Understanding the Robustness in Phoneme Production Mechanism in English and Bengali

TRISHITA DHARA, Sirshapan Mitra, Suparnakanti Das, Sudip Kumar Naskar

Abstract: Hidden Markov Modeling (HMM) is one of the most popular methods for speech recognition due to its capability to exploit the inherent speech production mechanism. However, the presence of noise in speech makes it difficult to model the phonemes. The accuracy of different phoneme models varies due to their production mechanism & often does not behave uniformly. In this paper, we model English & Bengali phonemes using ergodic (EHMM) & left to right (L2R HMM) HMMs in different background noise scenarios, with an aim to understand the degradation pattern of the phoneme models across models & background noise. Babble & car noise were used at different SNR levels for validating modeling accuracy at different scenarios. Results show that a few phonemes behave in a more uniform manner compared to the other phonemes. This may be attributed to the robustness of the phonemes due to their structural cues in presence of different background noise.