|
import os
|
|
import torch
|
|
import librosa
|
|
import numpy as np
|
|
import gradio as gr
|
|
from sonics import HFAudioClassifier
|
|
|
|
|
|
MODEL_IDS = {
|
|
"SpecTTTra-α (5s)": "awsaf49/sonics-spectttra-alpha-5s",
|
|
"SpecTTTra-β (5s)": "awsaf49/sonics-spectttra-beta-5s",
|
|
"SpecTTTra-γ (5s)": "awsaf49/sonics-spectttra-gamma-5s",
|
|
"SpecTTTra-α (120s)": "awsaf49/sonics-spectttra-alpha-120s",
|
|
"SpecTTTra-β (120s)": "awsaf49/sonics-spectttra-beta-120s",
|
|
"SpecTTTra-γ (120s)": "awsaf49/sonics-spectttra-gamma-120s",
|
|
}
|
|
|
|
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
|
|
model_cache = {}
|
|
|
|
|
|
def load_model(model_name):
|
|
"""Load model if not already cached"""
|
|
if model_name not in model_cache:
|
|
model_id = MODEL_IDS[model_name]
|
|
model = HFAudioClassifier.from_pretrained(model_id)
|
|
model = model.to(device)
|
|
model.eval()
|
|
model_cache[model_name] = model
|
|
return model_cache[model_name]
|
|
|
|
|
|
def process_audio(audio_path, model_name):
|
|
"""Process audio file and return prediction"""
|
|
try:
|
|
model = load_model(model_name)
|
|
max_time = model.config.audio.max_time
|
|
|
|
|
|
audio, sr = librosa.load(audio_path, sr=16000)
|
|
chunk_samples = int(max_time * sr)
|
|
total_chunks = len(audio) // chunk_samples
|
|
middle_chunk_idx = total_chunks // 2
|
|
|
|
|
|
start = middle_chunk_idx * chunk_samples
|
|
end = start + chunk_samples
|
|
chunk = audio[start:end]
|
|
|
|
if len(chunk) < chunk_samples:
|
|
chunk = np.pad(chunk, (0, chunk_samples - len(chunk)))
|
|
|
|
|
|
with torch.no_grad():
|
|
chunk = torch.from_numpy(chunk).float().to(device)
|
|
pred = model(chunk.unsqueeze(0))
|
|
prob = torch.sigmoid(pred).cpu().numpy()[0]
|
|
|
|
real_prob = 1 - prob
|
|
fake_prob = prob
|
|
|
|
|
|
return {
|
|
"Real": float(real_prob),
|
|
"Fake": float(fake_prob)
|
|
}
|
|
|
|
except Exception as e:
|
|
return {"Error": str(e)}
|
|
|
|
|
|
def predict(audio_file, model_name):
|
|
"""Gradio interface function"""
|
|
if audio_file is None:
|
|
return {"Message": "Please upload an audio file"}
|
|
return process_audio(audio_file, model_name)
|
|
|
|
|
|
|
|
css = """
|
|
:root {
|
|
--primary-bg: #000000;
|
|
--secondary-bg: #111111;
|
|
--panel-bg: #1e1e1e;
|
|
--text-color: #ffffff;
|
|
--text-secondary: #bbbbbb;
|
|
--border-color: #333333;
|
|
--analyze-button-color: #ffa500;
|
|
--analyze-button-hover: #ff8c00;
|
|
--accent-color: #4a78e5;
|
|
}
|
|
|
|
body, .gradio-container {
|
|
background-color: var(--primary-bg) !important;
|
|
color: var(--text-color) !important;
|
|
}
|
|
|
|
.footer, .header-container, .accordion-content {
|
|
background-color: var(--secondary-bg) !important;
|
|
color: var(--text-color) !important;
|
|
}
|
|
|
|
/* Headers and text */
|
|
h1, h2, h3 {
|
|
color: var(--text-color) !important;
|
|
}
|
|
|
|
p {
|
|
color: var(--text-secondary) !important;
|
|
}
|
|
|
|
/* Button styling */
|
|
button#submit_btn {
|
|
background-color: var(--analyze-button-color) !important;
|
|
color: white !important;
|
|
border: none !important;
|
|
font-weight: bold !important;
|
|
padding: 10px 20px !important;
|
|
font-size: 16px !important;
|
|
border-radius: 8px !important;
|
|
}
|
|
|
|
button#submit_btn:hover {
|
|
background-color: var(--analyze-button-hover) !important;
|
|
box-shadow: 0 4px 8px rgba(0, 0, 0, 0.5) !important;
|
|
transform: translateY(-2px) !important;
|
|
transition: all 0.2s ease !important;
|
|
}
|
|
|
|
/* Panel backgrounds */
|
|
.gr-panel, .gr-box, .gr-form, .gr-input-label, .gr-input {
|
|
background-color: var(--panel-bg) !important;
|
|
border: 1px solid var(--border-color) !important;
|
|
border-radius: 8px !important;
|
|
color: var(--text-color) !important;
|
|
}
|
|
|
|
/* Results panel */
|
|
#output {
|
|
background-color: var(--panel-bg) !important;
|
|
border-radius: 8px !important;
|
|
padding: 10px !important;
|
|
box-shadow: 0 2px 5px rgba(0, 0, 0, 0.3) !important;
|
|
border: 1px solid var(--border-color) !important;
|
|
}
|
|
|
|
/* Resource links */
|
|
.resource-link {
|
|
background-color: var(--secondary-bg) !important;
|
|
color: var(--accent-color) !important;
|
|
border: 1px solid var(--border-color) !important;
|
|
padding: 8px 16px !important;
|
|
border-radius: 20px !important;
|
|
margin: 5px !important;
|
|
text-decoration: none !important;
|
|
display: inline-block !important;
|
|
font-weight: 500 !important;
|
|
box-shadow: 0 2px 4px rgba(0, 0, 0, 0.3) !important;
|
|
}
|
|
|
|
.resource-link:hover {
|
|
transform: translateY(-2px) !important;
|
|
box-shadow: 0 3px 6px rgba(0, 0, 0, 0.4) !important;
|
|
transition: all 0.2s ease !important;
|
|
background-color: #222222 !important;
|
|
}
|
|
|
|
.header-container {
|
|
padding: 20px !important;
|
|
border-radius: 10px !important;
|
|
box-shadow: 0 2px 10px rgba(0, 0, 0, 0.3) !important;
|
|
margin-bottom: 20px !important;
|
|
border: 1px solid var(--border-color) !important;
|
|
}
|
|
|
|
/* Accordion styling */
|
|
.gr-accordion {
|
|
border: 1px solid var(--border-color) !important;
|
|
border-radius: 8px !important;
|
|
overflow: hidden !important;
|
|
background-color: var(--panel-bg) !important;
|
|
}
|
|
|
|
.gr-accordion-header {
|
|
background-color: var(--secondary-bg) !important;
|
|
padding: 10px 15px !important;
|
|
font-weight: 600 !important;
|
|
color: var(--text-color) !important;
|
|
}
|
|
|
|
/* Audio player */
|
|
.audio-player {
|
|
background-color: var(--panel-bg) !important;
|
|
border-radius: 8px !important;
|
|
overflow: hidden !important;
|
|
}
|
|
|
|
/* Dropdown & Input fields */
|
|
select, input, .gr-dropdown {
|
|
background-color: var(--panel-bg) !important;
|
|
color: var(--text-color) !important;
|
|
border: 1px solid var(--border-color) !important;
|
|
}
|
|
|
|
/* Labels */
|
|
label, .gr-label {
|
|
color: var(--text-secondary) !important;
|
|
}
|
|
|
|
/* Footer styling */
|
|
.footer {
|
|
border-top: 1px solid var(--border-color) !important;
|
|
margin-top: 30px !important;
|
|
padding: 15px !important;
|
|
}
|
|
"""
|
|
|
|
|
|
with gr.Blocks(css=css, theme=gr.themes.Default()) as demo:
|
|
|
|
gr.HTML(
|
|
"""
|
|
<div class="header-container">
|
|
<div style="display: flex; justify-content: center; margin-bottom: 20px;">
|
|
<img src="https://i.postimg.cc/3Jx3yZ5b/real-vs-fake-sonics-w-logo.jpg"
|
|
style="max-width: 150px; border-radius: 10px; box-shadow: 0 4px 8px rgba(0,0,0,0.3);">
|
|
</div>
|
|
<h1 style="text-align: center; font-size: 28px; margin-bottom: 10px; color: #ffffff;">SONICS: Synthetic Or Not - Identifying Counterfeit Songs</h1>
|
|
<h3 style="text-align: center; color: #bbbbbb; margin-bottom: 15px;">ICLR 2025 [Poster]</h3>
|
|
<p style="text-align: center; font-size: 16px; margin: 0; color: #aaaaaa;">
|
|
Detect if a song is real or AI-generated with our state-of-the-art models.
|
|
Simply upload an audio file to verify its authenticity!
|
|
</p>
|
|
</div>
|
|
"""
|
|
)
|
|
|
|
|
|
gr.HTML(
|
|
"""
|
|
<div style="display: flex; justify-content: center; flex-wrap: wrap; gap: 8px; margin-bottom: 25px;">
|
|
<a href="https://openreview.net/forum?id=PY7KSh29Z8" target="_blank" class="resource-link">
|
|
📄 Paper
|
|
</a>
|
|
<a href="https://huggingface.co/datasets/awsaf49/sonics" target="_blank" class="resource-link">
|
|
🎵 Dataset
|
|
</a>
|
|
<a href="https://huggingface.co/collections/awsaf49/sonics-spectttra-67bb6517b3920fd18e409013" target="_blank" class="resource-link">
|
|
🤖 Models
|
|
</a>
|
|
<a href="https://arxiv.org/abs/2408.14080" target="_blank" class="resource-link">
|
|
🔬 ArXiv
|
|
</a>
|
|
<a href="https://github.com/awsaf49/sonics" target="_blank" class="resource-link">
|
|
💻 GitHub
|
|
</a>
|
|
</div>
|
|
"""
|
|
)
|
|
|
|
|
|
with gr.Row(equal_height=True):
|
|
with gr.Column():
|
|
audio_input = gr.Audio(
|
|
label="Upload Audio File",
|
|
type="filepath",
|
|
elem_id="audio_input",
|
|
elem_classes="audio-player"
|
|
)
|
|
|
|
model_dropdown = gr.Dropdown(
|
|
choices=list(MODEL_IDS.keys()),
|
|
value="SpecTTTra-γ (5s)",
|
|
label="Select Model",
|
|
elem_id="model_dropdown"
|
|
)
|
|
|
|
submit_btn = gr.Button(
|
|
"✨ Analyze Audio",
|
|
elem_id="submit_btn"
|
|
)
|
|
|
|
with gr.Column():
|
|
|
|
output = gr.Label(
|
|
label="Analysis Result",
|
|
num_top_classes=2,
|
|
elem_id="output"
|
|
)
|
|
|
|
with gr.Accordion("How It Works", open=True):
|
|
gr.Markdown("""
|
|
## The SONICS classifier
|
|
|
|
The SONICS classifier analyzes your audio to determine if it's an authentic song (human created) or generated by AI. Our models are trained on a diverse dataset of real and AI-generated songs from Suno and Udio.
|
|
|
|
### Models available:
|
|
- **SpecTTTra-α**: Optimized for speed
|
|
- **SpecTTTra-β**: Balanced performance
|
|
- **SpecTTTra-γ**: Highest accuracy
|
|
|
|
### Duration variants:
|
|
- **5s**: Analyzes a 5-second clip (faster)
|
|
- **120s**: Analyzes up to 2 minutes (more accurate)
|
|
""")
|
|
|
|
|
|
with gr.Accordion("Example Audio Files", open=True):
|
|
gr.Examples(
|
|
examples=[
|
|
["demo/real_song.mp3", "SpecTTTra-γ (5s)"],
|
|
["demo/fake_song.mp3", "SpecTTTra-γ (5s)"],
|
|
],
|
|
inputs=[audio_input, model_dropdown],
|
|
outputs=[output],
|
|
fn=predict,
|
|
cache_examples=True,
|
|
)
|
|
|
|
|
|
gr.HTML(
|
|
"""
|
|
<div class="footer" style="text-align: center;">
|
|
<p style="color: #bbbbbb; font-size: 14px;">SONICS: Synthetic Or Not - Identifying Counterfeit Songs | ICLR 2025</p>
|
|
<p style="color: #777777; font-size: 12px;">For research purposes only</p>
|
|
</div>
|
|
"""
|
|
)
|
|
|
|
|
|
submit_btn.click(fn=predict, inputs=[audio_input, model_dropdown], outputs=[output])
|
|
|
|
if __name__ == "__main__":
|
|
|
|
demo.launch() |
