Mandarin Speech Recognition Based on Chinese Syllable Similarity for Children

Deep learning has facilitated the advancement and wide application of speech recognition technology. While mainstream speech recognition models are usually trained with adult standard speech data, which do not take into account the particularities of special groups, which is not suitable for all groups. Compared with adult speech, child speech recognition is more challenging because there are greater differences in acoustic and linguistic features between speakers. Furthermore, children have their own unique pronunciation habits and pronunciation methods and it’s hard for them to clear the difference between such as level and oblique tones, front nasal and back nasal. In addition, the lack of adequate and appropriate children’s voice resources increases the challenge of designing robust end-to-end neural structures. Therefore, we propose a novel model for Chinese children’s speech recognition, the Chinese children syllable similarity recognition model, which is used to help distinguish and identify children’s ambiguous pronunciation, in order to achieve a better recognition effect. Since the speech recognition of young children is usually short words and phrase recognition, the existing recognition evaluation metrics are not suitable. In the paper, we design a sliding window matching judgment method, which greatly improves the accuracy of children’s speech recognition and achieves very good recognition effect. Our model achieves a judgment accuracy of 91.7% on our dataset, which is higher than other models.