Abstract
This paper presents a robust and secure framework for text-independent speaker recognition designed to operate effectively in low Signal-to-Noise Ratio (SNR) environments. The proposed system addresses the dual challenges of environmental noise and biometric data security by integrating a three-layer encryption scheme with a modified autocorrelation algorithm for pitch extraction. By combining Pseudo-Random Noise Generators (PRNG) and Discrete Cosine Transforms (DCT) with Lagrange interpolation for precise pitch estimation, the system achieves a high identification accuracy of 92.39% while ensuring that biometric templates remain protected from unauthorized access and tampering.
Introduction
Speaker recognition systems are increasingly deployed in real-world scenarios ranging from forensic analysis to secure access control. However, two significant hurdles limit their widespread adoption: environmental noise and data vulnerability. Most recognition systems rely on pitch estimation, which tends to degrade rapidly in noisy conditions. Furthermore, the storage and transmission of biometric voice templates pose a security risk; if compromised, these templates can be used for identity spoofing. This research proposes a "noise-hardened" and "cryptographically secure" architecture that maintains high recognition performance even when the signal is distorted by noise or intercepted by malicious actors.
Methodology
1. Secure Multi-Level Encryption Framework
To protect sensitive biometric data, the system employs a three-level "sandwich" encryption process that secures the voice signal in both the time and frequency domains:
- Level 1 (Time Domain Scrambling): A user-defined 8-character password is processed via a Caesar cipher to generate two state keys (Key1 and Key2). Key1 seeds a Pseudo-Random Noise Generator (PRNG) to scramble the original voice signal, obscuring its temporal characteristics.
- Level 2 (Frequency Domain Transformation): The scrambled signal is transformed using the Discrete Cosine Transform (DCT). DCT is utilized because it uncorrelates the data, making the signal highly resistant to cryptanalytic attacks that exploit the natural redundancy of human speech. In this domain, the signal becomes effectively meaningless to an eavesdropper.
- Level 3 (Coefficient Scrambling): Key2 seeds a second PRNG layer that adds noise to the DCT coefficients. This final layer ensures that even if the frequency domain data is intercepted, it remains indecipherable without the correct cryptographic keys.
2. Robust Pitch Extraction via Modified Autocorrelation
For the recognition task, the authors developed a modified autocorrelation pitch extraction algorithm designed for resilience in low SNR environments. The process involves two primary optimizations:
- Computational Efficiency: The system incorporates the Average Magnitude Difference Function (AMDF). Unlike standard autocorrelation, which is mathematically intensive, AMDF replaces multiplications with subtractions, significantly reducing the computational load for real-time processing.
- Precision via Interpolation: To combat the flattening or shifting of pitch peaks caused by noise, the algorithm uses Lagrange interpolation. By applying 3-point interpolation around detected peaks of the autocorrelation function, the system can reconstruct the true fundamental frequency (F0) with much higher precision than standard methods.
Experimental Setup
Datasets and Noise Profiles
The system was evaluated using a custom speaker database containing voice samples from 50 subjects. Each subject provided six samples (three for training and three for testing) recorded with varying tones and volumes. To simulate real-world conditions, noise profiles including babble, car, and street noise were sourced from the NOISEX-92 database. Additionally, Additive White Gaussian Noise (AWGN) was applied to simulate transmission interference at SNR levels ranging from 16 dB to 20 dB.
Results and Discussion
Recognition Accuracy
The proposed modified autocorrelation method demonstrated superior performance compared to traditional feature extraction techniques. The system achieved an identification accuracy of 92.39%, significantly outperforming the following baselines:
- Statistical Methods: 74.87%
- Linear Predictive Coding (LPC): 72.42%
- Zero Crossing: 62.35%
- Fast Fourier Transform (FFT): 55.49%
Decryption Integrity
The reliability of the encryption-decryption cycle was measured using Mean Square Error (MSE). As the SNR increased, the MSE of the decrypted signal dropped exponentially. At 20 dB, the MSE reached a negligible $7.32 \times 10^{-7}$, indicating that the original biometric features can be recovered with near-perfect integrity for recognition purposes, provided the correct keys are used.
Conclusion and Future Work
This research demonstrates that it is possible to achieve high-accuracy speaker recognition without sacrificing data security. The combination of DCT-based encryption and modified autocorrelation provides a robust defense against both environmental noise and cyber threats.
Future work will focus on scaling the system to larger, standardized online databases to test its performance across thousands of subjects. Additionally, the authors aim to optimize the algorithmic framework for real-time, live-streamed applications and explore the use of longer encryption keys to ensure long-term resistance to brute-force attacks.
Summary updated