Multilingual shifting deep bottleneck features for low-resource ASR

Author

Quoc Bao Nguyen ; Gehring, Jonas ; Muller, Mathias ; Stuker, Sebastian ; Waibel, Alex

Author_Institution

Karlsruhe Inst. of Technol., Karlsruhe, Germany

fYear

2014

fDate

4-9 May 2014

Firstpage

5607

Lastpage

5611

Abstract

In this work, we propose a deep bottleneck feature architecture that is able to leverage data from multiple languages. We also show that tonal features are helpful for non-tonal languages. Evaluations are performed on a low-resource conversational telephone speech transcription task in Bengali, while additional data for DBNF training is provided in Assamese, Pashto, Tagalog, Turkish, and Vietnamese. We obtain relative reductions of up to 17.3% and 9.4% WER over mono-lingual GMMs and DBNFs, respectively.

Keywords

Gaussian processes; feature extraction; mixture models; natural language processing; speech recognition; Assamese; Bengali; DBNF training; Pashto; Tagalog; Turkish; Vietnamese; WER; low-resource ASR; low-resource conversational telephone speech transcription task; monolingual GMM; multilingual shifting deep bottleneck features; nontonal languages; tonal features; Feature extraction; Mel frequency cepstral coefficient; Speech; Speech recognition; Standards; Training; Deep Neural Networks; Low-Resource ASR; Multilingual Deep bottleneck features;

fLanguage

English

Publisher

ieee

Conference_Titel

Acoustics, Speech and Signal Processing (ICASSP), 2014 IEEE International Conference on

Conference_Location

Florence

Type

conf

DOI

10.1109/ICASSP.2014.6854676

Filename

6854676