An Optimized Signal Processing Pipeline for Syllable Detection and Speech Rate Estimation
Syllable detection is an important speech analysis task with applications in speech rate estimation, word segmentation, and automatic prosody detection. Based on the well understood acoustic correlates of speech articulation, it has been realized by local peak picking on a frequency-weighted energy contour that represents vowel sonority. While several of the analysis parameters are set based on known speech signal properties, the selection of the frequency-weighting coefficients and peak-picking threshold typically involves heuristics, raising the possibility of data-based optimisation. In this work, we consider the optimization of the parameters based on the direct minimization of naturally arising task-specific objective functions. The resulting non-convex cost function is minimized using a population-based search algorithm to achieve a performance that exceeds previously published performance results on the same corpus using a relatively low amount of labeled data. Further, the optimisation of system parameters on a different corpus is shown to result in an explainable change in the optimal values.