---
title: Linearhead Leaderboard
emoji: 🏃
colorFrom: gray
colorTo: yellow
sdk: gradio
sdk_version: 5.49.1
app_file: app.py
pinned: false
license: apache-2.0
short_description: Comprehensive comparison of Linear-Head classifiers
---

Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

# 🎙️ Linear-Head Model Leaderboard

This leaderboard presents a comprehensive comparison of **Linear-Head classifiers** trained on a variety of **Self-Supervised Learning (SSL)** speech representations from the **S3PRL** library. It highlights model performance across multiple spoofing datasets, codecs, and TTS attacks in the context of **audio deepfake detection**.

---

## Frontend – SSL Feature Extractors

The **frontend** of each model is a frozen SSL feature extractor from **S3PRL**, capable of generating rich speech embeddings.
These extractors are pre-trained on large-scale audio corpora and capture different aspects of speech acoustics and phonetic content.
The leaderboard includes models built with several SSL backbones such as:

* **WavLM-Large**
* **Wav2Vec 2.0 XLSR (xls_r_300m)**
* **NPC 960 hr**
* **HuBERT**, **APC**, and others

Each extractor converts input waveforms into frame-level representations, serving as the foundation for downstream spoof detection.

---

## Backend – Classifier Models

On top of these SSL embeddings, four **downstream classifier architectures** are implemented.
Among them, the **Linear-Head model** serves as a lightweight yet highly effective backend.
It projects the SSL features into spoof/bonafide decision scores using a single fully connected layer trained with binary classification loss.
The simplicity of this approach allows fast adaptation and fair benchmarking across different SSL frontends.

---

## What the Leaderboard Shows

The leaderboard summarizes key results from extensive evaluations.
It includes separate sections for:

* **Main Leader Board** – Overall ranking based on average EER or TNR.
* **Models Performance on Each Data** – Per-dataset or per-attack breakdowns.
* **TTS Difficulty Level Per Model** – Shows which TTS generators most effectively fool the models.
* **Performance on Codecs** – Evaluates robustness under various compression schemes.
* **Best Model per Attack** – Highlights the top-performing model for each individual attack type.

---

## Purpose

The goal of this leaderboard is to provide a transparent, unified view of **how SSL-based frontends and lightweight classifier backends perform in deepfake speech detection tasks**.
It enables researchers and engineers to identify the most robust combinations of feature extractors and classifier heads, supporting future improvements in generalization, efficiency, and security of speech authentication systems.