Andrei-Florian/salma2026-african-whisper

# Evaluating the Effect of Linguistic Relatedness on Cross-Lingual Transfer in Large Multilingual Automatic Speech Recognition Code accompanying the SALMA 2026 submission. Released under GPL-3.0. --- ## Overview This repository contains the full experimental pipeline for a study evaluating whether linguistic relatedness between languages reliably predicts cross-lingual transfer gains in large multilingual ASR. We adopt the two-stage sequential fine-tuning methodology of Nowakowski et al. (2023) and Khare et al. (2021), pre-adapting a base model on an auxiliary language before fine-tuning on a low-resource target, and extend it to large multilingual ASR through a systematic controlled experimental design spanning six factors, two Africa-centric corpora (AfriVoices KE and Google WaxalNLP), and four models (Whisper Small, Whisper Large v3, XLS-R 300M, OmniASR CTC 7B). Across all conditions, pre-adaptation on linguistically related auxiliary languages yields no practically meaningful improvement in target-language performance from as little as one hour of target-language data onward. --- ## Repository Structure | File | Description | |------|-------------| | `finetune_whisper_fft.py` | Full-model fine-tuning for Whisper Small and Whisper Large v3. Covers Factors 1, 3, 4, 5, and the Whisper Large condition of Factor 6. Two-stage via `PHASE` switch in USER CONFIG. | | `finetune_whisper_lora.py` | LoRA fine-tuning for Whisper Small. Covers Factor 2. Same two-stage `PHASE` switch. | | `finetune_wav2vec_fft.py` | Full-model fine-tuning for XLS-R 300M. Covers the XLS-R condition of Factor 6. Builds a CTC vocabulary from training transcripts; supports two-stage fine-tuning via `--model_dir`. | | `finetune_omni_fft.py` | Full-model fine-tuning for OmniASR CTC 7B. Covers the OmniASR condition of Factor 6. Loads the raw `.pt` checkpoint via a custom CTC wrapper compatible with the HuggingFace Trainer. | | `results_analysis.ipynb` | Produces Figures 3–5 and the results tabl …