VynoDePal/afri-voices-omniasr-kenya

# Afri-Voices omniASR for six Kenyan languages This repository documents our end-to-end development of a six-language automatic speech recognition system for the Afri-Voices East Africa ASR Hackathon. We began with joint and language-specialized omniASR checkpoints, explored balanced fine-tuning, checkpoint merging, low-rank approximations and LoRA, and converged on one BF16 acoustic base, one Maasai LoRA adapter, and language/domain-routed KenLM decoding. The final audited system has **985,573,552 active neural parameters**, stays below the **8 GiB edge-memory limit**, and obtained a public leaderboard WER of **0.36585**. The repository contains the complete research notebooks, including failed experiments, plus a deterministic K1–K6 language-model and inference pipeline. It contains no audio, competition test data, transcripts, credentials, model weights, KenLM binaries, or submission files. ## Final result | Submission | Acoustic system | Decoder | Status | Public LB WER | |---|---|---|---|---:| | Historical exploration | Multiple specialist checkpoints | KenLM V5 | Exploratory; not the final single-artifact system | **0.36529** | | First edge-audited system | Step-1250 BF16 base + Maasai LoRA | Production domain routing | Under 1B and under 8 GiB | 0.36878 | | **Final K6 system** | Same audited acoustic system | Hybrid production/KenLM V6 routing | **Under 1B and under 8 GiB** | **0.36585** | The historical 0.36529 score is reported for completeness. It relied on multiple specialized acoustic checkpoints and was not retained under our strict interpretation of the single-model, total-parameter, and edge-memory constraints. The 0.36585 submission is our best fully audited, reproducible configuration. KenLM V6 improved the compliant 0.36878 system by **0.00293 absolute WER** (approximately **0.8% relative**). ## Task and metric The task covers six languages: Swahili (`swa`/`sw`), Kikuyu (`kik`), Kalenjin (`kln`), Dholuo (`luo`), Somali (`som`), and Maa …