Activity Recognition (AR) from smartphone sensors has be-come a hot topic in the mobile computing domain since it can provide ser-vices directly to the user (health monitoring, fitness, context-awareness) as well as for third party applications and social network (performance sharing, profiling). Most of the research effort has been focused on direct recognition from accelerometer sensors and few studies have integrated the audio channel in their model despite the fact that it is a sensor that is always available on all kinds of smartphones. In this study, we show that audio features bring an important performance improvement over an accelerometer based approach. Moreover, the study demonstrates the interest of considering the smartphone location for on-line context-aware AR and the prediction power of audio features for this task. Finally, an-other contribution of the study is the collected corpus that is made avail-able to the community for AR recognition from audio and accelerometer sensors.