For me personally I wouldn't be interested in using anything besides the most accurate prediction. However, I would probably make the alternative hypotheses available to choose in case the best prediction is incorrect. In that case I can only assume that the hypotheses are listed in decreasing order of confidences.
If you have a relevant text corpus (e.g. previous transcripts of the same person), you could use some Markov-type modeling/analysis to verify the transcription or find the most suitable alternative in the list.
For me personally I wouldn't be interested in using anything besides the most accurate prediction. However, I would probably make the alternative hypotheses available to choose in case the best prediction is incorrect. In that case I can only assume that the hypotheses are listed in decreasing order of confidences.