Knowledge distillation bridges the gap for Nigerian language speech recognition

Automatic speech recognition systems struggle with Nigerian languages, typically achieving much lower accuracy than they do with English or French. Researchers have now developed a framework that substantially narrows this performance gap by combining knowledge distillation with iterative refinement using pseudo-labeled data.

The challenge#

Nigerian languages present a constellation of technical hurdles for ASR systems. The languages have limited training data available, lack standardized writing conventions, rely on tonal diacritics to distinguish meaning, and feature diverse regional accents. Code-switching—when speakers mix languages within a single utterance—and localized place names further complicate the task.

A two-stage distillation approach#

The team built their system through successive rounds of knowledge transfer. Initially, they transferred learning from established monolingual models into a multilingual framework, guided by language-specific N-gram language models. They then refined the system further by iteratively retraining on pseudo-labeled data—predictions the model made with high confidence on unlabeled audio.

Results and coverage#

The method achieved an average relative Word Error Rate reduction of 29% compared to monolingual baselines and surpassed existing multilingual models on standard benchmarks like Common Voice and FLEURS. The team introduced Sometin Beta Pass Notin (SBPN), a multilingual foundational model covering Yorùbá, Hausa, Igbo, Nigerian Pidgin, and Nigerian English.

Research details#

The work was submitted to arXiv in May 2026 and revised in September 2026.