Pricile/afrivoices-edge-asr

wbertcur — Unified Edge ASR for Six East African Languages

A single, edge-deployable speech-to-text model for six East African languages, built for the AfriVoices East Africa ASR Hackathon (Multilingual Edge ASR track).

Languages	Swahili (swa), Kikuyu (kik), Dholuo (luo), Kalenjin (kln), Maasai (mas), Somali (som)
Backbone	facebook/w2v-bert-2.0 (Conformer, 580 M params, MIT)
Head	character-level CTC
Weights	included in this repo (checkpoint wbertcur-step-16500)
Decoder	per-language 4-gram KenLM (shallow fusion, pyctcdecode, beam 100)
License	MIT

Edge compliance (validated)

Requirement	Limit	Measured	Status
Parameters	< 1 B	0.58 B	✅
Peak RAM	≤ 8 GB	5.53 GB (single language, int8)	✅
Compute	CPU only	CPU only	✅
Latency (RTF)	≤ 2.0	0.091 (int8, KenLM beam 100, end-to-end)	✅

See HARDWARE_VALIDATION_REPORT.md in this repo for the full methodology.

Results

Configuration	Dev macro-WER	Kaggle (public)
Greedy (no LM)	0.3547	—
+ web-augmented 4-gram KenLM + per-language α/β	0.2866	0.37236

Metric = unweighted mean of per-language WER (macro).

Usage

import numpy as np, torch, pyctcdecode, soundfile as sf
from transformers import Wav2Vec2BertForCTC, Wav2Vec2BertProcessor

REPO = "Pricile/afrivoices-edge-asr"      # self-contained: weights + LMs in this repo
model = Wav2Vec2BertForCTC.from_pretrained(REPO).eval()
proc  = Wav2Vec2BertProcessor.from_pretrained(REPO)

# character vocabulary sorted by index (required by pyctcdecode)
vocab = [t for t, i in sorted(proc.tokenizer.get_vocab().items(), key=lambda x: x[1])]
# per-language 4-gram KenLM binaries are in this repo under lm/ (huggingface_hub.hf_hub_download)
decoder = pyctcdecode.build_ctcdecoder(vocab, kenlm_model_path="lm/som_final_4gram.binary",
                                       alpha=0.5, beta=2.0)   # per-language α/β (see code/ab.json)

audio, sr = sf.read("clip.wav")     # 16 kHz mono
feats = proc.feature_extractor(audio, sampling_rate=16000, return_tensors="pt").input_features
with torch.no_grad():
    logits = model(feats).logits[0].numpy()
print(decoder.decode(logits, beam_width=100))

For CPU edge deployment, apply dynamic int8 quantisation:

torch.quantization.quantize_dynamic(model, {torch.nn.Linear}, dtype=torch.qint8, inplace=True)

Training (summary)

Fine-tuned from scratch on top of w2v-bert-2.0 with a character-level CTC head, using a length curriculum (max audio 15 s → 30 s → 45 s → 60 s) and language re-weighting (kln×4, mas×4, som×2) since the macro metric is dominated by the low-resource languages. LR 5e-5, cosine schedule, effective batch 32, bf16, EMA (0.9999), seed 42; best at step 16 500. Full methodology and ablations in TECHNICAL_REPORT.md.

Dataset & data card

Trained on AfriVoices-KE (African Next Voices — Pilot Data Collection in Kenya), released by the KenCorpus Consortium (Maseno University, USIU-Africa, Kabarak University, DeKUT, LDRI) and Digital Umuganda, funded by the Gates Foundation. License: CC BY 4.0. The organisers' full data card is reproduced in DATA_CARD.md.

• Competition: AfriVoices East Africa: ASR Hackathon (Kaggle · host: Digital Umuganda)
• Test set: digitalumuganda/anv-test-data-nt · Training sources: Anv-ke (kik/luo/kln/mas/som), DigitalUmuganda/Afrivoice_Swahili, DigitalUmuganda/Afrivoice (Somali Mogadishu)

Composition (organisers' figures):

Language	ISO	Dialects	Scripted (h)	Unscripted (h)	Total (h)
Swahili	swa	Nairobi/Kisii/Wajir/Mombasa/Nakuru, Tanzania	0	2 979	2 979
Kikuyu	kik	Gĩ-Kabete, Ki-Mathira, Ki-Muranga, Ki-Ndia, Gĩ-Gichugu	183	571	754
Dholuo	luo	Nyandwat, Milambo	195	528	723
Somali	som	Maxatire, Mogadishu	118	884	1 002
Kalenjin	kln	Nandi, Kipsigis	122	399	521
Maasai	mas	Kimasaai, Kisamburu	51	454	505

Splits are speaker-disjoint (train 85 % / dev 5 % / dev_test 5 % / test 5 %). For LM shallow fusion, extra public text was used for Somali (CC-100) and Kalenjin (Zenodo).

Please cite the dataset (Wanzare et al., 2026 — accepted at LREC SIGUL 2026): AfriVoices-KE: A Multilingual Speech Dataset for Kenyan Languages, arXiv:2604.08448.

Limitations

• Kalenjin and Maasai remain the hardest languages (genuine low-resource acoustic difficulty).
• The KenLM decoder is per-language; language identity is assumed known at inference (or run LID first).
• Trained and evaluated on the AfriVoices-KE distribution (read + spontaneous speech).

License & citation

MIT. Backbone w2v-bert-2.0 is likewise MIT. Please cite the AfriVoices Hackathon and this repository.