ziheng-zhang-aus/warden

WARDEN is a research pipeline for processing Wardaman, an indigenous Australian language. It tackles two closely linked tasks: automatically transcribing spoken Wardaman audio into text, then translating that text into English. The pipeline has three stages. The first is ASR (automatic speech recognition) fine-tuning, which adapts Whisper, an existing speech recognition model, to handle Wardaman audio. Utility scripts download source recordings, convert them, and build training datasets split into train, validation, and test sets. Audio segments are capped at 30 seconds and included only when both transcription and translation annotations are available. The second stage is lexicon retrieval and injection. A cleaned lexicon file with more than 2,000 manually reviewed Wardaman-English word entries augments transcriptions before translation. The injection module finds matching words using exact and fuzzy character-error-rate comparison, then adds glossary definitions alongside each transcribed sentence, giving the translation model explicit vocabulary hints for uncommon words. The third stage is translation fine-tuning, which trains a Qwen language model using LoRA (a method for efficiently adapting large models with fewer resources) via a tool called LLaMAFactory. The model is trained on transcribed Wardaman text paired with English translations. Pre-trained checkpoints for both the speech recognition and translation models are published on Hugging Face, so researchers can use them without retraining from scratch. The pipeline is written in Python, with conda used to manage separate environments for the ASR and translation stages.