Noise robustness using different acoustic units

This paper presents an evaluation of the use of different acoustic units for automatic speech recognition (ASR). Comparative experiments have indicated that the use of syllables as acoustic units leads to an improvement in the recognition performance of HMM-based ASR systems in noisy environments. The Hidden Markov Model Toolkit (HTK) was used throughout our experiments to test the use of the syllables for noisy ASR. A series of experiments on speaker-independent continuous-speech recognition have been carried out using subsets of the noisy speech corpus AURORA. The obtained results show that syllable-based recognition outperformed word-based recognition for a wide range of SNRs varying from 20 dB to -5 dB. The use of syllables did not only improve the performance of the ASR process in noisy environments, but also it limited the complexity of the computation (and consequently the running time) of the recognition process. This is due to the limited number of the syllables that has been used for the ASR compared to the number of words that represents the vocabulary of AURORA.