Reference-quality genomes are expected to provide a resource for studying gene structure, function, and evolution. However, often genes of interest are not completely or accurately assembled, leading to unknown errors in analyses or additional cloning efforts for the correct sequences. A promising solution is long-read sequencing. Here we tested PacBio-based long-read sequencing and diploid assembly for potential improvements to the Sanger-based intermediate-read zebra finch reference and Illumina-based short-read Anna’s hummingbird reference, 2 vocal learning avian species widely studied in neuroscience and genomics. With DNA of the same individuals used to generate the reference genomes, we generated diploid assemblies with the FALCON-Unzip assembler, resulting in contigs with no gaps in the megabase range, representing 150-fold and 200-fold improvements over the current zebra finch and hummingbird references, respectively. These long-read and phased assemblies corrected and resolved what we discovered to be numerous misassemblies in the references, including missing sequences in gaps, erroneous sequences flanking gaps, base call errors in difficult-to-sequence regions, complex repeat structure errors, and allelic differences between the 2 haplotypes. These improvements were validated by single long-genome and transcriptome reads and resulted for the first time in completely resolved protein-coding genes widely studied in neuroscience and specialized in vocal learning species. These findings demonstrate the impact of long reads, sequencing of previously difficult-to-sequence regions, and phasing of haplotypes on generating the high-quality assemblies necessary for understanding gene structure, function, and evolution.© The Authors 2017. Published by Oxford University Press.