Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

MGFF-TDNN: A Multi-Granularity Feature Fusion TDNN Model with Depth-Wise Separable Module for Speaker Verification

Ya Li, Bin Zhou, Bo Hu

TL;DR

Addresses speaker verification by integrating a 2D depthwise separable front-end with a Multi-Granularity TDNN (M-TDNN) to fuse local and global context. The front-end captures time-frequency patterns, while M-TDNN combines TDNN-derived context with phoneme-level pooling and SE-based fusion to produce robust embeddings. On VoxCeleb, the method achieves competitive $EER$ and $minDCF$ with fewer parameters and FLOPs than strong baselines, demonstrating practical efficiency. This approach offers a scalable, real-time capable solution for SV with strong discrimination across short and long utterances.

Abstract

In speaker verification, traditional models often emphasize modeling long-term contextual features to capture global speaker characteristics. However, this approach can neglect fine-grained voiceprint information, which contains highly discriminative features essential for robust speaker embeddings. This paper introduces a novel model architecture, termed MGFF-TDNN, based on multi-granularity feature fusion. The MGFF-TDNN leverages a two-dimensional depth-wise separable convolution module, enhanced with local feature modeling, as a front-end feature extractor to effectively capture time-frequency domain features. To achieve comprehensive multi-granularity feature fusion, we propose the M-TDNN structure, which integrates global contextual modeling with fine-grained feature extraction by combining time-delay neural networks and phoneme-level feature pooling. Experiments on the VoxCeleb dataset demonstrate that the MGFF-TDNN achieves outstanding performance in speaker verification while remaining efficient in terms of parameters and computational resources.

MGFF-TDNN: A Multi-Granularity Feature Fusion TDNN Model with Depth-Wise Separable Module for Speaker Verification

TL;DR

Addresses speaker verification by integrating a 2D depthwise separable front-end with a Multi-Granularity TDNN (M-TDNN) to fuse local and global context. The front-end captures time-frequency patterns, while M-TDNN combines TDNN-derived context with phoneme-level pooling and SE-based fusion to produce robust embeddings. On VoxCeleb, the method achieves competitive and with fewer parameters and FLOPs than strong baselines, demonstrating practical efficiency. This approach offers a scalable, real-time capable solution for SV with strong discrimination across short and long utterances.

Abstract

In speaker verification, traditional models often emphasize modeling long-term contextual features to capture global speaker characteristics. However, this approach can neglect fine-grained voiceprint information, which contains highly discriminative features essential for robust speaker embeddings. This paper introduces a novel model architecture, termed MGFF-TDNN, based on multi-granularity feature fusion. The MGFF-TDNN leverages a two-dimensional depth-wise separable convolution module, enhanced with local feature modeling, as a front-end feature extractor to effectively capture time-frequency domain features. To achieve comprehensive multi-granularity feature fusion, we propose the M-TDNN structure, which integrates global contextual modeling with fine-grained feature extraction by combining time-delay neural networks and phoneme-level feature pooling. Experiments on the VoxCeleb dataset demonstrate that the MGFF-TDNN achieves outstanding performance in speaker verification while remaining efficient in terms of parameters and computational resources.
Paper Structure (10 sections, 5 equations, 3 figures, 4 tables)

This paper contains 10 sections, 5 equations, 3 figures, 4 tables.

Table of Contents

  1. Introduction
  2. System description
  3. DSM block
  4. M-TDNN block
  5. Experiments and analysis
  6. Datasets and evaluation metrics
  7. Implementation details
  8. Results and analysis
  9. Complexity analysis
  10. Conclusion

Figures (3)

  • Figure 1: Overview of the proposed MGFF-TDNN architecture.
  • Figure 2: Illustration of depthwise separable residual block.
  • Figure 3: The t-SNE visualization depicts the extracted embeddings of five speakers. These 5-second speaker embeddings are derived from ECAPA-TDNN, Res2Net, and MGFF-TDNN models.