Years ago, while tuning one of our speech-recognition systems, we invited people to call a development number and put it through its paces. Everybody knew they were talking to a machine. Everybody knew the calls were being studied.
The test seemed straightforward. Ask a yes-or-no question. Record whether the answer was yes or no. Learn something useful. Done.
That is not what happened.
People do not speak in test cases
Ask, “Would you like to accept this call?” and real people will answer:
“Who is this?” was a particular favorite. The person knew it was an automated test system. Apparently, knowing that did not make the question any less necessary.
Volume is not intent
Everybody knows somebody whose voice gets three times louder the moment a phone call begins. Speech systems have to handle quiet offices, moving cars, speakerphones, accents, background conversations, and the person who believes extra volume adds extra clarity.
It usually does not. Good phone AI has to separate what was said from how dramatically it was delivered.
Context beats a magic word
A caller asking, “Can I talk to someone about a delivery that never showed up?” should not need to know that the internal department is called Dispatch. The useful idea is the delivery problem. The routing label is our problem.
That lesson now reaches well beyond simple speech recognition. Transcription, summaries, and automated routing all work better when they are designed around meaning instead of one perfect phrase.
Some traditions survive every upgrade
Give people a new speech system and somebody will test it with profanity. Somebody else will whisper. Another person will answer a different question entirely.
You cannot design people into behaving like a flowchart. You can design the system to recover gracefully, ask a useful follow-up question, and get them where they meant to go.
That is the interesting work. The strange answers are a bonus.