This document discusses improving the word accuracy of an automatic speech recognition (ASR) system for the Telugu language. It analyzes the substitution errors in the system using two different lexical models - one based on stress-timed English phonemes (CMU lexicon) and one handcrafted lexicon for syllable-timed Telugu (UOH lexicon). The UOH lexicon improves word accuracy by 20-30% compared to the CMU lexicon by better modeling the phonetic characteristics of Telugu. The paper also examines the effect of gender, accents, and non-native speakers on substitution errors and the resulting confusion matrices provide insight into the most commonly substituted phonemes.