Next-generation DNA sequencing technologies, such as RNA-Seq, currently dominate genome-wide gene expression studies. A standard approach to analyse this data requires mapping sequence reads to a reference and counting the number of reads which map to each gene. However, for many transcriptome studies, a suitable reference genome is unavailable, especially for meta-transcriptome studies which assay gene expression from mixed populations of organisms. Where a reference is unavailable, it is possible to generate a reference by the de novo assembly of the sequence reads. However, the high cost of generating high-coverage data for de novo assembly hinders this approach and more importantly the accurate assembly of such data is challenging, especially for meta-transcriptome data, and resulting assemblies frequently suffer from collapsed regions or chimeric sequences. As an alternative to the standard reference mapping approach, we have developed a k-mer-based analysis pipeline (DiffKAP) to identify differentially expressed reads between RNA-Seq datasets without the requirement for a reference. We compared the DiffKAP approach with the traditional Tophat/Cuffdiff method using RNA-Seq data from soybean, which has a suitable reference genome. We subsequently examined differential gene expression for a coral meta-transcriptome where no reference is available, and validated the results using qRT-PCR. We conclude that DiffKAP is an accurate method to study differential gene expression in complex meta-transcriptomes without the requirement of a reference genome.