For regional nuclear test monitoring applications, determining accurate locations for small-to-intermediate events (2.5 < mb < 4.0) using sparse stations or arrays is an important and difficult problem. The primary objective of this study was to develop a new and innovative wavelet technique for accurate, reliable, and semiautomatic arrival-time determination, specifically for secondary phases in complex regional seismograms. Since relatively few P-wave observations will be available for hypocenter location, secondary seismic phases, such as Lg and Pg, must be considered to improve location accuracy to the level required for nuclear test monitoring. For these events, single array or single station locations may be used to determine an epicenter for an event, based on estimates of back azimuth and epicentral distance. Because Lg propagates at a consistent group velocity of approximately 3.5±0.2 km/s, the distance estimates using the travel-time difference between this arrival and primary phases (e.g., Pn, Pg) are often quite reliable. However, Lg is a very complicated phase and identifying its onset time on seismograms remains one of the most difficult problems encountered in regional seismology. We have tested an improved version of the Lg wavelet phase-picker (Wavelet 1.0), based on a method developed by Tibuleac and Herrin (1999, 2001), and two other phase-picking methods: CUMSQ (Inclan and Tiao, 1994; Der and Shumway, 1999) and AR (Taylor et al., 1992). We tested the three methods on Lg and Pg (as secondary arrivals) from a data set of 97 shallow (mb < 4.0) well located mine explosions and collocated aftershocks recorded at the TXAR (Lajitas, Texas), NVAR (Mina, Nevada), and PDAR (Pinedale, Wyoming) seismic arrays. Based on a rigorous statistical evaluation, Wavelet 1.0 estimates were more consistent for six out of seven clusters located 300-700 km from three seismic arrays, and produced the lowest location sample standard deviations in the 4.1-5.8 km range, while the other two methods had sample standard deviations well above 5.8 km. The 95% confidence intervals for location standard deviations were estimated as follows: standard deviation between 4.1 and 5.8 km for Wavelet 1.0 picks, between 6.9 and 9.5 km for the analyst picks, and between 5.8 and 8.0 km for CUMSQ picks. Our results show that Wavelet 1.0 offers an improvement in automatic secondary phase-picking over the CUMSQ and AR algorithms and has the potential to be developed into a powerful and robust automatic tool for routine operations.