Simultaneous Speech Translation

Simultaneous speech translation studies how to translate spoken language before the speaker has finished, so listeners can follow in near real time. Systems use cascaded speech recognition and text translation or end-to-end speech and language models, with timing controlled by fixed chunking, adaptive read/write policies, attention-based alignment, or training on partial-speech prefixes. Evaluation compares translation quality with measures of delay on benchmarks such as MuST-C and CVSS, while long-form and speech-to-speech settings also expose accumulating delays and awkward pauses. The central challenge is to reduce latency without sacrificing accuracy, stable output, or natural speech flow; methods that work well on short segments may not transfer to continuous conversation.