Abstract
At present Automatic Speaker Recognition system is a very important issue due to its diverse applications. Hence, it becomes absolutely necessary to obtain models that take into consideration the speaking style of a person, vocal tract information, timbral qualities of his voice and other congenital information regarding his voice. The study of Bengali speech recognition and speaker identification is scarce in the literature. Hence the need arises for involving Bengali subjects in modelling our speaker identification engine. In this work, we have extracted some acoustic features of speech using non linear multifractal analysis. The Multifractal Detrended Fluctuation Analysis reveals essentially the complexity associated with the speech signals taken. The source characteristics have been quantified with the help of different techniques like Correlation Matrix, skewness of MFDFA spectrum etc. The Results obtained from this study gives a good recognition rate for Bengali Speakers.
Introduction
Automatic Speaker Recognition (ASR) systems are increasingly vital for security, forensics, and personalized technology. Most contemporary ASR systems rely on linear statistical characteristics, such as Mel-frequency Cepstral Coefficients (MFCCs), which transform signals into the frequency domain using Fourier analysis. However, human speech is fundamentally non-stationary and chaotic. Standard linear methods often overlook high-frequency harmonics and the geometric complexity inherent in a speaker's voice.
Furthermore, while speaker recognition is well-documented for English, research involving the Bengali language remains limited. This study addresses this gap by applying nonlinear dynamical modeling—specifically Multifractal Detrended Fluctuation Analysis (MFDFA)—to Bengali speech, achieving a 96% identification accuracy in a pilot study.
Methodology
The core of this research utilizes Multifractal Detrended Fluctuation Analysis (MFDFA) to capture the fractal geometry of speech signals. Unlike traditional methods that focus on pitch or volume, MFDFA examines how a signal scales across different time intervals.
The methodology follows several key steps:
- Signal Transformation: The raw, noise-like speech signal is converted into a "random walk" signal to facilitate fluctuation analysis.
- Fluctuation Calculation: Local Root Mean Square (RMS) variation is determined across various segment sizes.
- Singularity Spectrum Generation: A multifractal spectrum, $f(\alpha)$, is generated. This spectrum describes the statistical distribution of the signal's local scaling exponents ($\alpha$), representing the complexity of the vocal tract's output.
- Feature Extraction: Two primary statistical features are derived from the $f(\alpha)$ spectrum:
- Feature 1: The difference between the median and the mode of the spectrum.
- Feature 2: The skewness (asymmetry) of the spectrum.
- Classification: These nonlinear features are processed by a Support Vector Machine (SVM) using a Radial Basis Function (RBF) kernel to identify individual speakers.
Experimental Setup and Dataset
The study utilized a custom-built Bengali speech corpus consisting of:
- Participants: 5 speakers reading naturally without emotional inflection.
- Content: 5 paragraphs selected from popular Bengali novels.
- Data Volume: 100 total samples (20 samples per speaker).
- Technical Specs: Signals were digitized at 22,050 samples/sec in 16-bit format, with each phrase lasting approximately 10 seconds.
Results and Discussion
The SVM classifier achieved an impressive overall accuracy of 96%. The researchers validated these results by generating a 100x100 Cross-Correlation Matrix, which showed high correlation coefficients for samples from the same speaker and significantly lower coefficients between different speakers.
The "shape" of the MFDFA spectrum—its width, peak, and skew—acts as a robust biometric signature. While the system was nearly perfect for most participants, Speaker 2 showed a lower individual accuracy (80%), occasionally being confused with Speaker 4. This suggests a high degree of nonlinear similarity in their vocal profiles, which might be missed by traditional linear analysis.
Conclusion
This research demonstrates that nonlinear features extracted via MFDFA provide a powerful tool for speaker recognition in the Bengali language. By treating the human vocal tract as a complex system producing multifractal output, the model captures "congenital information" unique to each individual. Future work will focus on expanding the dataset to include a more diverse population and optimizing the computational efficiency of fractal analysis for real-time applications.
Summary updated