The degree of impact of environmental noise over phonemes is not uniform since it is dependent on their distinct acoustical properties. The criterion to decrease the impact of noise on speech varies across phonemes. The objective of this study is to suppress noise selectively from speech based on the knowledge of the phonemes, thereby producing an overall speech signal with improved speech quality.; A ROVER based enhancement framework is proposed in this study which employs Auto-LSP, a constrained iterative speech enhancement algorithm, as its baseline system. There are three significant contributions made from this study. First, the new algorithm addresses the issue of dependency of terminating iteration in Auto-LSP by generating a small set of enhanced utterances for every noisy utterance and selecting the best segments from this set using a hard decision scheme based on phoneme class specific constraints. Second, a soft decision method is also formulated to alleviate the effect of errors made in hard decisions. Finally, auditory masking threshold based enhancement technique is integrated into the hard and soft decision methods. The proposed enhancement algorithms are evaluated over the TIMIT speech corpus, using objective quality assessment tests based on the Itakura-Saito distortion, segmental SNR, and PESQ metrics. Results demonstrate that the proposed algorithms are more effective in achieving improved levels of speech quality for almost all phoneme classes and noise types considered in this study.
展开▼
