Speech recognition system based on deep neural network acoustic modeling for low resourced language-Amharic

Abstract

In this paper automatic speech recognition is investigated using deep neural network (DNN) acoustic modeling method for Amharic language at syllabic acoustic units. In grapheme based database; graphemes/characters are basic units of lexicon and language model. A large portion of them represents syllables which are a combination of consonants and vowels (CV). Grapheme to phoneme (G2P) conversion was required to represent all text corpuses into CV phoneme representations via G2P conversion algorithm developed for this purpose. This algorithm used to develop syllable based pronunciation dictionary and language modeling which are vital for speech recognizer. DNN based acoustic model (AM) such as tanh-DNNs, tanh-fast-DNNs, p-norm-DNNs and p-norm-fast-DNNs are also explored with different hidden layers, hidden units and parameter settings. These DNN AMs are trained with morpheme based Amharic read speech in order to develop models. The recognition performance of our methods is evaluated in testing data and the reduced WER is achieved in p-norm-fast(p=2) DNN AMs.