Non-Parametric Bayesian Subspace Models for Acoustic Unit Discovery

Non-Parametric Bayesian Subspace Models for Acoustic Unit Discovery

WoS Article

This work investigates subspace non-parametricmodels for the task of learning a set of acoustic units fromunlabeledspeech recordings. We constrain the base-measure of a Dirichlet-Process mixture with a phonetic subspaceestimated from othersource languagesto build an educated prior, thereby forcing thelearned acoustic units to resemble phones of known source languages.Two types of models are proposed: (i) the Subspace HMM(SHMM) which assumes that the phonetic subspace is the same forevery language, (ii) the Hierarchical-Subspace HMM (H-SHMM)which relaxes this assumption and allows to have a languagespecificsubspace estimated on the unlabeled target data. Thesemodels are applied on 3 languages: English, Yoruba and Mboshiand they are compared with various competitive acoustic unitsdiscovery baselines. Experimental results show that both subspacemodels outperform other systems in terms of clustering quality andsegmentation accuracy. Moreover, we observe that the H-SHMMprovides results superior to the SHMM supporting the idea thatlanguage-specific priors are preferable to language-agnostic priorsfor acoustic unit discovery.

This work investigates subspace non-parametricmodels for the task of learning a set of acoustic units fromunlabeledspeech recordings. We constrain the base-measure of a Dirichlet-Process mixture with a phonetic subspaceestimated from othersource languagesto build an educated prior, thereby forcing thelearned acoustic units to resemble phones of known source languages.Two types of models are proposed: (i) the Subspace HMM(SHMM) which assumes that the phonetic subspace is the same forevery language, (ii) the Hierarchical-Subspace HMM (H-SHMM)which relaxes this assumption and allows to have a languagespecificsubspace estimated on the unlabeled target data. Thesemodels are applied on 3 languages: English, Yoruba and Mboshiand they are compared with various competitive acoustic unitsdiscovery baselines. Experimental results show that both subspacemodels outperform other systems in terms of clustering quality andsegmentation accuracy. Moreover, we observe that the H-SHMMprovides results superior to the SHMM supporting the idea thatlanguage-specific priors are preferable to language-agnostic priorsfor acoustic unit discovery.

Unsupervised learning, non- parametricBayesian models, acoustic unit discovery

Unsupervised learning, non- parametricBayesian models, acoustic unit discovery

ONDEL YANG, L.; YUSUF, B.; BURGET, L.; SARAÇLAR, M.

2023

03.05.2022

2329-9290

IEEE-ACM Transactions on Audio Speech and Language Processing

30

5

United States of America

1902

1917

16

https://ieeexplore.ieee.org/document/9767690

ondel_ieee_acm-tslp2022_Non-Parametric_Bayesian_Subspace_Models_for_Acoustic_Unit_Discovery