Speech Interaction

Currently the speech recognition abilities of computers are limited in practice, due to

• inter-individual phonetic differences
• intra-individual phonetic differences (emotions, colds, time of day)
• ambient noise
• continuous speech
• interjections (ahem, hm), breaks, repetitions

Partial remedies

• train the system on speech data of individual user
  (speaker-dependent recognition, e.g. in dictation software)
• leave the system in control of the interaction (by, e.g., asking questions, telling the user what to say)
• limit what the user can say at each step (e.g., to "yes" and "no")

Some design recommendations for speech dialogues

• verify the correct understanding of every user input
• abbreviate instructions after user heard them repeatedly ("tapering")
• allow users to request a human agent (by saying "agent")
• redirect to human in case of repeated recognition failure