Genomic sequencing techniques introduce experimental errors into reads which can mislead sequence assembly efforts and complicate the diagnostic process. Here we present a method for detecting and removing sequencing errors from reads generated in genomic shotgun sequencing projects prior to sequence assembly. For each input read, the set of all length k substrings (k-mers) it contains are calculated. The read is evaluated based on the frequency with which each k-mer occurs in the complete data set (k-count). For each read, k-mers are clustered using the variable-bandwidth mean-shift algorithm. Based on the k-count of the cluster center, clusters are classified as error regions or non-error regions. For the 23 real and simulated data sets tested (454 and Solexa), our algorithm detected error regions that cover 99% of all errors. A heuristic algorithm is then applied to detect the location of errors in each putative error region. A read is corrected by removing the errors, thereby creating two or more smaller, error-free read fragments. After performing error removal, the error-rate for all data sets tested decreased (∼35-fold reduction, on average). EDAR has comparable accuracy to methods that correct rather than remove errors and when the error rate is greater than 3% for simulated data sets, it performs better. The performance of the Velvet assembler is generally better with error-removed data. However, for short reads, splitting at the location of errors can be problematic. Following error detection with error correction, rather than removal, may improve the assembly results.