TinyCAM: A Lightweight and Discriminative Network for Tibetan Speaker Verification

Speaker verification for minority languages faces considerable challenges, primarily due to the scarcity of training data and the structural complexities inherent to these languages. In particular, Tibetan speech is characterized by its rich plosive inventory and strong rhythmic patterns, which pose unique difficulties for robust speaker modeling. To address these issues, this paper proposes TinyCAM, an efficient and lightweight speaker verification framework tailored for such acoustic properties. Built upon the CAM++ architecture, TinyCAM introduces three structural enhancements aimed at improving feature representation and computational efficiency. First, the framework incorporates WDS-ResBlock, which fuses wavelet convolution with depthwise separable convolution to enable more effective multi-scale processing of local time-frequency details. Second, it employs a streamlined GS-TDNN module, which utilizes grouped and pointwise convolutions to capture diverse feature types and temporal dynamics with reduced complexity. Third, TinyCAM integrates the SLRT Layer, which combines low-rank compression with sparse constraints to minimize channel redundancy while enhancing information extraction. Extensive experiments and ablation studies conducted on a Tibetan speech dataset demonstrate that these architectural innovations not only improve recognition accuracy but also significantly reduce model size and computational overhead. Specifically, TinyCAM achieves an Equal Error Rate (EER) of 5.9905% and a minimum Detection Cost Function (minDCF) of 0.8287, with only 6.67 million parameters and 1.38 GFLOPs, marking reductions of 26.6% and 11.7 %, respectively, compared to the original model. These results highlight TinyCAM's strong potential for practical deployment in resource-constrained scenarios, particularly in applications involving minority languages.

• Publication year

  2025

• Venue

  2025 International Conference on Algorithms, Data Mining, and Information Technology (ADMIT)

• Publication date

  2025-10-24

• Fields of study

  Not labeled

• Identifiers
  DOI 10.1109/ADMIT67050.2025.11336995
• External record

  Open on Semantic Scholar

• Source metadata

  Semantic Scholar

• No claims are published for this paper.

• No concepts are published for this paper.

• An Enhanced Res2Net with Local and Global Feature Fusion for Speaker Verification

  2023cited by this paper
• Depth-First Neural Architecture With Attentive Feature Fusion for Efficient Speaker Verification

  2023cited by this paper
• CAM++: A Fast and Efficient Network for Speaker Verification Using Context-Aware Masking

  2023cited by this paper
• A Survey of Speaker Recognition: Fundamental Theories, Recognition Methods and Opportunities

  2021cited by this paper
• Densely Connected Time Delay Neural Network for Speaker Verification

  2020cited by this paper
• ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification

  2020cited by this paper
• X-Vectors: Robust DNN Embeddings for Speaker Recognition

  2018cited by this paper
• Exploring the Encoding Layer and Loss Function in End-to-End Speaker and Language Recognition System

  2018cited by this paper
• Semi-Orthogonal Low-Rank Matrix Factorization for Deep Neural Networks

  2018cited by this paper
• Deep Residual Learning for Image Recognition

  2015cited by this paper
• Vector Quantization Approach for Speaker Recognition

  2013cited by this paper
• Front-End Factor Analysis for Speaker Verification

  2011cited by this paper
• A Study of Interspeaker Variability in Speaker Verification

  2008cited by this paper
• Support vector machines for speaker and language recognition

  2006cited by this paper
• Speaker Verification Using Adapted Gaussian Mixture Models

  2000cited by this paper
• Robust text-independent speaker identification using Gaussian mixture speaker models

  1995cited by this paper
• Cepstral analysis technique for automatic speaker verification

  1981cited by this paper
• Dynamic programming algorithm optimization for spoken word recognition

  1978cited by this paper

• No citing papers are available for this paper.