analiza aiqbench
This research introduces MECT, a speaker verification architecture that successfully integrates the Mixture-of-Experts mechanism into a CNN-Transformer backbone to achieve state-of-the-art performance with compact parameters. By optimizing block structures and exploring both dense and sparse routing strategies, the model maintains low computational complexity while enabling efficient streaming inference at 100ms chunks. The study demonstrates significant gains over baselines across diverse datasets like VoxCeleb1 and CN-Celeb, highlighting the viability of MoE for specialized audio AI tasks.
MECT: Mixture of Experts with CNN-Transformer Network for Speaker verification
arXiv:2609.24061v1 Announce Type: cross Abstract: In this paper, we propose MECT, a speaker verification model that integrates the Mixture-of-Experts (MoE) mechanism into a CNN-Transformer backbone with optimized block structure and stacking scheme. Specifically, we investigated four MoE variants that span utterance-level and frame-level granularity with dense and sparse routing strategies. The MoE mechanism proves to be effective over the baseline without MoE with only a small increase in parameters. We further scale MECT to a series of model sizes, all maintaining compact parameters and low computational complexity. In particular, MECT-B2 achieves state-of-the-art performance on VoxCeleb1 and delivers strong results on CN-Celeb, demonstrating its effectiveness across diverse datasets. In addition, we establish a streaming inference paradigm through causal retraining, which maintains strong performance at a chunk size of 100ms.
Abstract:In this paper, we propose MECT, a speaker verification model that integrates the Mixture-of-Experts (MoE) mechanism into a CNN-Transformer backbone with optimized block structure and stacking scheme. Specifically, we investigated four MoE variants that span utterance-level and frame-level granularity with dense and sparse routing strategies. The MoE mechanism proves to be effective over the baseline without MoE with only a small increase in parameters. We further scale MECT to a series of model sizes, all maintaining compact parameters and low computational complexity. In particular, MECT-B2 achieves state-of-the-art performance on VoxCeleb1 and delivers strong results on CN-Celeb, demonstrating its effectiveness across diverse datasets. In addition, we establish a streaming inference paradigm through causal retraining, which maintains strong performance at a chunk size of 100ms.
Sound (cs.SD); Artificial Intelligence (cs.AI); Audio and Speech Processing (eess.AS)
arXiv:2609.24061 [cs.SD]
(or arXiv:2609.24061v1 [cs.SD] for this version)
https://doi.org/10.48550/arXiv.2609.24061 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 224 cuvinte
Subiecte conexe
NVIDIA GTC Taipei at COMPUTEX: Live Updates on What’s Next in AI
NVIDIA · 21 mai 2026
The J-Space Debate, Agent Swarms, and Pacing Frontier AI - Digital Minds Newsletter #4
LessWrong · 18 sept. 2026
From Self-Supervised Speech Models to Mixture-of-Experts for Robust Anti-Spoofing
arXiv · 15 iun. 2026
**Know Who Spoke When: Build Real-Time, Multi-Speaker AI with NVIDIA Nemotron 3 Diarization**
Hugging Face · 23 sept. 2026