The onset of a musical note is the earliest time at which a note can be reliably detected. Detection of these musical onsets pose challenges in the presence of ornamentation such as vibrato, bending, and if the attack of the note transient is slower. The legacy systems such as spectral difference or flux and complex domain functions suffer from the addition of false positives due to ornamentation posing as viable onsets. We propose that this can be solved by appropriately improving the resolution of the onset strength signal (OSS) and smoothening it to increase true positives and decrease false positives, respectively. An appropriate peak picking algorithm that works well in unison with the OSS generated is also desired. Since onset detection is a low-level process upon which many other tasks are built, computational complexity must also be reduced. We propose an onset detection alogrithm that is a combination of short-time spectral average-based OSS estimation, chirp group delay-based smoothening, and valley-peak distance-based peak picking. This algorithm performs on par with the state-of-the-art, superflux and convolutional neural networks-based onset detection, with an average F1 score of 0.88, across three datasets. Subsets from the IDMT-SMT-Guitar, Guitarset, and Musicnet datasets that fit the scope of the work, are used for evaluation. It is also found that the proposed algorithm is computationally 300\% more efficient than superflux. The positive effects of smoothening an OSS, in determining the onset locations, is established by refining the OSS produced by legacy algorithms, where consistent improvement in onset detection performance is observed. To provide insights into the performance of the proposed algorithms when different ornamentation styles are present in the recording, three levels of results are computed, by selecting different subsets of the IDMT dataset.