MECT-b2-finetuned

opensvbench macro EER: 7.2276

MECT-b2-finetuned is a speaker recognition and verification model based on MECT-b2 from AntResearch/AntSpeaker.

Built on a model pretrained on VoxCeleb2 dev and VoxBlink2, it is further fine-tuned on 3D-Speaker, with a focus on robustness to dialect variation and different recording distances.

Model File

• MECT-b2-finetuned.pt: A PyTorch checkpoint containing the model configuration and parameters, intended for use with the AntSpeaker MECT inference code.

Input and Output

• Input: 16 kHz mono speech audio.
• Output: Speaker embeddings for speaker similarity scoring and verification.

Acknowledgments

We thank the contributors to AntResearch/AntSpeaker and the VoxCeleb2, VoxBlink2, and 3D-Speaker datasets. Please follow the applicable licenses and terms of use of the original model and datasets.