基于Transformer的普通话语声识别模型位置编码选择

Transformer-based position coding selection of Mandarin speech recognition model

具有自注意机制的Transformer网络在语声识别研究领域渐渐得到广泛关注。该文围绕着将位置信息嵌入与语声特征相结合的方向,研究更加适合普通话语声识别模型的位置编码方法。实验结果得出,采用卷积编码的输入表示代替正弦位置编码,可以更好地融合语声特征上下文联系和相对位置信息,获得较好的识别效果。训练的语声识别系统是在Transformer模型基础上,比较4种不同的位置编码方法。结合3-gram语言模型,所提出的卷积位置编码方法,在中文语声数据集AISHELL-1上的误识率降低至8.16%。

The Transformer network with self-attention mechanism has gradually gained wide attention in the field of speech recognition research.This paper revolves around the direction of embedding location information and speech features,and studies the location coding method that is more suitable for Mandarin speech recognition model.The experimental results show that the input representation of convolutional coding instead of sinusoidal position coding can better integrate the contextual relationship of speech features and relative position information,and obtain better recognition results.The trained speech recognition system is based on the Transformer model and compares four different position coding methods.Combined with the 3-gram language model and the proposed convolutional position coding method,the word recognition error rate on the Chinese speech data set AISHELL-1 is reduced to 8.16%.