tiny-ast-mad-military-audio-classifier

Tiny-AST Military Audio Classifier

🎖️ State-of-the-art military audio classification model achieving 96.73% accuracy on the Military Audio Dataset (MAD).

🌐 Live Demo

This model is deployed and accessible via Azure cloud:

🔗 Try the Live Demo

• Infrastructure: Azure VM with FastAPI backend
• Runtime: ONNX Runtime for optimized inference
• Interface: Browser-based audio upload and real-time classification
• SSL: Self-signed certificate (accept the browser warning to access)

⚠️ The demo uses a self-signed SSL certificate. Your browser may show a security warning — click "Advanced" → "Proceed" to access.

Model Description

This model is a fine-tuned version of MIT/ast-finetuned-audioset-10-10-0.4593 on the Military Audio Dataset (MAD). It's designed for edge deployment on devices like Raspberry Pi 5 for military surveillance applications.

Key Features

• 🎯 96.73% accuracy on MAD dataset (7 military audio classes)
• 🚀 Edge-optimized for Raspberry Pi deployment
• ⚡ Fast inference (<200ms per sample)
• 🧠 Efficient (16.5% of parameters fine-tuned)
• 🔊 Robust to real-world military environments

Training Results

Progressive Training Performance:

• Phase 1 (Classifier only): 94.32% accuracy
• Phase 2 (Top 2 layers): 96.73% accuracy ← Best Model
• Phase 3 (Top 4 layers): 96.35% accuracy
• Phase 4 (Top 6 layers): 96.73% accuracy

Training Configuration:

• Method: Progressive unfreezing strategy
• Learning Rates: Conservative (1e-4 → 2e-5)
• Normalization: MAD-specific statistics (mean: -2.16, std: 2.85)
• Class Weighting: Balanced for imbalanced dataset
• Training Time: 40 minutes on RTX 3060

Model Classes

The model classifies 7 military audio categories:

Usage

Quick Start

from transformers import ASTForAudioClassification, ASTFeatureExtractor
import librosa
import torch

# Load model and feature extractor
model = ASTForAudioClassification.from_pretrained("Akashpaul123/tiny-ast-mad-military-audio-classifier")
feature_extractor = ASTFeatureExtractor.from_pretrained("Akashpaul123/tiny-ast-mad-military-audio-classifier")

# Load audio file (16kHz recommended)
audio, sr = librosa.load("military_audio.wav", sr=16000)

# Extract features
inputs = feature_extractor(audio, sampling_rate=16000, return_tensors="pt")

# Predict
with torch.no_grad():
    outputs = model(**inputs)
    predicted_class = torch.argmax(outputs.logits, dim=-1).item()

# Class mapping
classes = ['Communication', 'Footsteps', 'Gunshot', 'Shelling', 'Vehicle', 'Helicopter', 'Fighter']
print(f"Predicted class: {classes[predicted_class]}")

Edge Deployment (Raspberry Pi 5)

import onnxruntime as ort

# Load ONNX model for edge inference
session = ort.InferenceSession("tiny_ast_mad_optimized.onnx")
# ... inference code

Training Details

Dataset

• Source: Military Audio Dataset (MAD)
• Total Samples: 7,466 audio files
• Duration: 2-8 seconds per sample
• Sample Rate: 16kHz
• Augmentation: Military-specific (time stretch, pitch shift, noise injection)

Architecture

• Base Model: Audio Spectrogram Transformer (AST)
• Parameters: 86.2M total, 14.2M trainable (16.5%)
• Input: Log-Mel spectrograms (1024 x 128)
• Output: 7 military audio classes

Performance Metrics

• Accuracy: 96.73%
• F1-Macro: 96.84%
• F1-Weighted: 96.74%
• Precision: High across all classes
• Recall: Balanced performance

Hardware Requirements

Training

• GPU: RTX 3060 (12GB VRAM) or similar
• RAM: 16GB+ recommended
• Storage: 50GB for dataset and models

Inference (Edge)

• Device: Raspberry Pi 5 or similar ARM device
• RAM: 2GB minimum
• Inference Time: <200ms per sample
• Power: <5W continuous operation

Limitations and Considerations

• Domain-specific: Optimized for military audio contexts
• Language: Primarily English communication samples
• Environment: Trained on MAD dataset conditions
• Real-time: Designed for batch processing, not streaming

Citation

If you use this model in your research, please cite:

@misc{tiny-ast-mad-2024,
  title={Tiny-AST Military Audio Classifier: Progressive Fine-tuning for Edge Deployment},
  author={Paul, Akash},
  year={2024},
  howpublished={Hugging Face Model Hub},
  url={https://huggingface.co/Akashpaul123/tiny-ast-mad-military-audio-classifier}
}

License

This model is licensed under the Apache 2.0 License.

Contact

• Author: Akash Paul
• GitHub: @akashpaul123
• Hugging Face: @akashpaul123

---

Model trained as part of military audio surveillance research with focus on edge deployment and real-world robustness.