Abstract
This research investigates the mathematical distinction between ordinary speech and rhythmic recitation using non-linear dynamical analysis. By applying Detrended Fluctuation Analysis (DFA) to audio signals of Bengali poetry, the study identifies the Hurst exponent as a robust indicator of vocal performance style. The findings suggest a "phase transition" threshold at approximately 0.3, where the shift from flat reading to emotional recitation fundamentally alters the signal's underlying complexity.
Introduction
Human speech is a complex, non-stationary signal driven by intricate biological and cognitive processes. While listeners can intuitively distinguish between a monotone reading and an expressive recitation, quantifying this difference objectively remains a challenge in acoustic phonetics. Traditional linear features like pitch and intensity often fail to capture the holistic "texture" of emotional performance. This study proposes treating the transition from speech to recitation as a physical phase transition, seeking a critical mathematical value that defines this shift.
Methodology: Detrended Fluctuation Analysis
To analyze the non-linear characteristics of vocal signals, the researchers employed Detrended Fluctuation Analysis (DFA). This technique is particularly effective for detecting long-range correlations in non-stationary time series.
The process involves integrating the audio signal to create a "random walk" profile, dividing it into windows to remove local trends, and calculating the root-mean-square fluctuation. From this, the scaling exponent $\alpha$ (related to the Hurst exponent) is derived. An exponent of $\alpha = 0.5$ indicates uncorrelated white noise, while $\alpha > 0.5$ suggests positive autocorrelation and $\alpha < 0.5$ indicates an anti-correlated, more complex signal.
Experimental Setup
The study utilized a curated dataset of five Bengali poems representing diverse emotional genres: Fun (Nonsense Laughter), Bliss, Romance (Slow), Romance (Fast), and Sorrow. To control for speaker-specific variables, a single speaker recorded both a "flat" reading and a "rhythmic" recitation for each poem. The emotional validity of these recordings was standardized through a listening test involving 50 participants.
Results and Discussion
The analysis revealed a consistent pattern: the introduction of rhythm and emotion generally decreases the Hurst exponent, suggesting that recitation is a more "anti-correlated" or structured phase of speech compared to normal reading.
As illustrated in Figure 2, the Hurst exponent for normal reading typically resides above 0.3, while rhythmic recitation (with the notable exception of the "Sorrow" genre) drops below this value. For instance, in the "Fun" genre, the exponent shifted from ~0.31 during reading to ~0.21 during recitation.
The detailed data in Figure 6 confirms the stability of these measurements across different segments of the audio clips. The researchers propose a critical threshold of 0.3 as the point of phase transition. Values crossing this threshold indicate a fundamental change in the signal's non-linear dynamics, moving from the "speech phase" to the "recitation phase."
Interestingly, the "Sorrow" genre exhibited an inverse trend, where recitation actually increased the Hurst exponent (from ~0.37 to ~0.41). This suggests that the slow, heavy cadence of sorrowful performance may introduce different types of long-range correlations compared to more rhythmic genres.
Conclusion
This pilot study demonstrates that non-linear metrics like the Hurst exponent can effectively quantify the subjective experience of vocal performance. The identification of a 0.3 threshold provides a preliminary mathematical basis for distinguishing speech from recital. Future work will involve larger datasets and multiple speakers to determine if this "phase transition" value is a universal feature of human vocal expression across different languages and cultures.
Summary updated