Massively parallel (next-generation) sequencing provides a powerful method to analyze DNA from many different sources, including degraded and trace samples. A common challenge, however, is that many forensic samples are often known or suspected mixtures of DNA from multiple individuals. Haploid lineage markers, such as mitochondrial (mt) DNA, are useful for analysis of mixtures because, unlike nuclear genetic markers, each individual contributes a single sequence to the mixture. Deconvolution of these mixtures into the constituent mitochondrial haplotypes is challenging as typical sequence read lengths are too short to reconstruct the distinct haplotypes completely. We present a powerful computational approach for determining the constituent haplotypes in massively parallel sequencing data from potentially mixed samples. At the heart of our approach is an expectation maximization based algorithm that co-estimates the overall mixture proportions and the source haplogroup for each read individually. This approach, implemented in the software package mixemt, correctly identifies haplogroups from mixed samples across a range of mixture proportions. Furthermore, our method can separate fragments in a mixed sample by the most likely originating contributor and generate reconstructions of the constituent haplotypes based on known patterns of mtDNA diversity.