sproocht/afrivoices-multi6-edge-asr

AfriVoices East Africa ASR (2nd-place solution) # AfriVoices East Africa ASR — 2nd Place Solution Single 580M **W2v-BERT 2.0 + CTC** model for six East African languages — Swahili, Kikuyu, Dholuo, Kalenjin, Maasai, Somali — decoded with per-language 5-gram KenLM shallow fusion, deployed to CPU/edge as **CTranslate2 int8**. **Final result: 0.34819 macro-WER (2nd of 26 teams), AfriVoices East Africa ASR Hackathon (Kaggle · Digital Umuganda & Maseno CAAI).** Submitted model: Checkpoint 11000 of a 1-epoch low-LR continued fine-tune of a previously averaged checkpoint, on an external-data-extended mix. | | | |---|---| | **Model** | leophill/w2v-bert-2.0-afrivoices-ea-6l-asr | | **Edge (CT2 int8 + ONNX)** | leophill/w2v-bert-2.0-afrivoices-ea-6l-asr-int8-edge | | **Technical report** | `TECHNICAL_REPORT.md` | | **Hardware validation** | `HARDWARE_VALIDATION_REPORT.md` | | **Data card** | `DATA_CARD.md` | | **Transparency note** | `TRANSPARENCY_NOTE.md` | ## Edge compliance | Requirement | Limit | Measured | Status | |---|---|---|---| | Parameters | 30 s) degradation measurement, windowing + stitch │ ├── eval_slices.py sliced WER/CER by type/dialect/domain, incl. synthetic long-form │ ├── longform_decode.py canonical decode core: pyctcdecode+KenLM, window, overlap-aware stitch │ ├── qc_wer_analysis.py error-bucket profiling (merge/split/edit/OOV/deletion/insertion) │ └── submissions frozen per-language submission CSVs + the merged submission │ ├── ledgers machine-readable exclusion ledgers (one subtree per language w/ external data) │ ├── Dholuo per-stage drop logs + _reconciliation.txt (raw N → kept N accounts exactly) │ ├── Kalenjin " │ └── Kikuyu " │ ├── logs reproducibility artefacts: SLURM stdout/stderr for the shipped training runs │ ├── Launch_w2v2-Bert_CTC_Finetuning_Part_1_4801063.err │ ├── Launch_w2v2-Bert_CTC_Finetuning_P …