The development of machine learning technology enables more robust real-time earthquake monitoring through automated implementations. However, the application of machine learning to earthquake location problems faces challenges in regions with limited available training data. To address the issues of sparse event distribution and inaccurate ground truth in historical seismic datasets, we expand the training dataset by using a large number of synthetic envelopes that closely resemble real data and build an earthquake location model named ENVloc. We propose an envelope-based machine learning workflow for simultaneously determining earthquake location and origin time. The method eliminates the need for phase picking and avoids the accumulation of location errors resulting from inaccurate picking results. In practical application, ENVloc is applied to several data intercepted at different starting points. We take the starting point of the time window corresponding to the highest prediction probability value as the origin time and save the predicted result as the earthquake location. We apply ENVloc to observed data acquired in the southern Sichuan Basin, China, between September 2018 and March 2019. The results show that the average difference with the catalog in latitude, longitude, depth, and origin time is 0.02°, 0.02°, 2 km, and 1.25 s, respectively. These suggest that our envelope-based method provides an efficient and robust way to locate earthquakes without phase picking, and can be used in earthquake monitoring in near-real time.