Dual supervised learning for non-native speech recognition

Kacper Pawel Radzikowski, Robert Nowak, Le Wang, Osamu Yoshie

研究成果: Article

抄録

Current automatic speech recognition (ASR) systems achieve over 90–95% accuracy, depending on the methodology applied and datasets used. However, the level of accuracy decreases significantly when the same ASR system is used by a non-native speaker of the language to be recognized. At the same time, the volume of labeled datasets of non-native speech samples is extremely limited both in size and in the number of existing languages. This problem makes it difficult to train or build sufficiently accurate ASR systems targeted at non-native speakers, which, consequently, calls for a different approach that would make use of vast amounts of large unlabeled datasets. In this paper, we address this issue by employing dual supervised learning (DSL) and reinforcement learning with policy gradient methodology. We tested DSL in a warm-start approach, with two models trained beforehand, and in a semi warm-start approach with only one of the two models pre-trained. The experiments were conducted on English language pronounced by Japanese and Polish speakers. The results of our experiments show that creating ASR systems with DSL can achieve an accuracy comparable to traditional methods, while simultaneously making use of unlabeled data, which obviously is much cheaper to obtain and comes in larger sizes.

元の言語English
記事番号3
ジャーナルEurasip Journal on Audio, Speech, and Music Processing
2019
発行部数1
DOI
出版物ステータスPublished - 2019 12 1

Fingerprint

Supervised learning
speech recognition
Speech recognition
learning
English language
methodology
Reinforcement learning
reinforcement
Experiments
gradients

ASJC Scopus subject areas

  • Acoustics and Ultrasonics
  • Electrical and Electronic Engineering

これを引用

@article{3478629dc72d4136bb900dcc528783d2,
title = "Dual supervised learning for non-native speech recognition",
abstract = "Current automatic speech recognition (ASR) systems achieve over 90–95{\%} accuracy, depending on the methodology applied and datasets used. However, the level of accuracy decreases significantly when the same ASR system is used by a non-native speaker of the language to be recognized. At the same time, the volume of labeled datasets of non-native speech samples is extremely limited both in size and in the number of existing languages. This problem makes it difficult to train or build sufficiently accurate ASR systems targeted at non-native speakers, which, consequently, calls for a different approach that would make use of vast amounts of large unlabeled datasets. In this paper, we address this issue by employing dual supervised learning (DSL) and reinforcement learning with policy gradient methodology. We tested DSL in a warm-start approach, with two models trained beforehand, and in a semi warm-start approach with only one of the two models pre-trained. The experiments were conducted on English language pronounced by Japanese and Polish speakers. The results of our experiments show that creating ASR systems with DSL can achieve an accuracy comparable to traditional methods, while simultaneously making use of unlabeled data, which obviously is much cheaper to obtain and comes in larger sizes.",
keywords = "Artificial intelligence, Deep learning, Dual supervised learning, Machine learning, Non-native speaker, Policy gradients, Reinforcement learning, Speech recognition",
author = "Radzikowski, {Kacper Pawel} and Robert Nowak and Le Wang and Osamu Yoshie",
year = "2019",
month = "12",
day = "1",
doi = "10.1186/s13636-018-0146-4",
language = "English",
volume = "2019",
journal = "Eurasip Journal on Audio, Speech, and Music Processing",
issn = "1687-4714",
publisher = "Springer Publishing Company",
number = "1",

}

TY - JOUR

T1 - Dual supervised learning for non-native speech recognition

AU - Radzikowski, Kacper Pawel

AU - Nowak, Robert

AU - Wang, Le

AU - Yoshie, Osamu

PY - 2019/12/1

Y1 - 2019/12/1

N2 - Current automatic speech recognition (ASR) systems achieve over 90–95% accuracy, depending on the methodology applied and datasets used. However, the level of accuracy decreases significantly when the same ASR system is used by a non-native speaker of the language to be recognized. At the same time, the volume of labeled datasets of non-native speech samples is extremely limited both in size and in the number of existing languages. This problem makes it difficult to train or build sufficiently accurate ASR systems targeted at non-native speakers, which, consequently, calls for a different approach that would make use of vast amounts of large unlabeled datasets. In this paper, we address this issue by employing dual supervised learning (DSL) and reinforcement learning with policy gradient methodology. We tested DSL in a warm-start approach, with two models trained beforehand, and in a semi warm-start approach with only one of the two models pre-trained. The experiments were conducted on English language pronounced by Japanese and Polish speakers. The results of our experiments show that creating ASR systems with DSL can achieve an accuracy comparable to traditional methods, while simultaneously making use of unlabeled data, which obviously is much cheaper to obtain and comes in larger sizes.

AB - Current automatic speech recognition (ASR) systems achieve over 90–95% accuracy, depending on the methodology applied and datasets used. However, the level of accuracy decreases significantly when the same ASR system is used by a non-native speaker of the language to be recognized. At the same time, the volume of labeled datasets of non-native speech samples is extremely limited both in size and in the number of existing languages. This problem makes it difficult to train or build sufficiently accurate ASR systems targeted at non-native speakers, which, consequently, calls for a different approach that would make use of vast amounts of large unlabeled datasets. In this paper, we address this issue by employing dual supervised learning (DSL) and reinforcement learning with policy gradient methodology. We tested DSL in a warm-start approach, with two models trained beforehand, and in a semi warm-start approach with only one of the two models pre-trained. The experiments were conducted on English language pronounced by Japanese and Polish speakers. The results of our experiments show that creating ASR systems with DSL can achieve an accuracy comparable to traditional methods, while simultaneously making use of unlabeled data, which obviously is much cheaper to obtain and comes in larger sizes.

KW - Artificial intelligence

KW - Deep learning

KW - Dual supervised learning

KW - Machine learning

KW - Non-native speaker

KW - Policy gradients

KW - Reinforcement learning

KW - Speech recognition

UR - http://www.scopus.com/inward/record.url?scp=85060134244&partnerID=8YFLogxK

UR - http://www.scopus.com/inward/citedby.url?scp=85060134244&partnerID=8YFLogxK

U2 - 10.1186/s13636-018-0146-4

DO - 10.1186/s13636-018-0146-4

M3 - Article

VL - 2019

JO - Eurasip Journal on Audio, Speech, and Music Processing

JF - Eurasip Journal on Audio, Speech, and Music Processing

SN - 1687-4714

IS - 1

M1 - 3

ER -