initial

README copy.md

@@ -0,0 +1,48 @@
+# 🎙️ 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.
+

README.md

@@ -12,3 +12,52 @@ 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.
+

app.py

@@ -0,0 +1,145 @@
+# app.py
+# Clean, read-only leaderboard with a banner image and tabbed pages (no buttons/uploads).
+# Improved styling so the explanation text is clearly visible in both light & dark themes.
+#
+# Usage:
+#   pip install -r requirements.txt
+#   python app.py
+#
+# Optional: set env vars:
+#   LB_DATA_PATH   -> path to your Excel (defaults to "data.xlsx")
+#   LB_BANNER_PATH -> path to a top banner image (defaults to "banner.png")
+
+import os
+import pandas as pd
+import gradio as gr
+
+DATA_PATH = os.environ.get("LB_DATA_PATH", "data.xlsx")
+BANNER_PATH = os.environ.get("LB_BANNER_PATH", "banner.png")  # change if your file is different
+
+# ---- Explanations shown above each table ----
+SHEET_DESCRIPTIONS = {
+    "Main Leader Board": (
+        "<b>Overview.</b> This Table compares models at a glance across the full deepfake detection suite. "
+        "Use it to spot <i>overall leaders</i> and identify systems that maintain strong performance "
+        "under diverse conditions. Where present, <b>Avg&nbsp;EER&nbsp;(↓)</b> is the primary error metric "
+        "(lower is better). If you see robustness columns such as <b>Codec&nbsp;(↑)</b>, "
+        "<b>TTS&nbsp;(↑)</b>, or <b>Cross-Gen&nbsp;(↑)</b>, higher values indicate stronger generalization."
+    ),
+    "Models Performance on Each Data": (
+        "<b>Per-dataset breakdown.</b> Each column corresponds to a dataset, subset, or attack group "
+        "(e.g., ASVspoof splits, Famous Figures, MLAAD; sometimes A-IDs like A07, A15). "
+        "Interpretation: <b>EER&nbsp;(↓)</b> lower is better; <b>TNR&nbsp;(↑)</b> higher is better. "
+        "Use this table to pinpoint which datasets or attacks are <i>hardest</i> and which are <i>easiest</i> "
+        "for each model, and to diagnose domain-specific weaknesses."
+    ),
+    "TTS Difficultly Level Per Model": (
+        "<b>TTS stress-test.</b> This Table shows how challenging different <b>TTS generators</b> are for each "
+        "detection model. For <b>TNR</b>, lower values mean the TTS fools the model more (i.e., harder); "
+        "higher values mean the model rejects that TTS more reliably (easier).<br><br>"
+        "<b>Key finding (from mean TNR across all TTS systems & datasets):</b> "
+        "<span style='white-space:nowrap;'>Hardest → <b>ASVSpoof&nbsp;5&nbsp;Eval – A31</b></span> "
+        "with <b>Mean&nbsp;TNR&nbsp;= 0.0221</b>; "
+        "<span style='white-space:nowrap;'>Easiest → <b>tts_models_it_mai_female_vits</b> (MLAAD)</span> "
+        "with <b>Mean&nbsp;TNR&nbsp;= 0.9961</b>."
+    ),
+    "Performance On Codecs": (
+        "<b>Compression robustness.</b> Columns represent codec/bitrate conditions; performance reflects whether "
+        "compression hides or amplifies spoof cues. If metrics are <b>EER&nbsp;(↓)</b>, lower is better; "
+        "if <b>TNR&nbsp;(↑)</b>, higher is better. Use this Table to compare which models remain stable when "
+        "audio is encoded for streaming, storage, or telephony."
+    ),
+    # NEW: tab for per-attack winners
+    "Best Model per Attack": (
+        "<b>Per-attack winners.</b> For each attack (e.g., A07, A15, A31, etc.), this table lists the "
+        "<i>single best-performing model</i> along with its corresponding <b>TNR&nbsp;(↑)</b>. "
+        "Use it to quickly see which model you should trust most against each specific attack family."
+    ),
+}
+
+def load_sheets(path: str):
+    if not os.path.exists(path):
+        raise FileNotFoundError(
+            f"Excel file not found at '{path}'. "
+            "Place your workbook next to app.py as 'data.xlsx' or set LB_DATA_PATH."
+        )
+    xls = pd.ExcelFile(path)
+    # Read every sheet; if "Best Model per Attack" is present it will be included automatically.
+    return {name: pd.read_excel(path, sheet_name=name) for name in xls.sheet_names}
+
+def build_app():
+    sheets = load_sheets(DATA_PATH)
+
+    with gr.Blocks(
+        title="🎙️ Benchmarking Linear-Head Classifiers Built on S3PRL Embeddings",
+        css="""
+        .gradio-container { max-width: 1200px !important; }
+
+        /* Title */
+        #title h1 {
+            text-align: center;
+            font-size: 2.1em;
+            margin: 0.5rem 0 0.75rem 0;
+            line-height: 1.25;
+        }
+
+        /* Banner */
+        #banner { border-radius: 16px; margin: 0.5rem auto 0.25rem auto; }
+
+        /* Sheet description card — visible in both light & dark */
+        [data-theme="light"] .sheet-card {
+            background: #f9fafb;
+            color: #111827;
+            border: 1px solid #e5e7eb;
+            border-radius: 12px;
+            padding: 14px 16px;
+            box-shadow: 0 1px 0 rgba(0,0,0,0.02);
+        }
+        [data-theme="dark"] .sheet-card {
+            background: #111827;
+            color: #e5e7eb;
+            border: 1px solid #374151;
+            border-radius: 12px;
+            padding: 14px 16px;
+            box-shadow: 0 0 0 rgba(0,0,0,0);
+        }
+        .sheet-card p { margin: 0.4rem 0; font-size: 15px; line-height: 1.6; }
+
+        /* Dataframe spacing */
+        .gr-dataframe { margin-top: 10px; }
+        """
+    ) as demo:
+        # --- Banner Image and Title ---
+        if os.path.exists(BANNER_PATH):
+            gr.Image(value=BANNER_PATH, show_label=False, elem_id="banner")
+        gr.Markdown(
+            "<h1>🎙️ Benchmarking Linear-Head Classifiers Built on S3PRL Embeddings</h1>",
+            elem_id="title",
+        )
+
+        # --- Tabs / Pages ---
+        with gr.Tabs():
+            # Keep workbook order, but ensure our description card appears if we know the sheet name.
+            for sheet_name, df in sheets.items():
+                with gr.TabItem(sheet_name):
+                    desc = SHEET_DESCRIPTIONS.get(
+                        sheet_name,
+                        "This table shows the original sheet data from the analysis workbook."
+                    )
+                    # Explanation card
+                    gr.Markdown(f"<div class='sheet-card'>{desc}</div>")
+                    # Single raw table (read-only)
+                    gr.Dataframe(
+                        value=df,
+                        interactive=False,
+                        wrap=True,
+                        label=None,
+                        elem_id=f"df_{sheet_name.replace(' ', '_').lower()}",
+                    )
+    return demo
+
+demo = build_app()
+
+if __name__ == "__main__":
+    # For a public link while testing locally, use: demo.launch(share=True)
+    demo.launch()

requirements.txt

@@ -0,0 +1,3 @@
+gradio>=4.44.0
+pandas>=2.0.0
+openpyxl>=3.1.0