Publication Details
Automatic Speech Recognition and Topic Identification for Almost-Zero-Resource Languages
Liu Chunxi
ONDEL YANG, L.
HARMAN, C.
MANOHAR, V.
Trmal Jan, Ing., Ph.D.
HUANG, Z.
Dehak Najim
Khudanpur Sanjeev
Universal acoustic models, topic identification,cross-language information retrieval, transfer learning, lowresourcespeech recognition
Automatic speech recognition (ASR) systems often need to bedeveloped for extremely low-resource languages to serve endusessuch as audio content categorization and search. Whileuniversal phone recognition is natural to consider when no transcribedspeech is available to train an ASR system in a language,adapting universal phone models using very small amounts(minutes rather than hours) of transcribed speech also needs tobe studied, particularly with state-of-the-art DNN-based acousticmodels. The DARPA LORELEI program provides a frameworkfor such very-low-resource ASR studies, and provides anextrinsic metric for evaluating ASR performance in a humanitarianassistance, disaster relief setting. This paper presentsour Kaldi-based systems for the program, which employ a universalphone modeling approach to ASR, and describes recipesfor very rapid adaptation of this universal ASR system. Theresults we obtain significantly outperform results obtained bymany competing approaches on the NIST LoReHLT 2017 Evaluationdatasets
@inproceedings{BUT163405,
author="WIESNER, M. and LIU, C. and ONDEL YANG, L. and HARMAN, C. and MANOHAR, V. and TRMAL, J. and HUANG, Z. and DEHAK, N. and KHUDANPUR, S.",
title="Automatic Speech Recognition and Topic Identification for Almost-Zero-Resource Languages",
booktitle="Proceedings of Interspeech",
year="2018",
journal="Proceedings of Interspeech",
volume="2018",
number="9",
pages="2052--2056",
publisher="International Speech Communication Association",
address="Hyderabad",
doi="10.21437/Interspeech.2018-1836",
issn="1990-9772",
url="https://www.isca-speech.org/archive/Interspeech_2018/abstracts/1836.html"
}