An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR

Aku Rouhe; Astrid Van Camp; Mittul Singh; Hugo Van Hamme; Mikko Kurimo

doi:10.1007/978-3-030-87802-3_54

An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR

Aku Rouhe^*, Astrid Van Camp, Mittul Singh, Hugo Van Hamme, Mikko Kurimo

^*Corresponding author for this work

Speech Recognition

KU Leuven

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

4 Citations (Scopus)

140 Downloads (Pure)

Abstract

Standard end-to-end training of attention-based ASR models only uses transcribed speech. If they are compared to HMM/DNN systems, which additionally leverage a large corpus of text-only data and expert-crafted lexica, the differences in modeling cannot be disentangled from differences in data. We propose an experimental setup, where only transcribed speech is used to train both model types. To highlight the difference that text-only data can make, we use Finnish, where an expert-crafted lexicon is not needed. With 1500h equal data, we find that both ASR paradigms perform similarly, but adding text data quickly improves the HMM/DNN system. On a smaller 160h subset we find that HMM/DNN models outperform AED models.

Original language	English
Title of host publication	Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings
Editors	Alexey Karpov, Rodmonga Potapova
Publisher	Springer
Pages	602-613
Number of pages	12
ISBN (Print)	9783030878016
DOIs	https://doi.org/10.1007/978-3-030-87802-3_54
Publication status	Published - 2021
MoE publication type	A4 Conference publication
Event	International Conference on Speech and Computer - Virtual, Online Duration: 27 Sept 2021 → 30 Sept 2021 Conference number: 23

Publication series

Name	Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
Volume	12997 LNAI
ISSN (Print)	0302-9743
ISSN (Electronic)	1611-3349

Conference

Conference	International Conference on Speech and Computer
Abbreviated title	SPECOM
City	Virtual, Online
Period	27/09/2021 → 30/09/2021

Keywords

Attention-based Encoder-Decoder
Equal data
HMM/DNN

Access to Document

10.1007/978-3-030-87802-3_54

E2E_vs_HMM_SPE_ComAccepted author manuscript, 242 KB

Cite this

Rouhe, A., Van Camp, A., Singh, M., Van Hamme, H., & Kurimo, M. (2021). An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR. In A. Karpov, & R. Potapova (Eds.), Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings (pp. 602-613). (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12997 LNAI). Springer. https://doi.org/10.1007/978-3-030-87802-3_54

Rouhe, Aku ; Van Camp, Astrid ; Singh, Mittul et al. / An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR. Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings. editor / Alexey Karpov ; Rodmonga Potapova. Springer, 2021. pp. 602-613 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)).

@inproceedings{53a7261a503e4a5a9683849902529a59,

title = "An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR",

abstract = "Standard end-to-end training of attention-based ASR models only uses transcribed speech. If they are compared to HMM/DNN systems, which additionally leverage a large corpus of text-only data and expert-crafted lexica, the differences in modeling cannot be disentangled from differences in data. We propose an experimental setup, where only transcribed speech is used to train both model types. To highlight the difference that text-only data can make, we use Finnish, where an expert-crafted lexicon is not needed. With 1500h equal data, we find that both ASR paradigms perform similarly, but adding text data quickly improves the HMM/DNN system. On a smaller 160h subset we find that HMM/DNN models outperform AED models.",

keywords = "Attention-based Encoder-Decoder, Equal data, HMM/DNN",

author = "Aku Rouhe and {Van Camp}, Astrid and Mittul Singh and {Van Hamme}, Hugo and Mikko Kurimo",

note = "| openaire: EC/H2020/780069/EU//MeMAD; International Conference on Speech and Computer, SPECOM ; Conference date: 27-09-2021 Through 30-09-2021",

year = "2021",

doi = "10.1007/978-3-030-87802-3_54",

language = "English",

isbn = "9783030878016",

series = "Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)",

publisher = "Springer",

pages = "602--613",

editor = "Alexey Karpov and Rodmonga Potapova",

booktitle = "Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings",

address = "Germany",

}

Rouhe, A, Van Camp, A, Singh, M, Van Hamme, H & Kurimo, M 2021, An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR. in A Karpov & R Potapova (eds), Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings. Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics), vol. 12997 LNAI, Springer, pp. 602-613, International Conference on Speech and Computer, Virtual, Online, 27/09/2021. https://doi.org/10.1007/978-3-030-87802-3_54

An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR. / Rouhe, Aku; Van Camp, Astrid; Singh, Mittul et al.
Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings. ed. / Alexey Karpov; Rodmonga Potapova. Springer, 2021. p. 602-613 (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics); Vol. 12997 LNAI).

Research output: Chapter in Book/Report/Conference proceeding › Conference article in proceedings › Scientific › peer-review

TY - GEN

T1 - An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR

AU - Rouhe, Aku

AU - Van Camp, Astrid

AU - Singh, Mittul

AU - Van Hamme, Hugo

AU - Kurimo, Mikko

N1 - Conference code: 23

PY - 2021

Y1 - 2021

N2 - Standard end-to-end training of attention-based ASR models only uses transcribed speech. If they are compared to HMM/DNN systems, which additionally leverage a large corpus of text-only data and expert-crafted lexica, the differences in modeling cannot be disentangled from differences in data. We propose an experimental setup, where only transcribed speech is used to train both model types. To highlight the difference that text-only data can make, we use Finnish, where an expert-crafted lexicon is not needed. With 1500h equal data, we find that both ASR paradigms perform similarly, but adding text data quickly improves the HMM/DNN system. On a smaller 160h subset we find that HMM/DNN models outperform AED models.

AB - Standard end-to-end training of attention-based ASR models only uses transcribed speech. If they are compared to HMM/DNN systems, which additionally leverage a large corpus of text-only data and expert-crafted lexica, the differences in modeling cannot be disentangled from differences in data. We propose an experimental setup, where only transcribed speech is used to train both model types. To highlight the difference that text-only data can make, we use Finnish, where an expert-crafted lexicon is not needed. With 1500h equal data, we find that both ASR paradigms perform similarly, but adding text data quickly improves the HMM/DNN system. On a smaller 160h subset we find that HMM/DNN models outperform AED models.

KW - Attention-based Encoder-Decoder

KW - Equal data

KW - HMM/DNN

UR - http://www.scopus.com/inward/record.url?scp=85116359931&partnerID=8YFLogxK

U2 - 10.1007/978-3-030-87802-3_54

DO - 10.1007/978-3-030-87802-3_54

M3 - Conference article in proceedings

AN - SCOPUS:85116359931

SN - 9783030878016

T3 - Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)

SP - 602

EP - 613

BT - Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings

A2 - Karpov, Alexey

A2 - Potapova, Rodmonga

PB - Springer

T2 - International Conference on Speech and Computer

Y2 - 27 September 2021 through 30 September 2021

ER -

Rouhe A, Van Camp A, Singh M, Van Hamme H, Kurimo M. An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR. In Karpov A, Potapova R, editors, Speech and Computer - 23rd International Conference, SPECOM 2021, Proceedings. Springer. 2021. p. 602-613. (Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)). doi: 10.1007/978-3-030-87802-3_54

An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

MeMAD: Methods for Managing Audiovisual Data: Combining Automatic Efficiency with Human Accuracy

Science-IT

Cite this

An Equal Data Setting for Attention-Based Encoder-Decoder and HMM/DNN Models: A Case Study in Finnish ASR

Abstract

Publication series

Conference

Keywords

Access to Document

Other files and links

Fingerprint

Projects

MeMAD: Methods for Managing Audiovisual Data: Combining Automatic Efficiency with Human Accuracy

Equipment

Science-IT

Cite this