AfriVoices Edge ASR — six East African languages, one edge-ready model

Final result: macro-WER 0.34778 — 1st place (public leaderboard at close), AfriVoices Multilingual Edge ASR Track.

One 971 M-parameter acoustic model (omniASR-CTC-1B v2, Meta, Apache-2.0, fine-tuned on ~1,035 h of deduplicated speech) covering Swahili, Kikuyu, Dholuo, Somali, Maasai and Kalenjin, decoded with one purified KenLM per language (shallow fusion). Runs on 4 ARM cores at ~0.57x real time within 5.8 GB RAM.

What is in this repository

File(s)	What it is
`checkpoint_step12750/model/pp_00/tp_00/sdp_00.pt`	the fine-tuned acoustic model (fairseq2 state dict, bf16)
`omni_ctc_int8.onnx`	int8 ONNX export for edge/CPU inference
`kenlm/.bin` + `kenlm/_unigrams.txt`	six purified per-language KenLM models + word lexicons

How to use it

Full pipeline, configs, decoding parameters, training recipe, technical report and the exact 3-step reproduction of the winning submission:

➡️ https://github.com/Izzoudine/Afrivoices-Edge-Asr

Quick outline: load the checkpoint over the omniASR-CTC-1B-v2 card (fairseq2), decode with pyctcdecode beam search + the per-language KenLM at the alpha/beta/beam values committed in the GitHub repo (configs/decode_params.json); clips longer than 38 s are decoded whole (single pass) for best accuracy.

Data, transparency, licenses

Training data: AfriVoices / KenCorpus datasets (Anv-ke, Digital Umuganda) — see the data cards in the GitHub repo.
The KenLM models are purified: a leak of test transcriptions found in the provided dev_test splits was removed and disclosed (TRANSPARENCY_NOTE in the repo).
Base model: Omnilingual ASR (Meta), Apache-2.0. This release: Apache-2.0.

Downloads last month: -; Downloads are not tracked for this model. How to track