Noisy Speech Based Temporal Decomposition to Improve Fundamental Frequency Estimation

This paper introduces a novel method to separate voiced frames of noisy speech signals into low-frequency or high-frequency. separation improves the accuracy fundamental frequency (F0) estimators. In this proposal, target signal is analyzed by means ensemble empirical mode decomposition. Next, pitch information extracted from first decomposition modes. feature indicates region where F0 should be located, thus separating The then applied correct candidates detection method, improving estimation accuracy. proposed and baseline approach are evaluated considering four different algorithms. Experiments conducted with CSTR TIMIT databases, six noises various signal-to-noise ratios. Gross Error (GE) Mean Absolute (MAE) metrics adopted evaluate solutions in terms errors. Results show that outperforms baseline, low/high Moreover, solution able better improve under conditions.