Implementation of a Large Vocabulary Continuous Speech Recognition System for Brazilian Portuguese

  • Rafael Teruszkin
  • Fernando Gil Vianna Resende Junior

Abstract

This work presents the implementation of a large vocabulary speech recognition system for Brazilian Portuguese. The implemented system uses tools available on HTK and ATK toolkits. Tests were conducted in order to check the correlation on the context of continuous speech recognition among the following variables: word recognition rate, perplexity, distinct language models, computational complexity and vocabulary size. A speech database was used to train the stochastic acoustic models based on continuous HMMs, and a textual database was developed to train language models based on n-grams. Vocabularies ranging between 3.528 and 60.000 words were tested. The best accuracy rate obtained with a dictionary size of 3.528 words was 90% when recognizing sentences with 9 to 12 words, and 81% with 60.0000 words, both of them being speaker dependent, with perplexities ranging between 250 and 350, and processing times less than one minute per sentence.
Published
18-06-2015
How to Cite
Teruszkin, R., & Gil Vianna Resende Junior, F. (2015). Implementation of a Large Vocabulary Continuous Speech Recognition System for Brazilian Portuguese. Journal of Communication and Information Systems, 21(3). https://doi.org/10.14209/jcis.2006.18
Section
Regular Papers