Muliya Pritibahen Jagjivanbhai, Ashwin Patni · International Journal of Electronics and Communication Engineering 2026 · 2026
DOI: 10.14445/23488549/ijece-v13i9p113
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
Speaker identification in multiple languages has received increasing attention in speech processing because of the growing use of voice-based security applications in multilingual settings. Traditional speaker identification systems have encountered difficulties ensuring reliable performance in bilingual and cross-language speech scenarios due to their concentration on local acoustic models and neglecting long-range contextual information. This research presents a novel hybrid architecture referred to as ECAPA-TF, which combines ECAPA-TDNN-based local acoustic modelling with transformer-based global contextual modelling for multilingual speaker identification. The proposed model is trained using the Native and Indian Speakers of Indian Phonetics (NISP) database consisting of 345 bilingual speakers of India, who are proficient in both English and native Indian languages, including Malayalam, Hindi, Kannada, Telugu and Tamil. The first step involves extracting Mel-Frequency Cepstral Coefficient (MFCC) feature vectors (40-dimensional) from the speech signal and normalizing the data prior to applying ECAPA residual blocks equipped with squeeze-excitation attention. Subsequently, the Transformer self-attention mechanism is implemented to identify long-range temporal and contextual dependencies in speech data. Attention pooling and Additive Angular Margin SoftMax (AAM-SoftMax) are also leveraged to create discriminative speaker embeddings. Performance is evaluated using Accuracy, recall, precision, confusion matrix, F1-score and ROC-AUC. The developed ECAPA-TF model achieved an accuracy of 92.93%, precision of 93.64%, recall of 93.02%, macro F1-score of 92.94%, and an AUC of 0.965. Under the same NISP experimental protocol, the proposed model outperformed the internally implemented Deep 1D-CNN, CNN-LSTM, and ECAPA baseline models. Compared with the ECAPA baseline, ECAPA-TF improved accuracy by 1.24 percentage points and macro F1-score by 1.36 percentage points. The confusion matrix and per-speaker evaluation further validate the effectiveness and reliability of the proposed approach under various multilingual speech scenarios.
No comments yet — start the discussion below.