Introduction.

Deploying Automatic Speech Recognition (ASR) for African languages is critically
hindered by the incompatibility of large-scale models with mobile hardware. Models like
NLLB exceed 1 GB in size and cannot load on typical devices with 2-4 GB RAM, preventing
real-time conversational ASR deployment in resource-constrained settings. This technological
barrier disproportionately affects African linguistic communities.
Methods.

We formulate ASR deployment as a constrained optimization problem: minimize
Word Error Rate (WER) subject to bounds on model size, latency, and computational complexity.
Our solution integrates three complementary techniques: (1) Knowledge Distillation,
transferring knowledge from large teacher networks to compact students; (2) Low-Rank Factorization
(W ≈ UVT ), reducing parameters from O(p) to O(rp); and (3) Post-training 8-
bit Quantization for 4× memory reduction. Theoretical analysis provides sample complexity
bounds of O(k log n) and quantifies output perturbation from compression.
Results.

On a 20M-parameter baseline (85 MB), our framework achieves 16× total compression
to approximately 5 MB. Across iSiZulu, Setswana, and Sesotho datasets, mainly South African langauges, we maintain competitive
WER with only 5-6% degradation while increasing inference throughput from 0.8× to 6× in real time
on commodity CPUs. This 7.5× latency improvement enables previously impossible mobile
deployment.
Conclusions.

This work demonstrates that algorithm design grounded in mathematical optimization
and hardware-aware compression enables scalable, practical ASR for underserved languages.
The framework provides a general blueprint for efficient edge AI deployment, directly
advancing speech technology accessibility and digital inclusion for African linguistic communities
globally.