The majority of the world's 7,000 languages lack a standardized writing system. In this paper we consider one such language, Bambara, which is rarely written down but is widely spoken in Mali and neighboring countries. We explore the task of using automatic speech recognition (ASR) to transcribe culturally significant recordings focused on two domains: archival linguistic and anthropological fieldwork and contemporary oral histories performed by griots, the traditional Mande history keepers. We describe our two 6.5-hour corpora then experiment with different data configurations and multi-stage tuning from pretrained multilingual models within two neural ASR architectures. We find that while the diversity in content, style, and recording quality across the two corpora presents challenges, their commonalities can sometimes be leveraged to improve ASR accuracy. We note, however, that the diverse qualities of these corpora diminish their utility for cross-domain ASR training.