Sphinx-Based Evaluation of Efficient Acoustic Modeling Parameters for LibriSpeech Corpus

Abstract

In this paper we are assessing the efficient parameters i.e., the number of senones and number of gaussian densities for a well-known audiobook corpus "LibriSpeech" based Automatic Speech Recognition System (ASR) using the open-source tool Sphinx. Sphinx is a Hidden Markov Model (HMM) based offline large vocabulary language and speaker independent continuous ASR system with a support for low-resource handheld devices. We have trained the acoustic model by varying the parameters and examined the quality of the models using Word Error Rate (WER). The best achieved WER of the model is observed as 9.5% with 2000 senones and 64 gaussian distributions.