Overview
A Whisper large-v3 LoRA fine-tune for Nigerian speech recognition, covering Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian-accented English.
NaijaVox-V1 is a fine-tuned Whisper large-v3 model (via LoRA) built under Axiveri for Nigerian multilingual speech recognition — trained on a dual-T4 Kaggle setup, covering Yoruba, Hausa, Igbo, Nigerian Pidgin, and Nigerian-accented English.
Along the way this involved resolving real production ML bugs: a dtype mismatch between float32 inputs and fp16 weights under DataParallel, a gradient-checkpointing incompatibility (resolved via use_reentrant=False), a silent augmentation flag bug, and a hardcoded dtype cast at eval time.