军事知识服务平台
基于极端随机树加权模糊概念认知的语音情感识别
Speech Emotion Recognition Based on Weighted Fuzzy Concept Cognition Using Extremely Randomized Trees
摘要 语音情感识别(Speech Emotion Recognition,SER)在提升人机交互体验中扮演着关键角色.当前主流深度学习模型虽在 SER 任务上取得了优异精度,但普遍存在可解释性不足的问题.为突破这一局限,本文借鉴概念认知学习(Concept-cognitive Learning,CCL)思想,将形式概念分析(Formal Concept Analysis,FCA)与粒计算(Granular Computing,GrC)理论相结合,提出一种基于极端随机树的加权模糊概念认知分类模型(Extremely Ran-domized Trees based Weighted Fuzzy Concept-cognitive Learning,ERT-WFCCL).该模型采用大规模预训练语音模型 Whisper 作为特征提取器,从原始语音信号中获取高层语义表征,并通过多层感知机制(Multi-layer Percep-tron,MLP)将高维特征解耦为具有明确语义的低维属性向量,以构建形式背景;在分类阶段,利用极端随机树(Ex-tremely Randomized Trees,ERT)评估各属性对情感分类的贡献度,据此构造加权模糊概念空间,并在该概念空间中完成分类推理.该框架以感知到认知的路径为主线,实现了从连续声学信号到可解释概念空间的映射,为 SER任务提供了一种具备内生可解释性的建模思路.本文在多伦多情感语音集(Toronto Emotional Speech Set,TESS)和众包情感多模态演员数据集(Crowd-sourced Emotional Multimodal Actors Dataset,CREMA-D)上进行了实验.结果表明,ERT-WFCCL 在 TESS 上取得 99.60%的分类准确率和 99.50%的宏 F1 值,在更具挑战性的 CREMA-D上取得 85.71%的准确率和 85.81%的宏 F1 值,在多项指标上优于支持向量机(Support Vector Machine,SVM)、逻辑回归(Logistic Regression,LR)等对比方法.此外,增量学习实验验证了所提加权模糊概念认知更新策略能够在保持较低计算开销的前提下稳定更新概念空间,展现出较强的鲁棒性和可扩展能力.
Abstract:
Speech Emotion Recognition(SER)plays a key role in improving the naturalness and effectiveness of human-computer interaction by enabling machines to perceive and respond to the emotional states conveyed in human speech.In recent years,deep learning models such as convo-lutional neural networks and recurrent architectures have achieved high recognition accuracy on various SER benchmarks.However,these models generally operate as black boxes and suffer from insufficient interpretability,making it difficult to provide transparent and traceable explana-tions for their classification decisions.This interpretability limitation restricts the deployment of SER systems in trust-sensitive and safety-critical applications such as mental health monitoring,intelligent customer service,and affective computing.To overcome this limitation,this paper draws on the theory of Concept-cognitive Learning(CCL),which integrates Formal Concept Analysis(FCA)and Granular Computing(GrC)to model human-like cognitive processes of con-cept formation and classification.Based on this theoretical foundation,we propose a novel classi-fication framework termed Extremely Randomized Trees based Weighted Fuzzy Concept-cognitive Learning(ERT-WFCCL).The proposed framework comprises three main stages:feature extrac-tion,feature decoupling,and concept-cognitive classification.In the feature extraction stage,the model employs Whisper,a large-scale pre-trained speech model,as the backbone feature extrac-tor to obtain high-level semantic representations from raw audio waveforms.Whisper benefits from extensive multilingual and multitask supervised training,enabling it to capture rich acoustic and linguistic information relevant to emotional expression.In the feature decoupling stage,a Multi-layer Perceptron(MLP)based module is designed to transform the high-dimensional Whis-per features into low-dimensional attribute vectors with explicit semantic meaning.Each dimen-sion of the decoupled vector corresponds to a specific interpretable attribute,which enables the construction of a formal context as the standard input structure required by FCA.This decou-pling step is critical for bridging the gap between deep feature representations and concept-based reasoning.In the classification stage,the Extremely Randomized Trees(ERT)algorithm is ap-plied to evaluate the discriminative contribution of each decoupled attribute to emotion classifica-tion.Based on the resulting feature importance scores,a weighted fuzzy concept space is con-structed,in which each attribute is assigned a weight reflecting its relevance to the target emotion categories.Classification inference is then performed by navigating this weighted fuzzy concept space according to the Concept-cognitive Learning strategy.This design endows the model with inherent interpretability,as each classification decision can be traced back through the concept lattice to specific attributes and their weighted contributions.The overall framework follows a perception-to-cognition pathway,achieving a mapping from continuous acoustic signals to an in-terpretable concept space and providing an inherently interpretable modeling approach for SER tasks.Comprehensive experiments are conducted on two widely used public SER datasets:the Toronto Emotional Speech Set(TESS),which contains 2800 utterances covering seven emotion categories,and the Crowd-sourced Emotional Multimodal Actors Dataset(CREMA-D),which comprises 7442 utterances from 91 actors across six emotion categories with greater speaker and recording variability.The proposed ERT-WFCCL model achieves a classification accuracy of 99.60%and a macro F1 score of 99.50%on TESS,and an accuracy of 85.71%with a macro F1 score of 85.81%on the more challenging CREMA-D,outperforming comparison methods inclu-ding Support Vector Machine(SVM)and Logistic Regression(LR)across multiple evaluation metrics.Furthermore,incremental learning experiments verify that the proposed weighted fuzzy concept-cognitive update strategy can stably update the concept space with low computational overhead,demonstrating strong robustness and scalability.
刊 名 计算机学报 ISTICEIPKU
年,卷(期): 2026, 49(9)
分类号: TP18
机标分类号: TP311 R338.1 TP18
在线出版日期: 2026年10月8日