|
--- |
|
license: mit |
|
datasets: |
|
- ProgramComputer/voxceleb |
|
base_model: |
|
- espnet/xeus |
|
- facebook/w2v-bert-2.0 |
|
--- |
|
# 🗣️ Speaker Verification Checkpoints |
|
|
|
This repository hosts pretrained and fine-tuned model checkpoints used in the course project [The Impact of Language on Automatic Speaker Verification Systems](https://github.com/Stefan2417/CourseWork). |
|
The project explores the effectiveness of multilingual self-supervised speech models adapted for **speaker verification** using **PMFA** (Partial Multi-scale Feature Aggregation), especially in **low-resource and multi-lingual** scenarios. |
|
|
|
--- |
|
|
|
## 📦 Available Checkpoints |
|
|
|
| File | Description | |
|
|--------------------------------------------------|------------------------------------------------------------| |
|
| `tunned_xeus.pth` | Fine-tuned XEUS encoder with full PMFA pipeline | |
|
| `xeus_adapter_checkpoint.pth` | Adapter-only checkpoint for XEUS | |
|
| `wav2vec-bert_adapter.pth` | Adapter-only for Wav2Vec2-BERT | |
|
| `wav2vec2-bert_tunned_without_large_margin.pth` | Fine-tuned Wav2Vec2-BERT (standard margin) | |
|
| `wav2vec2-bert-large-margin.pth` | Fine-tuned Wav2Vec2-BERT using large-margin classification | |
|
|
|
> All checkpoints are compatible with the training and evaluation scripts in [Stefan2417/CourseWork](https://github.com/Stefan2417/CourseWork) |
|
|
|
--- |
