Automatic Detection of Subharmonics in Acoustic Voice Signal

Objective: Among many possible modes of irregular vocal fold oscillation, those with subharmonics warrant special attention because both normal and pathological vocal folds can produce such voice. As such, a capable subharmonic detector can enhance the clinical acoustic analysis either by discriminating subharmonics from other irregularities or by identifying the abnormal attributes during subharmonic phonation. Currently, none of the available voice analysis software provides reliable subharmonic detection for acoustic signals to our knowledge. This presentation presents a novel model-based detector with minimum description length (MDL) information criterion and its performance in Monte Carlo experiment with synthetic voice signals.

Methods: The proposed algorithm is based on a mathematical model of a harmonic signal with the fundamental frequency ω0 in autoregressive (AR) noise. The harmonic signal represents both voice signal at the fundamental frequency fo without subharmonics (H0: ω0=2πfo) and those with subharmonics with period Ts (H1: ω0=2πfo/Ts, Ts>1). The AR noise represents the turbulent glottal flow with vocal tract effect. Given a good fo estimate, the detector generates N candidate models with Ts=1,2,…,N. The MDL of each candidate is computed, and the detector selects the model with the lowest MDL. Additional selection mechanisms are also proposed to minimize overestimation of Ts.

Monte Carlo experiments were conducted to evaluate the detector. Acoustic voice stimuli were numerically synthesized with a kinematic vocal fold model, which is aerodynamically and acoustically coupled to the wave-reflection trachea and vocal tract models. The subharmonics were introduced by amplitude modulating vocal fold displacements. A male voice (fo=100 Hz) with sustained /ɑ/ was considered for the study. The rate of correct modulation period classification, Pc, was used as the performance metric, and the changes in Pc were observed as the analysis duration and modulation extent and period were varied. Each outcome was evaluated from 1000 trials with randomized noise samples, initial and modulation phases.

Results and Conclusions: The proposed detector classified the subharmonic period (or absence) correctly for over 90% of the time with a 50-millisecond analysis duration. The observed level of performance makes the proposed detector a viable tool for acoustic analysis.