MarkNwilliam/uganda-dysarthric-asr

First ever Luganda and Ugandan English dysarthric ASR models # Uganda Dysarthric Speech Recognition First-ever ASR models for Ugandan speakers with dysarthria — supporting both **Luganda** and **Ugandan English**. Built for people with motor speech disorders (dysarthria) caused by neurological conditions like cerebral palsy, stroke, and Parkinson's disease. ## Results | Model | Language | Architecture | WER | |-------|----------|-------------|-----| | whisper-ugandan-english-dysarthric-v8 | Ugandan English | Whisper-Small + LoRA | 32.13% | | wav2vec2-luganda-dysarthric-v7 | Luganda | Wav2Vec2 + KenLM | **24.03%** | Both models run **completely offline** on CPU or GPU. ## Repository Structure ``` ├── english/ │ ├── train_v8_final.py # English Whisper LoRA training (32.13% WER) │ └── augment_english.py # English dysarthric augmentation ├── luganda/ │ ├── train_luganda_wav2vec2_v6.py # Luganda v6 baseline (63.11% WER) │ ├── train_luganda_wav2vec2_v7.py # Luganda v7 final (27.87% WER) │ ├── augment_luganda_v2.py # 12-type aggressive augmentation │ └── build_kenlm.sh # KenLM language model build script ├── inference/ │ ├── transcribe_luganda.py # Offline Luganda transcription │ └── transcribe_english.py # Offline English transcription └── README.md ``` ## Quick Start ### Luganda (24.03% WER) ```python from transformers import Wav2Vec2Processor, Wav2Vec2ForCTC import torch, librosa processor = Wav2Vec2Processor.from_pretrained("nugwa-mark/wav2vec2-luganda-dysarthric-v7") model = Wav2Vec2ForCTC.from_pretrained("nugwa-mark/wav2vec2-luganda-dysarthric-v7") audio, _ = librosa.load("speech.wav", sr=16000) inputs = processor(audio, sampling_rate=16000, return_tensors="pt") with torch.no_grad(): logits = model(**inputs).logits print(processor.decode(torch.argmax(logits, dim=-1)[0])) ``` ### Ugandan English (32.13% WER) ```python from transformers import WhisperProcessor, WhisperForConditionalGeneration from peft import PeftModel import torch, librosa base …