Accent Analysis for Mandarin Large Vocabulary Continuous Speech Recognition

摘要

This paper presents our work on accent issues in Mandarin large vocabulary continuous speech recognition. Across a vast region and a huge population, there are varieties of accented Mandarin spoken in China, which are mainly caused by speakers' dialects. What we want to address in this paper are two questions about Mandarin: whether accents affect speech recognition greatly; how we can solve the problem. For the first question, we focus on three types of mispronunciations as the dominant problems of accented Mandarin. We analyze their effects on speech recognition for each speaker. For the second question, we perform maximum likelihood linear regression (MLLR) adaptation for each speaker and then analyze the recognition results. Experimental results show that up to 45% of the accent related errors get corrected for accented speakers and there is no such improvement for standard speakers. Our experimental analysis and results support us to conclude that the accent is a serious problem in Mandarin speech recognition and the MLLR adaptation is effective in reducing the mismatch caused by accents.