Pricile21/afrivoices-edge-asr

Unified edge-deployable ASR for 6 East African languages (AfriVoices EAC Hackathon) - w2v-bert-2.0 + char CTC + per-language KenLM. Training, decoding, submission & edge-validation scripts. # AfriVoices Edge ASR — Unified Speech-to-Text for Six East African Languages Training, decoding, submission and edge-validation code for the **AfriVoices East Africa ASR Hackathon** (Digital Umuganda / KenCorpus). A **single, edge-deployable** ASR model covering **Swahili (`swa`), Kikuyu (`kik`), Dholuo (`luo`), Kalenjin (`kln`), Maasai (`mas`) and Somali (`som`)**. - 🤗 **Model weights + KenLM + full cards:** huggingface.co - 📄 **Technical report:** `TECHNICAL_REPORT.md` · `TECHNICAL_REPORT.pdf` - 🔧 **Hardware validation:** `HARDWARE_VALIDATION_REPORT.md` - 🗂️ **Data card:** `DATA_CARD.md` · **Model card:** `MODEL_CARD.md` ## Model | | | |---|---| | **Backbone** | `facebook/w2v-bert-2.0` (Conformer, 580 M, MIT) | | **Head** | character-level CTC | | **Decoder** | per-language 4-gram KenLM (shallow fusion, `pyctcdecode`) + per-language α/β | | **Checkpoint** | `wbertcur-step-16500` | | **License** | MIT | ## Edge compliance (validated — see hardware report) | Requirement | Limit | Measured | |---|---|---| | 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 | **0.156** (4 cores, Raspberry-Pi-class) | ## 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). ## Approach (summary) - **Length curriculum** — max audio 15 s → 30 s → 45 s → 60 s (training on long clips from scratch collapses to blank). - **Language re-weighting** — `kln`×4, `mas`×4, `som`×2 (the macro metric is dominated by the low-resource languages). - **Canonical `normalize()`** applied identically at train / LM-build / inference, incl. a mojibake repair for the Dholuo `ng'` apostrophe. - **Shallow-fusion KenLM** per language (4-gram, web-augmented for `som`/`kln`), beam 100, …