Given the high data volumes in time series applications, or simply the need for fast response times, it is usually necessary to rely on alternative, shorter representations of time series, usually with information loss. This incurs approximate comparisons of time series where precision is a major issue. We propose a new representation approach called ASAX, coming with two techniques ASAX EN and ASAX SAE, for segmenting time series before their transformation into symbolic representations. Our solution can reduce significantly the error incurred by possible splittings at different steps of the representation calculation, by taking into account the entropy of the representations (ASAX EN) or the sum of absolute errors (ASAX SAE), particularly for datasets with unbalanced (non-uniform) distributions. This is particularly useful for time series similarity search, which is the core of many data analytics tasks. We provide theoretical guarantees on the lower bound of similarity measures, and our experiments illustrate that our approach can improve significantly the time series representation quality.