DSE-vTAD

This is the AuroraLab system for the Voice Timbre Attribute Detection (vTAD) 2025 Challenge. We developed our system based on the official vTAD baseline.

vTAD Challenge 2025

Please visit the challenge website for more information about the Challenge. Decription of the baseline methods can be found in: baseline paper link.

Data

The VCTK-RVA dataset is employed in our work, wherein the publicly available VCTK database was annotated for timbre intensity.

The Dataset/attribute_pair/train.txt contains the annotation information for the training set of this competition. The annotation information for some of the data is as follows:

明亮_F: p314|p268, p262|p257, p294|p250...
粗_F: p333|p269, p253|p225, p303|p295...
明亮_M: p363|p334, p364|p263, p263|p232...
单薄_M: p285|p363, p226|p298, p275|p364...
...

attribute_gender: speaker A|speaker B represents the comparison of the strength of the two speakers of the same gender(F OR M) in this attribute. In the annotation, the speaker on the right, speakerB, is stronger in this attribute compared to the speaker on the left, speakerA.

Dependencies

Our experiment was conducted in a Python==3.8.20 environment for training and testing. Please run the following command :

pip install torch==1.12.1+cu113 torchaudio==0.12.1 --extra-index-url https://download.pytorch.org/whl/cu113

Then run the following command to install other dependencies:

pip install -r requirements.txt

Download Pretrained Model

In this experiment, we used ECAPA-TDNN and FACodec as pre-trained speaker encoder to extract speaker embeddings. The checkpoint for these two pre-trained speaker encoder can be downloaded from the link.

In addition, we also conducted experiments using the pretrained models provided by Wespeaker:

• SimAM-ResNet34
• SimAM-ResNet100
• ResNet293_LM

Run

We take the SimAM-ResNet100 pretrained speaker recognition embedding extractor as an example.

Path Configuration

configs/baseline.json

• /path/to/VCTK-Corpus-0.92/wav16
• /path/to/test/anonymous_VCTK_RVA
• /path/to/voxblink2_samresnet100/voxblink2_samresnet100

voxblink2_samresnet100_wespeaker/voxblink2_samresnet100_test.py

• /path/to/configs/baseline.json

voxblink2_samresnet100_wespeaker/voxblink2_samresnet100.py

• /path/to/configs/baseline.json

Run DSE-vTAD

bash run_voxblink2_samresnet100_wespeaker.sh