Audio Classification
MECT
PyTorch
English
speaker-verification
voiceprint
speaker-recognition
mixture-of-experts
CNN-Transformer
audio
streaming
Instructions to use AntResearch/AntSpeaker with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- MECT
How to use AntResearch/AntSpeaker with MECT:
# No code snippets available yet for this library. # To use this model, check the repository files and the library's documentation. # Want to help? PRs adding snippets are welcome at: # https://github.com/huggingface/huggingface.js
- Notebooks
- Google Colab
- Kaggle
Upload README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,61 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
library_name: mect
|
| 3 |
+
license: cc-by-nc-sa-4.0
|
| 4 |
+
language:
|
| 5 |
+
- en
|
| 6 |
+
tags:
|
| 7 |
+
- speaker-verification
|
| 8 |
+
- voiceprint
|
| 9 |
+
- speaker-recognition
|
| 10 |
+
- mixture-of-experts
|
| 11 |
+
- CNN-Transformer
|
| 12 |
+
- audio
|
| 13 |
+
- pytorch
|
| 14 |
+
pipeline_tag: audio-classification
|
| 15 |
+
---
|
| 16 |
+
|
| 17 |
+
# MECT
|
| 18 |
+
|
| 19 |
+
Official model checkpoints for the paper **"MECT: Mixture of Experts with CNN-Transformer Network for Speaker Verification"**.
|
| 20 |
+
|
| 21 |
+
MECT is a novel speaker verification architecture that combines the strengths of **Convolutional Neural Networks (CNNs)** and **Transformer networks** through a **Mixture of Experts (MoE)** approach. The model leverages expert routing to enhance performance and robustness in speaker recognition tasks.
|
| 22 |
+
|
| 23 |
+
## π Key Features
|
| 24 |
+
|
| 25 |
+
- **CNN + Transformer fusion** β combines local feature extraction (CNN) with global context modeling (Transformer) in a unified architecture
|
| 26 |
+
- **Mixture of Experts** β dynamic expert routing enables specialized feature processing for diverse speakers and acoustic conditions
|
| 27 |
+
- **Lightweight and scalable** β four model sizes ranging from 3.78M to 14.43M parameters, all producing 192-dimensional embeddings
|
| 28 |
+
- **Strong performance** β competitive EER on VoxCeleb1 benchmarks
|
| 29 |
+
|
| 30 |
+
## π¦ Model Variants
|
| 31 |
+
|
| 32 |
+
This repository provides **4 model sizes** with **5 checkpoints** in total:
|
| 33 |
+
|
| 34 |
+
| Model | Parameters | Embedding Dim | Training Data | Checkpoint File |
|
| 35 |
+
|-------|-----------|---------------|---------------|-----------------|
|
| 36 |
+
| MECT-A1 | 3.78M | 192 | VoxCeleb2 | `mect_a1_vc2.pt` |
|
| 37 |
+
| MECT-A2 | 4.12M | 192 | VoxCeleb2 | `mect_a2_vc2.pt` |
|
| 38 |
+
| MECT-B1 | 8.26M | 192 | VoxCeleb2 | `mect_b1_vc2.pt` |
|
| 39 |
+
| MECT-B2 | 14.43M | 192 | VoxCeleb2 | `mect_b2_vc2.pt` |
|
| 40 |
+
| MECT-B2 | 14.43M | 192 | VoxCeleb2 + VoxBlink2 | `mect_b2_vb2.pt` |
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
|
| 44 |
+
## π License
|
| 45 |
+
|
| 46 |
+
This model is released under the **Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License (CC-BY-NC-SA 4.0)**.
|
| 47 |
+
|
| 48 |
+
- **Non-Commercial** β you may not use this model for commercial purposes
|
| 49 |
+
- **Attribution** β you must give appropriate credit
|
| 50 |
+
- **ShareAlike** β derivatives must be licensed under the same terms
|
| 51 |
+
|
| 52 |
+
See the [LICENSE](https://creativecommons.org/licenses/by-nc-sa/4.0/) for full details.
|
| 53 |
+
|
| 54 |
+
## π Links
|
| 55 |
+
|
| 56 |
+
- **GitHub Repository**: [https://github.com/ant-research/AntSpeaker](https://github.com/ant-research/AntSpeaker)
|
| 57 |
+
- **HuggingFace Model**: [https://huggingface.co/AntResearch/AntSpeaker](https://huggingface.co/AntResearch/AntSpeaker)
|
| 58 |
+
|
| 59 |
+
## π Acknowledgements
|
| 60 |
+
|
| 61 |
+
We thank the authors of [VoxCeleb2](https://www.robots.ox.ac.uk/~vgg/data/voxceleb/) and [VoxBlink2](https://github.com/VoxBlink/VoxBlink2) for providing the training corpora.
|