端到端维吾尔语语音识别研究

近几年来,基于端到端模型的语音识别系统因其相较于传统混合模型的结构简洁性和易于训练性而得到广泛的应用,并在汉语和英语等大语种上取得了显著的效果.本文将自注意力机制和链接时序分类损失代价函数相结合,将这种端到端模型应用到维吾尔语语音识别上.考虑到维吾尔语属于典型的黏着语,其丰富的构词形式使得维吾尔语的词汇量异常庞大,本文引入字节对编码算法进行建模单元的生成,从而获得合适的端到端建模输出单元.在King-ASR450维吾尔语数据集上,提出的算法明显优于基于隐马尔可夫模型的经典混合系统和基于双向长短时记忆网络的端到端模型,最终识别词准确率为91.35%.

Research on End-to-end Speech Recognition System for Uyghur

Compared with the conventional hybrid models,the end-to-end frameworks have recently been widely used in the automatic speech recognition(ASR)fields for their simple structure and ease of training,and have achieved remarkable results in large languages such as Chinese and English.In this paper,the end-to-end model which integrates self-attention mechanism and Connectionist Temporal Classification(CTC)loss function is applied to Uyghur speech recognition.Uyghur is a typical adhesive language with extremely large vocabulary.This paper introduces Byte Pair Encoding(BPE)to generate modeling units for CTC output layer.Experiments are carried out on King-ASR450 Uyghur corpus,the proposed methods can achieve better performance than the conventional hybrid system based on Hidden Markov Model and the end-to-end model based on Bi-directional long-short memory network,and we can final obtain a 91.35%word accuracy in this corpus.