EML-CNN reports 96%+ speaker-identification accuracy across three speech benchmarks

A proposed Enhanced Multi-Layer Convolutional Neural Network, or EML-CNN, achieved speaker-identification accuracies of 99.4%, 96.1% and 96.4% on three benchmark datasets designed to test emotional speech, dialect diversity and noisy multi-speaker audio. The results, reported in an accepted research article in Complex & Intelligent Systems, place all three scores above 96% and show how the model performed across distinct acoustic conditions.

The framework combines dense convolutional blocks, gated recurrent units and an attention mechanism. The authors report that these components were integrated to produce highly discriminative speaker representations and improve performance under noisy, emotional and cross-domain conditions. In the study’s comparisons, EML-CNN consistently outperformed both traditional approaches and recent state-of-the-art methods. The work is presented as a model evaluation for automated speaker identification, with testing conducted on IEMOCAP, THUYG-20 and VoxCeleb rather than as a report from an operational access-control or forensic system.

Each dataset contributed a particular challenge to the evaluation. IEMOCAP was used for the emotional-speech setting, where the reported accuracy reached 99.4%. THUYG-20, described by the researchers as representing dialect diversity, produced an accuracy of 96.1%. VoxCeleb supplied noisy, multi-speaker conditions, on which the model reached 96.4%. Together, the three tests indicate that the proposed architecture retained strong measured performance across emotional delivery, dialect variation and noisy recordings, although the scores describe results on the named benchmarks and should not be treated as universal estimates of performance in every recording environment.

The research also included audio augmentation intended to strengthen robustness. The reported augmentations were additive Gaussian noise at signal-to-noise ratios from 10 to 30 decibels, temporal shifting, frequency stretching and pitch variation. Along with the architectural changes, these perturbations formed part of the study’s training strategy. The authors report that EML-CNN outperformed the comparison methods under the experimental setup, supporting the conclusion that the combined model was effective on the three evaluated datasets. Further evidence would be needed to determine how the same approach performs under deployment conditions beyond those represented by the benchmarks.

Subject of Research: Benchmark evaluation of an EML-CNN for robust speaker identification in emotional, dialectal and noisy speech

Article Title: Noise-robust speaker recognition using an enhanced multi-layer CNN with attention and recurrent modeling

Article References: Amjad, A., Fan, S.-Y., Chang, H.-T., Lin, Y.-W., & Tai, L.-C. (2026). Noise-robust speaker recognition using an enhanced multi-layer CNN with attention and recurrent modeling. Complex & Intelligent Systems. https://doi.org/10.1007/s40747-026-02543-6

Image Credits: AI Generated

DOI: 10.1007/s40747-026-02543-6

Keywords: speaker identification, speech processing, convolutional neural network, recurrent modeling, attention mechanism, audio noise, audio augmentation, emotional speech, dialect diversity, deep learning, Noise-robust, speaker