Method, device and medium for constructing Chinese speech recognition model

Abstract

The application discloses a kind of construction methods, device and medium of Chinese speech recognition model, belong to speech recognition technical field.The method includes: obtaining the pinyin sequence corresponding to speech data, each pinyin is split into initial and final, according to pronunciation feature initial is divided into multiple initial groups, initial group is combined with final to obtain modeling unit, with modeling unit as recognition object training first recognition task, with initial as recognition object training second recognition task, first recognition task and second recognition task are jointly trained to obtain speech recognition model.The above scheme is used, the number of modeling unit classes is less than the number of full pinyin classes, the recognition granularity is greater than single phoneme, initial information is still identifiable when missing at fast speed, second recognition task confirms the specific initial in initial group, and the recognition accuracy of Chinese speech recognition model is improved.

Images