Language Recognition for Dialects and Closely Related Languages - - PowerPoint PPT Presentation

1 / 3 – Language Recognition for Dialects and Closely Related Languages (G. Gelly, et al., LIMSI-Vocapia)

Language Recognition for Dialects and Closely Related Languages –

NIST OpenLRE 2015

• G. Gelly, J.L. Gauvain, L. Lamel, A. Laurent,

V.B. Le, A. Messaoudi

2 / 3 – Language Recognition for Dialects and Closely Related Languages (G. Gelly, et al., LIMSI-Vocapia)

Pitch

Main objectives Being able to identify the dialect when given excerpts of spoken conversations Benchmark different LID approaches in preparation for NIST LRE15 Techniques studied Phonotactic I-vectors LSTM-RNN Lexical

3 / 3 – Language Recognition for Dialects and Closely Related Languages (G. Gelly, et al., LIMSI-Vocapia)

Pitch

Main results LSTM-RNN can lead to a lower LER than I-vectors Phonotactic is the most robust method Phonotactic and LSTM-RNN fusion can lead to an important LER reduction

speech duration in seconds [0;5] ]5;10] ]10;20] ]20;30] > 30 all average LER 5 10 15 20 25 30 35 40 45 50 PHO RNN IVC LEX RNN+PHO RNN+PHO+IVC RNN+PHO+IVC+LEX target dialect ara-arz ara-acm ara-apc ara-ary ara-arb zho-yue zho-cmn zho-cdo zho-wuu eng-gbr eng-usg eng-sas fre-waf fre-hat qsl-pol qsl-rus spa-car spa-eur spa-lac por-brz system output ara-arz ara-acm ara-apc ara-ary ara-arb zho-yue zho-cmn zho-cdo zho-wuu eng-gbr eng-usg eng-sas fre-waf fre-hat qsl-pol qsl-rus spa-car spa-eur spa-lac por-brz