An End-to-End Speech Recognition System Based on Shared Encoder

Abstract

With the development of streaming media, automatic speech recognition (ASR) has been widely used in online education, live broadcast and other fields. However, for a better recognition effect in the real scenario, it is necessary to combine various technologies, such as front-end voice endpoint detection and back-end language model. In order to filter sensitive words in real scenarios, we require good online recognition and decoding methods. This paper presents an End-to-End speech recognition system, which unifies stream and non-stream speech recognition based on a shared encoder, and contains an additional CTC structure in the middle layer. Based on the monosyllable feature of mandarin, we calculate the probability distribution of syllables in the middle layer. The results show that our method is reliable for recognition in educational scenarios. We have achieved good results on aishell-l and audio in real scenarios provided by the company. At the same time, this system provides accurate syllable information to analyze sensitive words further.