Scholarlink

<p>Speech emotion classification (SEC) systems increasingly rely on computationally intensive signal decomposition and feature extraction techniques, which pose scalability challenges when applied to large speech corpora. In this work, we propose an efficient speech emotion classification framework based on empirical wavelet transform (EWT)—driven feature extraction combined with Deep Neural Network (DNN) classifier. The proposed pipeline operates at the frame level, performing three-band EWT decomposition followed by feature computation (20 Mel-frequency cepstral coefficients (MFCC), 3 approximate entropy (AE), and 3 permutation entropy (PE) features per mode) resulting in a high computational load per speech utterance. To address this challenge, the feature extraction stage is reformulated as an embarrassingly parallel workload and executed on a shared-memory High Performance Computing (HPC) platform using Simple Linux Utility for Resource Management (SLURM)-managed batch jobs. Parallelization is performed at the speech-sample level, enabling independent processing of individual speech files without synchronization overhead. Next core level parallelism was achieved by executing independent cross-validation folds concurrently on separate cpu cores using HPC job array approach. Experiments conducted on EMODB, EMOVO, and TESS achieve emotion classification accuracies of 87.56 ± 1.15%, 86.12 ± 2.01%, and 99.75 ± 0.08%, respectively, under a speaker-dependent five-fold cross-validation setup. To assess generalization to unseen speakers, a strict speaker-independent Leave-One-Speaker-Out (LOSO) evaluation is additionally performed on EMODB, yielding an average accuracy of 75.65% ± 3.24%. For EMODB dataset the total wall-clock time for training is approximately 268&#xa0;s on CPU-based HPC nodes using core level parallel execution, demonstrating that the proposed framework is both computationally efficient and scalable for large-scale speech emotion recognition experiments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Emotion classification from speech using empirical wavelet transform derived spectral and entropy features with DNN classifier

  • 期刊论文

  • 发表日期：

    2026年6月25日

查看全文
  • Devi Prasad Pattnaik,
  • Bala Sai Srilatha Indira Dutt Vemuri

摘要

Speech emotion classification (SEC) systems increasingly rely on computationally intensive signal decomposition and feature extraction techniques, which pose scalability challenges when applied to large speech corpora. In this work, we propose an efficient speech emotion classification framework based on empirical wavelet transform (EWT)—driven feature extraction combined with Deep Neural Network (DNN) classifier. The proposed pipeline operates at the frame level, performing three-band EWT decomposition followed by feature computation (20 Mel-frequency cepstral coefficients (MFCC), 3 approximate entropy (AE), and 3 permutation entropy (PE) features per mode) resulting in a high computational load per speech utterance. To address this challenge, the feature extraction stage is reformulated as an embarrassingly parallel workload and executed on a shared-memory High Performance Computing (HPC) platform using Simple Linux Utility for Resource Management (SLURM)-managed batch jobs. Parallelization is performed at the speech-sample level, enabling independent processing of individual speech files without synchronization overhead. Next core level parallelism was achieved by executing independent cross-validation folds concurrently on separate cpu cores using HPC job array approach. Experiments conducted on EMODB, EMOVO, and TESS achieve emotion classification accuracies of 87.56 ± 1.15%, 86.12 ± 2.01%, and 99.75 ± 0.08%, respectively, under a speaker-dependent five-fold cross-validation setup. To assess generalization to unseen speakers, a strict speaker-independent Leave-One-Speaker-Out (LOSO) evaluation is additionally performed on EMODB, yielding an average accuracy of 75.65% ± 3.24%. For EMODB dataset the total wall-clock time for training is approximately 268 s on CPU-based HPC nodes using core level parallel execution, demonstrating that the proposed framework is both computationally efficient and scalable for large-scale speech emotion recognition experiments.