Acoustic emission (AE) signals have tremendous promise for monitoring machining processes, but current understanding of AE generation is very limited. In addition, existing techniques to analyze these signals are inadequate. Presented here is a new approach to analyze AE signals based on thorough characterization of the dynamics underlying the signals. The signals were collected from extensive experimentation on five different lathes, performed under various cutting conditions, with different tool--work combinations and sampling rates. First, a battery of statistical tests was applied on the collected AE signals. Tests revealed that AE in turning is chaotic with fractal dimensions ranging between three and six. But owing to nonstationarity, chaos was not clearly revealed in the lag plots. Next, the transients were quantified, based on a compact representation scheme that was developed---called suboptimal wavelet packet representation---that captured all salient features of the AE signals. This representation scheme was used in developing a neural network-based estimator for on-line monitoring of tool wear.