FAST-CAP Framework Reduces Speech-to-Speech Translation Latency by 26%
September 28, 2026
The FAST-CAP framework improves simultaneous speech-to-speech translation through a factorized architecture and a causality-aware adaptive policy. Experiments on Spanish, German, and French datasets show a 26% relative latency reduction and a 1.2 BLEU increase.
HOW THIS AFFECTS YOU
●
builderYou can implement this architecture to achieve better quality-latency trade-offs in real-time translation products.
●
researcherYou can use the causality-aware latency metric to evaluate simultaneous translation models more effectively.