Edinburgh Research Explorer Fusing ASR Outputs in Joint Training for Speech Emotion Recognition