YECS: A120-Hour Yoruba–English Code-Switching Corpus and a Diagnostic Framework for Low-Resource Speech Recognition

Code-switching (CS) speech resources remain heavily skewed toward high-resource language pairs, leaving tonal, morphosyntactically complex Sub-Saharan African languages almost entirely unrepresented. We introduce YECS, a 120-hour, 140-speaker Yoruba-English CS corpus, the largest naturalistic CS speech dataset for any Sub-Saharan African language pair, built through a community-centered "data-farming" pipeline with word-level language ID and emotion annotations (mean CMI = 34.22). Using YECS, we show that state-of-the-art multilingual speech models (Whisper, MMS, Omnilingual ASR) fail to generalize to natural code-switching in zero-shot settings (WER up to 320%). We introduce a diagnostic evaluation framework, Switch-Point WER, Point of Interest Error Rate (PIER), and switch-direction analysis, that localizes these failures to asymmetric errors on English-to-Yoruba transitions rather than uniform degradation. Fine-tuning on YECS is highly effective (a 244M-parameter Whisper model reaches 19.53% WER, outperforming zero-shot models 5x its size) but induces catastrophic forgetting whose severity depends on pretraining paradigm rather than fine-tuning data alone. The YECS corpus and our fine-tuned models are already publicly released online, offering a ready-to-use corpus and benchmark suite, along with a replicable, ethical data-collection template for under-resourced language pairs beyond Yoruba-English.