Abstract
Traditional speech processing relies on linear assumptions that often overlook the inherent complexity of human vocal production. This paper proposes a non-linear approach to emotion recognition using multifractal analysis. By applying Multi-Fractal Detrended Fluctuation Analysis (MFDFA) and Wavelet Transform Modulus Maxima (WTMM), we extract emotional "fingerprints" from speech signals. Our model, validated on English and German datasets, achieves a 96.3% average accuracy in distinguishing happy, sad, and neutral states, suggesting that non-linear features provide a robust, language-independent basis for emotion quantification.
Introduction
Standard emotion recognition systems typically utilize Mel-frequency cepstrum coefficients (MFCCs) or Linear Prediction Cepstrum Coefficients (LPCCs). These methods assume the human vocal tract is a linear, time-varying system. However, speech production is fundamentally non-linear and chaotic, involving complex air-puff oscillations and time-varying resonances. Linear models often discard high-frequency information and subtle complexities essential for emotional nuance. Furthermore, these systems frequently struggle with cross-linguistic generalization. This research seeks a mathematical "fingerprint" of emotion that remains consistent across different languages.
Methodology: Non-linear Multifractal Analysis
The researchers employed Non-linear Multifractal Analysis to quantify signal complexity. Two primary frameworks were used:
- Multi-Fractal Detrended Fluctuation Analysis (MFDFA): Detects long-range correlations in non-stationary time series to calculate a "singularity spectrum."
- Wavelet Transform Modulus Maxima (WTMM): Uses wavelets to identify local maxima across scales, effectively uncovering the multifractal distribution.
Statistical features, such as the area under the curve in the "low fluctuation region" and spectral width, were extracted from these spectra and fed into a Support Vector Machine (SVM) with a linear kernel for classification.
Experimental Setup
The model's language independence was tested using two databases:
- EMO-DB (German): Ten actors performing sentences in seven emotional states.
- TESS (English): 2,800 voice samples from two actresses across various emotional contexts.
The study focused on three core states: Happiness, Sadness, and Neutral.
Results and Discussion
The analysis confirmed that speech is "strongly multifractal," with a spectral width consistently exceeding 0.5. A key finding was that emotions are most distinguishable in the "low fluctuation" region of the multifractal spectrum. In this region, happy speech typically occupies low Holder exponent values, sad speech occupies high values, and neutral speech falls in between.
The WTMM method generally outperformed MFDFA, providing clearer segregation of emotional attributes. The system achieved an average classification accuracy of 96.3%. In a specific test run of 36 clips (12 per emotion), the model reached 97.2% accuracy.
Conclusion
This study demonstrates that non-linear, multifractal modeling of speech provides a highly accurate and language-independent method for emotion recognition. While the pilot study focused on three primary emotions in controlled environments, future work will expand to include more nuanced emotional states (e.g., fear, boredom) and test the model's robustness in noisy, real-world settings and across tonal languages.
Summary updated