Speech Translation

Speech translation maps spoken input to translated text or speech, supporting applications from multilingual communication to live interpretation. Systems either cascade speech recognition with machine translation and speech synthesis, or translate directly with end-to-end models and speech-capable language models. Current methods use multilingual training, acoustic-text alignment, and chunking or adaptive read-write policies to improve low-resource coverage and stream translations with less delay. A central challenge is balancing latency and semantic accuracy with speech-specific qualities such as speaker identity, prosody, emotion, and disfluencies, while developing evaluations that measure real-world usability as well as text quality.