Paper: Algorithm Design and Mathematical Modeling for Efficient Automatic Speech Recognition in Low-Resource African Languages Author: Simanga Comfort Mchunu (47483571@mylife.unisa.ac.za), University of South Africa
This project benchmarks ASR models on three South African languages:
| Language | HuggingFace config | ISO |
|---|---|---|
| isiZulu | fleurs/zu_za |
zu |
| Setswana | fleurs/tn_za |
tn |
| Sesotho | fleurs/st_za |
st |
Three compression techniques are evaluated:
- Low-Rank Factorization – W ≈ U·Vᵀ, reduces parameters O(p·q) → O(r·(p+q))
- Knowledge Distillation – teacher → student transfer
- Post-training 8-bit Quantization – 4× memory reduction
asr_benchmark/
├── src/
│ ├── benchmark.py # Main benchmark runner
│ ├── compression.py # Mathematical compression implementations
│ └── visualize.py # Result tables (LaTeX + Markdown)
├── results/
│ └── benchmark_results.json # Auto-generated after run
├── configs/
│ └── default.json # Configurable run parameters
├── requirements.txt
└── README.md
python -m venv .venv && source .venv/bin/activate # (recommended)
pip install -r requirements.txt# Works without GPU or internet – uses synthetic audio
python src/benchmark.py --samples 5
# With a real HuggingFace model (downloads ~300 MB)
python src/benchmark.py \
--model openai/whisper-tiny \
--samples 10
# With compression enabled
python src/benchmark.py \
--model openai/whisper-tiny \
--samples 10 \
--low-rank \
--rank-ratio 0.25 \
--quantizepython src/visualize.pyOutputs a Markdown table and a ready-to-paste LaTeX \table{} block.
| Flag | Default | Description |
|---|---|---|
--model |
facebook/wav2vec2-base |
HuggingFace model ID |
--samples |
10 |
Samples per language |
--quantize |
off | Apply 8-bit quantization |
--low-rank |
off | Apply LRF (W ≈ UVᵀ) |
--rank-ratio |
0.25 |
Fraction of singular values kept |
--output |
results |
Output directory |
--device |
cpu |
cpu or cuda |
| Model | Size | Notes |
|---|---|---|
openai/whisper-tiny |
~75 MB | Best for replicating paper baseline |
openai/whisper-base |
~145 MB | Higher accuracy |
facebook/wav2vec2-base |
~360 MB | Requires fine-tuning for SA languages |
mbzuai-paris/Whisper-tiny-ZU |
~75 MB | Fine-tuned isiZulu |
| Metric | Definition |
|---|---|
| WER | Word Error Rate (%) |
| CER | Character Error Rate (%) |
| RTF | Real-Time Factor (latency / audio duration; <1 = faster than real-time) |
| Throughput | 1/RTF – multiples of real-time |
| Model size | Estimated MB from parameter count |
| Compression | Ratio vs 85 MB baseline |
W ∈ ℝ^(p×q)
Truncated SVD: W ≈ U_r · diag(σ₁...σ_r) · Vᵀ_r
Parameters: p·q → r·(p+q) where r << min(p,q)
m ≥ O(k log n) samples sufficient for PAC learning
k = vocabulary size, n = feature dimension
‖W - W_q‖_F / ‖W‖_F ≤ ε (bounded perturbation)
8-bit: 4× memory reduction from fp32
@misc{mchunu2026asr,
title = {Algorithm Design and Mathematical Modeling for Efficient
Automatic Speech Recognition in Low-Resource African Languages},
author = {Mchunu, Simanga Comfort},
year = {2026},
institution = {University of South Africa}
}