efficient-sequence-transduction-by-jointly-predicting-tokens-and-durations