I was traveling a lot last week and ended up calling United frequently for flight information. They use a voice recognition system to provide that data. Horrible idea.
In a previous job, I worked on speech recognition systems and know that in general they work OK in quiet environments. But in noisy places, they have a really hard time because they can’t filter out background noise from the person speaking.
With the sound of jet engines, other passengers and frequent announcements, I would often get bad information or be asked to repeat myself. If I did get through a step, it would often ask for confirmation (“I think you said San Francisco, is that correct?”), giving the system another chance to get lost by interpreting background noise as a “no”.
In another exchange, for a flight that made multiple stops, the system prompted me for the city I wanted. Again, it gets lost in background noise. It would have been faster to play the information for both cities.
The designers of speech rec systems, including the United one, try to inject personality into the prompts, making them longer and again increasing the opportunity for misrecognition. They need to keep the prompts as short as possible and provide the information as fast as possible.