Multilingual ASR

Multilingual ASR studies how to turn speech into text reliably across languages, dialects, accents, and mixed-language conversations, including languages with little training data. Systems range from CTC and transducer models to encoder-decoder recognizers such as Whisper and speech-language models connected through lightweight projectors; researchers adapt them with fine-tuning, language-specific adapters, distillation, and synthetic speech. Work increasingly tests real-world conditions such as spontaneous speech, medical terminology, speaker overlap, and code-switching, while benchmarks examine low-resource languages and regional variation. A central challenge is balancing shared cross-language knowledge with language-specific needs, while improving evaluation beyond WER to account for scripts, spelling variation, meaning, and demographic disparities.