🧩 MultiConvAdapter: A Parameter-Efficient Multi-Scale Convolutional Adapter for Synthetic Speech Detection

A Parameter-Efficient Multi-Scale Convolutional Adapter for Synthetic Speech Detection

Recent synthetic speech detection models typically adapt a pre-trained self-supervised learning (SSL) model via full fine-tuning, which is computationally demanding. Parameter-Efficient Fine-Tuning (PEFT) offers a lightweight alternative, but existing methods lack the inductive biases required to model the multi-scale temporal artifacts characteristic of spoofed audio.

This repository introduces MultiConvAdapter, a parameter-efficient architecture designed to address this limitation. MultiConvAdapter integrates parallel convolutional modules within the SSL encoder, enabling the model to capture short-term artifacts and long-term distortions simultaneously.

With only 3.17M trainable parameters (~1% of the SSL backbone), MultiConvAdapter achieves state-of-the-art performance on five public datasets, outperforming both full fine-tuning and existing PEFT methods.

---

🚀 Installation

# Clone the repository
git clone https://github.com/<your_username>/MultiConvAdapter.git
cd MultiConvAdapter

# Create a conda environment
conda create -n multiconvadapter python=3.10
conda activate multiconvadapter

# Install dependencies
pip install -r requirements.txt

---

🧠 Training

1. Configure Paths
   Open config.py and set the following:

   • Training and validation data paths
   • Feature files (FALCS) paths
   • Wav2Vec/XLS-R model weights path

2. Start Training
   Run the training script:

   python3 main.py

---

🔍 Evaluation

1. Configure Test Set
   Before running evaluation, open config.py and update the path to the TEST set.

2. Run Evaluation
   Launch the evaluation script:

   python3 evaluation/evaluation.py -o output_dir

Note: Ensure that the TEST set path in config.py is correctly set before starting evaluation.

🏆 Best Reported Weights with AASIST

You can download the best-performing model checkpoints below:

• ASV19-trained model
• ASV5-trained model

Both models were trained using the configuration files available here:
🔗 Google Drive Folder