Header menu link for other important links
X
Addressing data sparsity in DNN acoustic modeling
Published in Institute of Electrical and Electronics Engineers Inc.
2017
Abstract
This paper presents our work on developing acoustic models using deep neural networks (DNN) for low resource languages. This is considered one of the challenging problems in automatic speech recognition (ASR) as DNNs need large amount of data for building efficient models. The techniques explored in this approach use a common idea of transferring knowledge from models of high resource language to a low resource language. These methods include: (i) cross-lingual approach of building a common DNN model with data from a low resource language pooled with a linguistically close high resource language, (ii) transfer learning approach by adapting the top one or more layers of a pooled DNN model with low resource data, (iii) multitask learning approach by treating each language as a separate task for a common model trained over multiple related tasks simultaneously, which alleviates over-fitting problems. The multi-task method was found to be the most effective of all. Experiments were done on Kannada (low resource) and Telugu data and 46% relative improvement is observed in multi-task framework over its mono-lingual DNN. © 2017 IEEE.
About the journal
JournalData powered by Typeset2017 23rd National Conference on Communications, NCC 2017
PublisherData powered by TypesetInstitute of Electrical and Electronics Engineers Inc.
Open AccessNo
Concepts (12)
  •  related image
    Deep learning
  •  related image
    Multitasking
  •  related image
    Speech recognition
  •  related image
    Automatic speech recognition
  •  related image
    CROSS-LINGUAL APPROACHES
  •  related image
    DATA POOLING
  •  related image
    Low resource languages
  •  related image
    LOW-RESOURCE
  •  related image
    Multitask learning
  •  related image
    OVER FITTING PROBLEM
  •  related image
    Transfer learning
  •  related image
    Deep neural networks