基于神经形态电路的音频场景特征提取及识别技术

Audio feature extraction and context recognition based on neuromorphic circuit

传统的冯·诺依曼架构在处理语音等复杂信息时能效较低,神经形态电路更适合于语音等复杂信息的智能处理。常用的音频场景识别方式中的长时特征和短时特征都有其不足之处,卷积神经网络可通过训练提取适合后续分类任务的特征,在特征提取方面有更大的优势。针对四层的卷积神经网络的特征提取及分析方法在语谱图上进行了音频场景识别的研究,并验证了音频场景识别在神经形态电路—类脑计算芯片上的可实现性。

The traditional Von Neumann architecture has low energy efficiency while dealing with complex information. Neuromorphic circuits is more suitable for intelligent processing of complex information. The long and short time features commonly used in audio context recognition methods have shortcomings,deep neural network can be trained to extract features for subsequent classification task,and has greater advantages in the feature extraction. This paper used the feature extraction and analysis method through deep neural network to recognize audio context,and verified the implementation of the recognition on the neuromorphic circuit.