Researchers present an open-source automatic speech recognition (ASR) solution designed to assess children's reading in Bambara, addressing the lack of developed tools for this African language. The system was built through an end-to-end process involving field data collection, benchmark construction, model adaptation, and classroom validation.

  • A mobile app collected 55 hours of raw reading speech from 60 children to create a public benchmark for Bambara child-reading assessment.
  • Fine-tuning experiments compared Soloni, a Bambara-adapted Fast-Conformer ASR framework, with QuartzNet, a compact convolutional ASR architecture.
  • The best Soloni model reduced Word Error Rate (WER) from 0.42 to 0.22 and Character Error Rate (CER) from 0.15 to 0.08, substantially outperforming QuartzNet.
  • Disaggregated analysis identified children under 10 as the main source of residual errors, motivating targeted data collection from younger readers.
  • Ten classroom trials supported the continued use of the application for literacy assessment.

The work provides a reproducible framework for literacy assessment in low-resource languages and highlights specific architectural benefits and data needs for improving ASR performance on child speech.