A Chinese Speech Recognition System Based on Articulatory Features

In this paper, we build a Chinese speech recognition system based on acoustic features and articulatory features (AFs). We modify the original Chinese phonemes and define several language-independent AF groups based on International Phonetic Alphabet (IPA) and linguistic knowledge. To eliminate the need to force alignment information and simplify training, a technique using connectionist temporal classication (CTC) criterion is adopted. The Mandarin speech recognition system is built by the joint features which are concatenated with AFs and MFCC. The results show that using of AFs can improve the performance of ASR task on THCHS-30 30 hours data set.