Code-Switching Speech Recognition

Code-switching speech recognition studies how automatic speech recognizers handle utterances that mix languages, dialects, and specialized vocabulary, where a system may omit or normalize words from one language. Common approaches adapt multilingual pretrained models with parameter-efficient fine-tuning, synthetic mixed-language speech, pseudo-labeling, preference or reinforcement learning, and language-aware decoding or hotword biasing. Evaluation uses bilingual datasets and measures such as word or mixed error rate, alongside switch-localized errors, language-specific accuracy, and script fidelity. Systems are improving on supported language pairs, but transfer to unseen pairs remains limited, and a central challenge is improving switch recognition without sacrificing monolingual accuracy or relying on aggregate metrics that hide language-specific failures.