taosanna/Gawarbati-ASR

ASR for low-resource language Gawarbati (automatic orthographic and IPA transcription) # Gawarbati-ASR ASR for low-resource language Gawarbati (automatic orthographic and IPA transcription). Code accompanying the Master's thesis: "Machine Learning for Gawarbati: Scaling multilingual ASR and extending towards automatic linguistic annotation" ## Contents - Data preprocessing - ELAN extraction - Whisper fine-tuning (LoRA) - Orthographic ASR evaluation - STIPA evaluation - Evaluation and analysis ## Data Availability The Gawarbati recordings, ELAN files, generated manifests, model outputs, and model checkpoints are not included due to privacy and data-access restrictions. ## Script Order ### 1. Data extraction and inspection - `extract_gawarbati_segments.py`: extracts segment metadata and annotations from ELAN/WAV pairs. - `inspect_dataset_final.py`: computes dataset statistics and checks missing values. - `split_dataset_final.py`: creates speaker-disjoint train/dev/test splits. ### 2. Feature preparation - `precompute_features.py`: prepares Whisper features for orthographic ASR. - `precompute_features_ipa.py`: prepares Whisper features for STIPA. - `merge_datasets.py`: merge output of precompute_features.py. - `merge_datasets_ipa.py`: merge output of precompute_features_ipa.py. ### 3. Baselines - `zero_shot_whisper_gawarbati.py`: runs zero-shot Whisper ASR. - `ipa_stipa_zero_shot.py`: runs zero-shot STIPA evaluation. - `epitran_vs_gold_ipa_updated.py`: evaluates Epitran G2P against gold IPA. - `segment_examples_and_perspeaker_averages.py`: get auto-language detection stats. ### 4. Fine-tuning - `train_whisper_lora.py`: LoRA fine-tuning for orthographic output. - `train_whisper_lora_ipa.py`: LoRA fine-tuning for Whisper-Tiny IPA output. - `train_whisper_lora_ipa_large.py`: LoRA fine-tuning for Whisper-Large IPA output. - `train_whisper_array.sh`: SLURM array job for training. ### 5. Evaluation - `evaluate_whisper_tiny_finetune.py`: evaluates fine-tuned Whisper-Tiny ASR. - `evaluate_whisper_large_finetune.py`: evaluates fine-tuned Whisper-Large …