Samcil/ghana-asr-lab

Transfer-learning ASR experiment lab for Akan/Twi: rethink Khaya/DONDO with WaxalNLP, Common Voice, and shared WER evals # ghana-asr-lab Rethinking the Khaya/DONDO architecture for Ghanaian-language speech recognition — by measuring, not guessing. Transfer learning on transformer speech encoders, evaluated on one fixed, general-domain Akan test set. ## The rethink, in one page Khaya AI's DONDO models (Jul 2026) = Meta's **w2v-BERT 2.0** encoder + a CTC character head, fine-tuned on **religious read speech** with stepped learning-rate annealing. That recipe is radically capital-efficient, but it has three known weaknesses: 1. **Domain narrowness** — trained and scored on Bible-style speech; real usage is worse. 2. **CTC-only decoding** — no language-model decoder, so homophones/punctuation suffer. 3. **Manual language selection** — no LID; and it rides entirely on Meta's encoder. This lab tests the alternatives head-to-head on general-domain Ghanaian speech: | Axis | Question | Experiments | |---|---|---| | Architecture | Does an encoder-decoder (Whisper) or adapter model (MMS) beat encoder+CTC at equal data? | E1 vs E3 vs E4 vs E5 | | Initialization | Do Khaya's open Akan-cluster weights transfer better than Meta's raw encoder? | E1 vs E2 | | Data domain | How much does Bible-only training cost on real speech? (DONDO's drawback, quantified) | E1 vs E6 | | Efficiency | WER per parameter and real-time factor — what can serve **offline on cheap hardware**? | all rows | The eval harness comes first: existing Hub models (MMS zero-shot, asr-africa's XLS-R Akan series, two community Whisper fine-tunes) get scored on our test set **before we train anything**, so every training run must beat a real number, not a vibe. ## Data | Source | What it is | License | Role | |---|---|---|---| | google/WaxalNLP `aka_asr` | ~10K labeled + 175K unlabeled Akan utterances, built on Univ. of Ghana's UGSpeechData | CC-BY-SA-4.0 | anchor train/val/test | | Common Voice Twi | crowdsourced read speech (Mozilla Data Collective download — no longer on HF) | CC0 | extra train + out-of-domain eval | | ghana …