一种基于注意力机制的语音情感识别算法研究

Research on Speech Emotion Recognition Method Based on Attention Mechanism

随着语音识别和自然语言处理技术的成熟,智能语音技术逐步落地到各行各业,但目前智能语音产品缺乏情感的交互,未实现真正的智能。为了提升智能语音产品的拟人性,学术界对语音情感识别的研究热度越来越高,但是多局限在使用传统特征工程以及通用深度学习模型上。在应用深度学习模型的基础上引入注意力机制对语音情感识别展开研究,选用语谱图作为输入特征,并对CNN输出的特征进行通道维度的注意力关注,分析注意力计算过程中不同池化方式对识别结果的影响,并引入残差,提升模型表现力,最终UA实现了2.83%的提升。

With the maturity of speech recognition and natural language processing technology,intelligent speech technology is gradually applied to all walks of life.However,at present,intelligent speech products lack emotional interaction and do not realize real intelligence.In order to improve the personification of intelligent speech products,the academic research on speech emotion recognition is becoming more and more popular,but it is mostly limited to the use of traditional feature engineering and general deep learning model.Based on the application of the deep learning model,this paper introduces the attention mechanism to study speech emotion recognition,selects the spectrogram as the input feature,pays attention to the features output by CNN in the channel dimension,analyzes the impact of different pooling methods on the recognition results in the process of attention calculation,and introduces the residual to improve the expressiveness of the model,Finally,UA achieved a 2.83%improvement.