“Speech-to-speech models, which bypass an intermediate text layer, are primarily used to achieve lower latency but suffer from lower reliability compared to cascaded models.”