Although the past decade has witnessed rapid advancements in automatic speech recognition (ASR) technologies, how prosodic variation impacts errors and why systems return degraded performance for African American English (AAE) speech are still not well-understood. The present study conducted two sets of analysis to offer insights into these issues. First, we computed seven quantitative measures of rhythmic variation (%V, ΔC, ΔV, VarcoC, VarcoV, nPVI-V, and rPVI-C) in a reading task produced by AAE speakers and tested the effect of articulation rate on these metrics. The results reveal intricate interactions between articulation rate and prosodic rhythm as found in non-AAE speech, while at the same time showing timing properties specific to AAE speech. We then examined the seven metrics and their relationships to word error rates. Results show that utterances exhibiting shorter %V and greater VarcoV values had higher error rates. We argue that shorter %V and greater VarcoV can be explained through vowel reduction in unstressed vowels, repetition reduction effect, and monophthongal /aɪ/ and /ɔɪ/, a well-documented AAE feature that may contribute to recognition errors. The results suggest that adding rhythmic variation to ASR acoustic models can provide additional information for developers interested in mitigating racial bias in voice technology.