# app.py — DeepFake AI Forensics (Audio-Only, Fixed Label Mapping)
import os
import subprocess
import tempfile
import warnings
warnings.filterwarnings('ignore')
import torch
import numpy as np
import librosa
import librosa.display
import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
import gradio as gr
from transformers import AutoFeatureExtractor, AutoModelForAudioClassification
# ==========================================
# 1. MODEL LOADING
# ==========================================
MODEL_NAME = "Hemgg/Deepfake-audio-detection"
print("[+] Loading AI forensic model...")
extractor = AutoFeatureExtractor.from_pretrained(MODEL_NAME)
model = AutoModelForAudioClassification.from_pretrained(MODEL_NAME)
model.eval()
device = "cuda" if torch.cuda.is_available() else "cpu"
model.to(device)
print(f"[+] Model loaded on {device}")
# ==========================================
# 2. AUDIO PREPROCESSING (ROBUST)
# ==========================================
def normalize_audio(file_path):
"""
Converts ANY audio to standard 16kHz mono WAV via FFmpeg.
Fixes WhatsApp voice notes (Opus/OGG disguised as MP3),
corrupt headers, and exotic codecs.
"""
out = tempfile.mktemp(suffix=".wav")
cmd = [
"ffmpeg", "-y",
"-i", file_path,
"-vn", # no video
"-acodec", "pcm_s16le", # 16-bit PCM
"-ar", "16000", # 16 kHz
"-ac", "1", # mono
"-af", "loudnorm=I=-16:TP=-1.5:LRA=11", # normalize levels
out
]
result = subprocess.run(cmd, stdout=subprocess.PIPE, stderr=subprocess.PIPE)
if result.returncode != 0:
err = result.stderr.decode('utf-8', errors='ignore')[:300]
raise RuntimeError(f"FFmpeg could not decode this file. It may be corrupted or use an unsupported codec.\nDetails: {err}")
return out
def convert_to_audio(file_path):
ext = os.path.splitext(file_path)[1].lower().lstrip('.')
# Supported audio formats (including WhatsApp Opus)
audio_exts = ["wav", "mp3", "flac", "m4a", "ogg", "opus", "aac", "wma", "oga"]
if ext in audio_exts:
print(f"[+] Audio detected ({ext}) → normalizing via ffmpeg...")
return normalize_audio(file_path)
# Unknown extension? Try ffmpeg anyway as last resort
print(f"[+] Unknown format ({ext}) → attempting ffmpeg decode...")
try:
return normalize_audio(file_path)
except Exception:
raise ValueError(f"Unsupported file format: {ext}. Please upload MP3, WAV, M4A, OGG, OPUS, or FLAC.")
def load_audio(path):
audio, _ = librosa.load(path, sr=16000)
audio = librosa.util.normalize(audio)
return audio
# ==========================================
# 3. INFERENCE & DSP
# ==========================================
def predict(audio):
inputs = extractor(audio, sampling_rate=16000, return_tensors="pt", padding=True).to(device)
with torch.no_grad():
logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)[0]
# CORRECTED: config.json says id2label: 0="AIVoice", 1="HumanVoice"
ai_p = float(probs[0])
human_p = float(probs[1])
return ai_p, human_p
def audio_features(audio):
mfcc = librosa.feature.mfcc(y=audio, sr=16000, n_mfcc=13)
return {
"mfcc_var": float(np.mean(np.var(mfcc, axis=1))),
"energy": float(np.mean(audio ** 2)),
"zcr": float(np.mean(librosa.feature.zero_crossing_rate(audio))),
"spectral_centroid": float(np.mean(librosa.feature.spectral_centroid(y=audio, sr=16000))),
}
def analyze(file_path):
audio_path = convert_to_audio(file_path)
audio = load_audio(audio_path)
ai_p, human_p = predict(audio)
feats = audio_features(audio)
# Calibrated ensemble: neural model is primary (80%), DSP is secondary (20%)
# DSP anomaly score — lower variance in MFCC and unnatural spectral centroid can indicate AI
dsp_score = min(1.0, max(0.0,
(feats["mfcc_var"] / 800.0) * 0.5 +
(1.0 - min(feats["zcr"] * 5, 1.0)) * 0.3 +
(feats["energy"] * 1.5) * 0.2
))
# Weighted fusion: trust the neural model more, use DSP as a soft modifier
ai_score = np.clip((ai_p * 0.80 + dsp_score * 0.20), 0.0, 1.0)
# Also compute human confidence for display
human_score = 1.0 - ai_score
if ai_score < 0.40:
verdict = "HUMAN VOICE"
level = "LOW RISK"
color = "#059669"
icon = "🧑"
glow = "rgba(5,150,105,0.18)"
elif ai_score < 0.60:
verdict = "UNCERTAIN / MIXED"
level = "MEDIUM RISK"
color = "#d97706"
icon = "⚠️"
glow = "rgba(217,119,6,0.18)"
else:
verdict = "AI / SYNTHETIC VOICE"
level = "HIGH RISK"
color = "#dc2626"
icon = "🤖"
glow = "rgba(220,38,38,0.18)"
confidence = int(max(ai_score, human_score) * 100)
return verdict, level, confidence, ai_score, human_score, feats, audio_path, color, icon, glow
# ==========================================
# 4. VISUALIZATION
# ==========================================
def generate_audio_plots(audio_path):
y, sr = librosa.load(audio_path, sr=16000, duration=10)
fig, (ax1, ax2) = plt.subplots(2, 1, figsize=(12, 7))
fig.patch.set_facecolor('#f0f5ff')
ax1.set_facecolor('#ffffff')
librosa.display.waveshow(y, sr=sr, ax=ax1, color='#2563eb', alpha=0.85)
ax1.set_title('Waveform Analysis', color='#1e293b', fontsize=13, fontweight='bold', pad=12)
ax1.tick_params(colors='#64748b')
for spine in ax1.spines.values():
spine.set_color('#cbd5e1')
ax2.set_facecolor('#ffffff')
mel = librosa.feature.melspectrogram(y=y, sr=sr, n_mels=128)
mel_db = librosa.power_to_db(mel, ref=np.max)
img = librosa.display.specshow(mel_db, sr=sr, ax=ax2, x_axis='time', y_axis='mel', cmap='viridis')
cbar = plt.colorbar(img, ax=ax2, format='%+2.0f dB')
cbar.ax.yaxis.set_tick_params(color='#64748b')
plt.setp(plt.getp(cbar.ax.axes, 'yticklabels'), color='#64748b')
ax2.set_title('Mel Spectrogram', color='#1e293b', fontsize=13, fontweight='bold', pad=12)
ax2.tick_params(colors='#64748b')
ax2.yaxis.label.set_color('#64748b')
ax2.xaxis.label.set_color('#64748b')
for spine in ax2.spines.values():
spine.set_color('#cbd5e1')
plt.tight_layout()
plot_path = '/tmp/audio_analysis.png'
plt.savefig(plot_path, facecolor='#f0f5ff', bbox_inches='tight', dpi=150)
plt.close()
return plot_path
# ==========================================
# 5. HTML BUILDERS
# ==========================================
def confidence_circle(percentage, color):
radius = 50
circumference = 2 * 3.14159 * radius
offset = circumference - (percentage / 100) * circumference
return f"""
"""
def probability_bar(label, percentage, color, icon):
return f"""
{icon} {label}
{percentage:.1f}%
"""
EQUALIZER_HTML = """
"""
# ==========================================
# 6. GRADIO HANDLERS
# ==========================================
def detect_audio(audio_file):
if audio_file is None:
return (
None,
'❌ No audio file provided
',
"Waiting...",
"#64748b",
EQUALIZER_HTML + 'Upload audio to begin forensic analysis
'
)
try:
verdict, level, confidence, ai_score, human_score, feats, audio_path, color, icon, glow = analyze(audio_file)
plot_path = generate_audio_plots(audio_path)
ai_pct = ai_score * 100
human_pct = human_score * 100
status_emoji = "🟢" if ai_pct < 40 else "🟡" if ai_pct < 60 else "🔴"
status_text = "LIKELY REAL" if ai_pct < 40 else "SUSPICIOUS" if ai_pct < 60 else "HIGH RISK"
circle = confidence_circle(confidence, color)
result_html = f"""
Status
{status_emoji} {status_text}
📊 Probability Breakdown
{probability_bar("AI / Synthetic", ai_pct, "#dc2626", "🤖")}
{probability_bar("Human / Real", human_pct, "#059669", "🧑")}
🔬 DSP Forensic Signatures
MFCC Variance
{feats['mfcc_var']:.4f}
Signal Energy
{feats['energy']:.6f}
Zero Crossing
{feats['zcr']:.4f}
Spectral Centroid
{feats['spectral_centroid']:.1f} Hz
Interpretation Guide:
● 0–40% Very likely genuine human voice |
● 40–60% Mixed signal, manual review advised |
● 60–100% Strong synthetic / AI indicators detected
"""
return plot_path, result_html, f"{ai_pct:.1f}%", color, ""
except Exception as e:
err_msg = str(e)
if "FFmpeg" in err_msg:
err_html = f"❌ File Decode Error
{err_msg}
WhatsApp voice notes are often .opus or .ogg files disguised as .mp3. Try renaming the file to .ogg or exporting it differently.
"
else:
err_html = f"❌ Analysis Error: {err_msg}
"
return None, err_html, "Error", "#dc2626", ""
# ==========================================
# 7. GRADIO UI — ENHANCED AUDIO-ONLY
# ==========================================
CUSTOM_CSS = """
@import url('https://fonts.googleapis.com/css2?family=Inter:wght@300;400;500;600;700;800&display=swap');
.gradio-container {
max-width: 1250px !important;
margin: auto !important;
font-family: 'Inter', sans-serif !important;
background: #e8f0fe !important;
}
body { background: #e8f0fe !important; }
/* Upload zones */
.upload-container {
background: #ffffff !important;
border: 2px dashed #93c5fd !important;
border-radius: 16px !important;
transition: all 0.3s ease !important;
}
.upload-container:hover {
border-color: #2563eb !important;
background: #eff6ff !important;
box-shadow: 0 0 30px rgba(37, 99, 235, 0.12) !important;
}
/* Buttons */
button.primary {
background: linear-gradient(135deg, #2563eb 0%, #7c3aed 100%) !important;
border: none !important;
border-radius: 12px !important;
font-weight: 700 !important;
letter-spacing: 0.5px !important;
padding: 14px 32px !important;
box-shadow: 0 4px 24px rgba(37, 99, 235, 0.25) !important;
transition: all 0.3s cubic-bezier(0.4, 0, 0.2, 1) !important;
}
button.primary:hover {
transform: translateY(-2px) !important;
box-shadow: 0 8px 32px rgba(37, 99, 235, 0.4) !important;
}
/* Equalizer Animation */
@keyframes eq-bounce {
0%, 100% { transform: scaleY(0.25); opacity: 0.5; }
50% { transform: scaleY(1); opacity: 1; }
}
/* Scrollbar */
::-webkit-scrollbar { width: 8px; }
::-webkit-scrollbar-track { background: #dbeafe; }
::-webkit-scrollbar-thumb { background: #93c5fd; border-radius: 4px; }
::-webkit-scrollbar-thumb:hover { background: #2563eb; }
/* Format badges */
.format-badge {
display: inline-block;
background: #eff6ff;
border: 1px solid #bfdbfe;
color: #2563eb;
padding: 4px 12px;
border-radius: 20px;
font-size: 0.75em;
font-weight: 600;
letter-spacing: 0.5px;
}
/* Text inputs */
input, textarea {
color: #1e293b !important;
background: #ffffff !important;
border: 1px solid #bfdbfe !important;
}
/* Audio player styling */
audio {
border-radius: 12px !important;
width: 100% !important;
}
/* Result cards */
.result-card {
background: #ffffff;
border-radius: 16px;
padding: 24px;
border: 1px solid #bfdbfe;
box-shadow: 0 2px 8px rgba(0,0,0,0.04);
}
"""
def build_ui():
with gr.Blocks(
title="DeepFake AI Forensics — Audio Detector",
theme=gr.themes.Base(
primary_hue="blue",
neutral_hue="slate",
),
css=CUSTOM_CSS,
) as demo:
# Header
gr.HTML("""
Neural + DSP ensemble detection for synthetic voice identification.
Audio-only analysis
""")
with gr.Row():
# ═══════════════════════════════════════════
# LEFT COLUMN — Upload
# ═══════════════════════════════════════════
with gr.Column(scale=1, min_width=360):
gr.Markdown("### 📤 Upload Audio File", elem_classes="section-title")
gr.HTML("""
MP3
WAV
M4A
FLAC
OGG
OPUS
AAC
📎
Maximum file size: 50 MB
💡 WhatsApp voice notes: If your file fails to upload, try renaming it from .mp3 to .ogg or .opus before uploading.
""")
audio_input = gr.Audio(
label="",
type="filepath",
elem_classes="upload-container"
)
audio_waves = gr.HTML(value=EQUALIZER_HTML + 'Audio waveform ready for analysis
')
audio_btn = gr.Button("🔍 Analyze Audio", variant="primary", size="lg")
audio_score_text = gr.Textbox(
label="",
value="--%",
interactive=False
)
# ═══════════════════════════════════════════
# RIGHT COLUMN — Results
# ═══════════════════════════════════════════
with gr.Column(scale=2):
gr.Markdown("### 📊 Forensic Analysis Report", elem_classes="section-title")
audio_plot = gr.Image(
label="",
show_label=False,
elem_classes="result-image"
)
audio_result = gr.HTML(
value="""
📊
Results will appear here
Upload an audio file and click analyze to begin
"""
)
audio_btn.click(
fn=detect_audio,
inputs=[audio_input],
outputs=[audio_plot, audio_result, audio_score_text, audio_score_text, audio_waves]
)
# ═══════════════════════════════════════════
# HOW IT WORKS SECTION
# ═══════════════════════════════════════════
gr.HTML("""
🧠 Detection Pipeline
🧠
Transformer Classifier
Hemgg/Deepfake-audio-detection
Wav2Vec 2.0 base model running on GPU/CPU with HuggingFace Transformers.
📊
DSP Ensemble
MFCC variance + signal energy + zero-crossing rate + spectral centroid fused with neural output (80/20 weighting).
🔧
Universal Decoder
FFmpeg extracts and normalizes audio from any format — including WhatsApp voice notes with disguised extensions.
📈 Score Interpretation
0–40%
🟢 Very likely genuine / human-created
40–60%
🟡 Uncertain / mixed signal — manual review recommended
60–100%
🔴 Strong AI-generated / synthetic voice indicators
⚠️ Important Limitations
No automated detector is 100% accurate. Adversarial AI models may evade detection.
Compressed or noisy audio reduces reliability. Always use human expert judgment for critical decisions.
""")
gr.HTML("""
Neural Audio Forensics • Powered by HuggingFace Transformers & DSP Signal Processing
""")
return demo
if __name__ == "__main__":
demo = build_ui()
demo.launch(
server_name="0.0.0.0",
server_port=7860,
share=False,
ssr_mode=False
)