This applies to applications that involve the creation of text or the formulation of very long queries. That is, in cases where the chances of an extended input being recognized perfectly are only moderate.
Users may not be certain of what they said originally and may be led astray by the erroneous decoding.