Mohamed, Abdelrahman

AU - Dupoux, Emmanuel

PY - 2023/3/14

Y1 - 2023/3/14

N2 - We introduce dGSLM, the first ‘‘textless’’ model able to generate audio samples of naturalistic spoken dialogues. It uses recent work on unsupervised spoken unit discovery coupled with a dual-tower transformer architecture with cross-attention trained on 2000 hours of two-channel raw conversational audio (Fisher dataset) without any text or labels. We show that our model is able to generate speech, laughter, and other paralinguistic signals in the two channels simultaneously and reproduces more naturalistic and fluid turn taking compared to a text-based cascaded model.1,2.

AB - We introduce dGSLM, the first ‘‘textless’’ model able to generate audio samples of naturalistic spoken dialogues. It uses recent work on unsupervised spoken unit discovery coupled with a dual-tower transformer architecture with cross-attention trained on 2000 hours of two-channel raw conversational audio (Fisher dataset) without any text or labels. We show that our model is able to generate speech, laughter, and other paralinguistic signals in the two channels simultaneously and reproduces more naturalistic and fluid turn taking compared to a text-based cascaded model.1,2.

UR - http://www.scopus.com/inward/record.url?scp=85150992661&partnerID=8YFLogxK

U2 - 10.1162/tacl_a_00545

DO - 10.1162/tacl_a_00545

M3 - Article

AN - SCOPUS:85150992661

SN - 2307-387X

VL - 11

SP - 250

EP - 266

JO - Transactions of the Association for Computational Linguistics

JF - Transactions of the Association for Computational Linguistics

ER -

Generative Spoken Dialogue Language Modeling

Abstract

Bibliographical note

Access to Document

Other files and links

Fingerprint

Cite this