| --- |
| base_model: badrex/w2v-bert-2.0-kinyarwanda-asr |
| library_name: peft |
| license: apache-2.0 |
| language: |
| - rw |
| tags: |
| - automatic-speech-recognition |
| - kinyarwanda |
| - lora |
| - health |
| pipeline_tag: automatic-speech-recognition |
| --- |
| |
| # UBUZIMA AI — Kinyarwanda ASR LoRA adapter |
|
|
| A LoRA adapter over [badrex/w2v-bert-2.0-kinyarwanda-asr](https://huggingface.co/badrex/w2v-bert-2.0-kinyarwanda-asr), |
| fine-tuned for a Kinyarwanda voice health assistant. Part of the UBUZIMA AI capstone |
| project at the African Leadership University, Kigali. |
|
|
| ## Model details |
|
|
| - **Developed by:** Ganza Didier (African Leadership University) |
| - **Model type:** LoRA adapter for Wav2Vec2-BERT 2.0 CTC speech recognition |
| - **Language:** Kinyarwanda (rw) |
| - **License:** Apache 2.0 |
| - **Fine-tuned from:** badrex/w2v-bert-2.0-kinyarwanda-asr |
|
|
| ### Adapter configuration |
|
|
| r = 32, alpha = 64, dropout 0.05, target modules `linear_q` and `linear_v`. |
| 3,145,728 trainable parameters (0.54% of the 583.6M base). Feature extractor and |
| feature projection frozen. |
|
|
| ## Training |
|
|
| Fine-tuned in two sequential stages on pooled Afrivoice (health-domain) and Common |
| Voice (general-domain) Kinyarwanda speech: |
|
|
| - **Stage 1:** 100 hours at a 1:1 Afrivoice:Common Voice ratio. |
| - **Stage 2:** warm-started from Stage 1, 500 hours at a 1:4 ratio. This adapter is |
| the Stage 2 result. |
|
|
| Training regime: fp16 mixed precision on an NVIDIA T4 (Google Colab). |
|
|
| ## Evaluation |
|
|
| | Test set | n clips | WER | 95% CI | CER | |
| |---|---|---|---|---| |
| | Afrivoice (in-domain, health) | 1,560 | 6.76% | [6.44, 7.11] | 1.39% | |
| | Common Voice (cross-domain, general) | 16,205 | 31.29% | [30.86, 31.72] | 8.99% | |
|
|
| WER = word error rate, CER = character error rate. 95% confidence intervals from |
| 500-resample bootstrapping. |
|
|
| **Contamination note.** The base badrex checkpoint was pretrained on Kinyarwanda |
| audio that included portions of the Digital Umuganda health corpus from which the |
| Afrivoice subset is drawn. The in-domain figure therefore reflects both this |
| adaptation and residual pretraining familiarity; the Common Voice figure is the |
| cleaner cross-domain measure of what the adapter added. |
|
|
| ## How to use |
|
|
| ```python |
| from transformers import Wav2Vec2BertForCTC, AutoProcessor |
| from peft import PeftModel |
| |
| base = "badrex/w2v-bert-2.0-kinyarwanda-asr" |
| model = Wav2Vec2BertForCTC.from_pretrained(base) |
| model = PeftModel.from_pretrained(model, "dganza/Ubuzima_AI") |
| model = model.merge_and_unload() # fold LoRA in for faster inference |
| processor = AutoProcessor.from_pretrained("dganza/Ubuzima_AI") |
| ``` |
|
|
| ## Intended use and limitations |
|
|
| Research prototype for Kinyarwanda health-*information* delivery. **Not a medical |
| device; does not diagnose.** Dialect coverage is limited to the standard Kinyarwanda |
| represented in the training corpora. Cross-domain performance (Common Voice, 31% WER) |
| is substantially weaker than in-domain (Afrivoice, 7% WER), so out-of-domain use |
| should expect that gap. |
|
|
| ## Citation |
|
|
| Ganza, D. (2026). *UBUZIMA AI: A Kinyarwanda voice health assistant* [Capstone |
| project]. African Leadership University. Supervisor: Emmanuel Adjei. |
|
|
| ### Framework versions |
|
|
| - PEFT 0.12.0 |
| - Transformers 4.44.2 |
