Coupling Speech Encoders with Downstream Text Models

Ciprian Chelba, Johan Schalkwyk

We present a modular approach to building cascade speech translation (AST) models that guarantees that the resulting model performs no worse than the 1-best cascade baseline while…