XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models

Official implementation for the paper "XAI-Grounded Explanation Generation for Speech Deepfake Detection with Training-Free Multimodal Large Language Models".

This repository provides scripts for speech deepfake detection, XAI feature attribution generation, and training-free multimodal large language model based explanation generation.

Overview

The pipeline contains three main stages:

1. Prediction: run speech deepfake detection models and obtain correctly predicted samples.
2. Feature-to-XAI: generate explainable AI visualizations from model features, including Integrated Gradients, Saliency, LIME, and SHAP-style explanations.
3. XAI-to-Text: use a multimodal large language model to convert XAI images into grounded natural-language explanation regions.

Repository Structure

.
|-- database/                  # Dataset directory. Download from Zenodo before running scripts.
|-- res/                       # Generated explanation results.
|   |-- one_model_4XAI/        # Results based on one-model XAI settings.
|   `-- three_model_4XAI/      # Results based on three-model XAI settings.
`-- script/
    |-- predict/               # Model feature extraction and prediction scripts.
    |   |-- environment.yaml
    |   |-- opensmile_feature.py
    |   |-- MLP_opensmile.py
    |   |-- hubert_try.py
    |   |-- wav2vec_try.py
    |   `-- wavLM_try.py
    |-- feature2xai/           # Convert model features/predictions to XAI images.
    |   |-- opensmile2xai.py
    |   |-- hubert2xai.py
    |   |-- wav2vec2xai.py
    |   `-- wavlm2xai.py
    `-- xai2txt.py             # Convert XAI images to textual explanations.

Dataset

This project uses the PartialSpoof Database - Partially Spoofed Audio Dataset for Anti-spoofing.

Download the dataset from Zenodo:

https://zenodo.org/records/4817532

After downloading and extracting the dataset, place it under database/ with the expected split structure:

database/
|-- dev/
|-- eval/
`-- train/

For the evaluation archive, Zenodo provides split files. Concatenate them before extraction:

cat database_eval.tar.gz.a* > database_eval.tar.gz
tar -zxvf database_eval.tar.gz

Environment

A Conda environment file is provided in script/predict/environment.yaml.

conda env create -f script/predict/environment.yaml
conda activate antideepfake

The scripts use common speech, deep learning, XAI, and multimodal LLM packages, including:

• torch
• torchaudio
• transformers
• fairseq
• huggingface_hub
• captum
• librosa
• scikit-learn
• matplotlib
• qwen-vl-utils
• opensmile

Depending on your CUDA/PyTorch setup, some packages may need to be installed manually.

Usage

1. Run Speech Deepfake Detection

Prediction scripts are stored in script/predict/.

Examples:

python script/predict/wav2vec_try.py
python script/predict/hubert_try.py
python script/predict/wavLM_try.py

For OpenSMILE-based prediction, first extract OpenSMILE features and then run the MLP detector:

python script/predict/opensmile_feature.py
python script/predict/MLP_opensmile.py

These scripts evaluate deepfake detection performance and can be used to identify correctly predicted samples for downstream XAI generation.

2. Generate XAI Images

XAI image generation scripts are stored in script/feature2xai/.

Examples:

python script/feature2xai/wav2vec2xai.py
python script/feature2xai/hubert2xai.py
python script/feature2xai/wavlm2xai.py
python script/feature2xai/opensmile2xai.py

The scripts generate attribution visualizations such as:

• Integrated Gradients
• Saliency
• LIME
• SHAP/OpenSMILE-based explanations

Before running, check and update the path variables near the top of each script, such as:

• database_root
• file_list_path
• output_root

3. Convert XAI Images to Text

The XAI-to-text stage is implemented in:

python script/xai2txt.py

This script uses Qwen2.5-VL-7B-Instruct to read generated XAI images and produce grounded textual explanation regions, such as the most important time and frequency ranges.

Before running, update the input and output paths in script/xai2txt.py:

INPUT_ROOT = "./XAI_Image/wav2vec/eval"
OUTPUT_ROOT = "./XAI_text/wav2vec/eval"

Results

Generated results are stored in res/.

• res/one_model_4XAI/: explanation outputs from the one-model XAI setting.
• res/three_model_4XAI/: explanation outputs from the three-model XAI setting.

Each text file contains the explanation output generated from XAI-grounded multimodal reasoning.

Setting	Generated Explanation	Comments
Pure audio	The audio clip begins with a faint, low-frequency thump, likely a microphone pop or handling noise, followed by a brief pause. The voice is monotonous, with a flat, robotic intonation and a lack of emotional inflection or dynamic range...	Too superficial
IG	...at the word "time" (1.5–2.0 seconds), the voice exhibits a sudden, unnatural pitch drop and a brief, glitch-like distortion, accompanied by a sharp, high-frequency artifact in the 6000–7000 Hz range... The XAI explanation using IG highlights this same time and frequency region...	Specific but limited to single-model evidence
LIME	The audio exhibits a noticeable acoustic anomaly in the form of a high-frequency, synthetic-sounding artifact occurring between 2.5 and 3 seconds... The XAI explanation from LIME identifies the TIME_REGION: [2.5–3] seconds as a region of interest, which aligns with the observed anomaly...	Specific but biased
Saliency	The audio exhibits a noticeable acoustic anomaly in the form of a sharp, high-frequency transient at approximately 4.5–5 seconds, coinciding with the word "insisted." The XAI saliency map confirms this region as a point of high gradient importance...	Specific but biased
All XAI (Single Model)	...The speech is: "This time, believe the silent good and insisted the lesser." The most notable acoustic anomaly occurs between 4.5 and 5.0 seconds, where a subtle, unnatural pitch drop is observed in the word "insisted." The Saliency analysis confirms this region as a key area of abnormality. Other XAI evidence, such as the SHAP values for loudness...	Specific and analysed
All XAI (Three Model)	The audio begins with a female voice speaking in a synthetic tone: "This time, believe ..." The speech is clear and evenly paced, with no emotional inflection or natural pauses, suggesting a text-to-speech (TTS) origin. At approximately 1.5 seconds, a subtle but noticeable high-frequency artifact emerges, lasting until 2.0 seconds... The wav2vec model identifies... The HuBERT and WavLM models do not provide any evidence for this region...	Specific, analysed and reasoned