I just dealt with a similar problem with semi-random 30 bp barcodes. I found this recent review extremely helpful, as it was my first time dealing with random barcodes. Barcode synthesis is an imperfect process, especially for runs of homopolymers or dinucleotides - indels are extremely common. Barcode correction is necessary if you want accurate counts.

My approach to look at engraftment efficiency of a cell line in a xenograft mouse model was basically:

• Take a look at the reads and trim primer/adapter sequences via cutadapt, leaving only the supposed barcode.
• Correct barcodes & count with starcode.
• For each sample, rank barcodes by count and consider the barcodes accounting for the first 90% of cumulative reads as "reasonably expressed". In practice, this results in most barcodes with very few counts being ignored:

Adjust as necessary for your application and experimental setup.