Hierarchical Phoneme Classification for Improved Speech Recognition

Speech recognition consists of converting input sound into a sequence phonemes, then finding text for the using language models. Therefore, phoneme classification performance is critical factor successful implementation speech system. However, correctly distinguishing phonemes with similar characteristics still challenging problem even state-of-the-art methods, and errors are hard to be recovered in subsequent processing steps. This paper proposes hierarchical clustering method exploit more suitable models different phonemes. The TIMIT database carefully analyzed confusion matrix from baseline model. Using automatic results, set optimized generated groups constructed integrated method. According results number experiments, proposed group improved over by 3%, 2.1%, 6.0%, 2.2% fricative, affricate, stop, nasal sounds, respectively. average accuracy was 69.5% 71.7% models, showing overall improvement.