Multi-band summary correlogram-based pitch detection for noisy speech

Online Monaural Speech Enhancement Based on Periodicity Analysis and A Priori SNR Estimation

“…Previous research [20] showed that combining the methods of autocorrelation and AMDF can improve the accuracy of pitch detection and estimation for noisy speech. Recently Tan and Alwan [21] proposed a multi-band summary correlogram (MBSC) based pitch detection algorithm for noisy speech. The MBSC algorithm calculated the harmonic-to-subharmonic energy ratio (HSR) by comb filter in frequency domain and used this ratio to weight the autocorrelation to achieve a peak-enhanced summary correlogram and improve the pitch detection.…”

“…proposed method. (e) P 0 detection result by a recently published method [21]. The circles denote the ground truth.…”

“…Before evaluating the whole enhancement effect of the proposed algorithm, the P 0 detection part is evaluated in comparison to a recently published multi-band summary correlogram-based (MBSC) algorithm [21]. The MBSC algorithm was compared with several representative algorithms in [21] and was shown to perform best. Therefore, it is used as a benchmark here.…”

“…The error rate is calculated as the percentage of misses (when a periodic frame is detected as aperiodic), false alarms (when an aperiodic frame is detected as periodic), and deviations (when the difference between the detected F 0 and the true F 0 is larger than 20% of the true F 0 ). This calculation method is the same as that in [21]. Panels (a), (b), and (c) show the results in car, train, and babble noise, respectively.…”

“…The true P 0 of each frame (the circles) is also shown in the figure. For comparison, the detected result (as the dots) of a recently published multi-band summary correlogram-based (MBSC) pitch detection algorithm [21] is also shown in panel (e). The P 0 detection error rate (defined at part A of section III) is 23.4% for the proposed algorithm and 32.2% for the MBSC algorithm for this example.…”