Cross-lingual articulation information transfer with progressive networks for speech recognition

A system for the lateral transfer of information from recurrent end-to-end neural networks recognizing articulatory feature classes to similarly structured networks recognizing phone tokens is proposed. The system uses connections between recurrent layers at the same depth between feature detectors pre-trained on a base language and a phone recognizer for different target languages, inspired primarily by the progressive neural network scheme. While initial experiments using detectors for four articulatory feature classes—consonant place, consonant manner, vowel height, and vowel backness—trained on Bengali speech, attached to phone recognizers for four other Asian languages (Javanese, Nepali, Sinhalese, Sundanese), do not currently suggest consistent performance improvements across different low-resource settings for target languages, irrespective of their genealogic or phonological relatedness to Bengali, they do suggest the need for further trials with different language sets, altered data sources and data configurations, and slightly altered network setups.

Name thesis:degree_name

M.S.

Level thesis:degree_level

Thesis

Discipline thesis:degree_discipline

Electrical & Computer Engr

Grantor

University of Illinois at Urbana-Champaign

Year dc:date

2022

Author dc:creator

• Morshed, Mahir

Contributors dc:contributor

• Hasegawa-Johnson, Mark A

• articulatory feature detection
• transfer learning
• progressive neural networks

Statement dc:rights

• Copyright 2022 Mahir Morshed

Language dc:language

en, eng

Handle dc:identifier

https://hdl.handle.net/2142/115607

OAI record oai:identifier

https://www.ideals.illinois.edu/oai-pmh?verb=GetRecord&metadataPrefix=etdms&identifier=oai%3Awww.ideals.illinois.edu%3A2142%2F115607