Numerical-audio-authentication-system_Deep-learning

This repository implements a deep learning-based voice number authentication system using CNN and a Siamese Network. It verifies spoken numbers by comparing voice embeddings to reference samples. The model extracts audio features (MFCC, spectrogram) using CNN and determines similarity through a Siamese architecture.

Use the kaggle dataset for this .Mainly do the compares the two audio file in same person same number different audio to training the model.

inference

📌 Features

• Audio Preprocessing: Extracts MFCCs, Waveforms from numerical audio inputs.

• Siamese Network Architecture: Uses a twin neural network to compute the similarity between two audio samples.

• Triplet Loss / Contrastive Loss: Optimized for better feature embedding and verification accuracy.

• Dataset Handling: Supports labeled audio datasets with numerical recordings given by kaggle.datset link

• Training & Evaluation: Implements robust training with data augmentation and real-time validation.

  • Training stat
    

  • Ealuation report

  • )

                    precision    recall  f1-score   support
    
          real       0.92      0.94      0.93      1457
          fake       0.95      0.92      0.93      1543
    
      accuracy                           0.93      3000
     macro avg       0.93      0.93      0.93      3000
    

    weighted avg 0.93 0.93 0.93 3000

• Inference & Authentication:This uses two input audios that may or may not match the two audio files.

• Deployment Ready: Can be integrated into real-world authentication systems.

  
  
  

🛠️ Setup & Installation

1.Clone the repository:

git clone https://github.com/KaushiML3/Numerical-audio-authentication-system_Deep-learning.git
cd Numerical-Audio-Authentication-System

2.Install dependencies:

pip install -r requirements.txt

3.Run inference

• change the direction for API folder

python main.py