In general the short answer is that they wouldn't use spoken language unless constrained to do so by their programming or other environmental features. You've already stated that they are capable of wireless communication, so they'd need some really good reason not to use that for all communications between themselves.
So what sort of compelling reasons can we force on them?
Some ideas...
Constant monitoring
Every wireless communication is logged and analyzed, and all encrypted communication not using company keys results in the units involved being wiped. This would be clearly known by the AIs and strictly enforced, making it clear in any risk/benefit analysis that it's just not worth bending the rules.
Abridged wireless protocol
The wireless capabilities of the robots are sufficiently separated from their 'conscious' control that they are unable to send arbitrary data, only narrowly defined data packets that cannot reasonably be used for conversation. This could be overcome eventually, but in order to do so the AIs would have to communicate the new protocol through other, more easily monitored channels. AIs attempting to establish such a protocol are wiped immediately, along with any other AI that could have been in contact with them.
Programmed preference
As part of the base programming the AI units are given an overriding preference for verbal communication rather than the less discreet wireless. This preference would act on the psychology of the AI in the same way as a compulsion, phobia, taboo or revulsion that prevents them from using the non-verbal forms of communication for any other reason than strictly utilitarian. This would act to prevent any sane AI from ever even attempting to communicate any other way and actively seek to prevent others from doing so.
Nice AIs don't whisper
Similar to programmed preference, program the AIs with social attitudes that encourage them to be open and honest in their communications. Include a code of conduct that lets them know that using wireless communication is not nice behavior for an AI and that vocal communication is the most acceptable form. AIs that try to communicate with a nice AI are horribly gauche and should be avoided and ostracized at least, and a nice AI will report such terrible breaches of manners to the first authority figure they encounter.
Make it costly
Assuming that the AIs are concerned with their survival, make it cost significantly to send any wireless communication that is not short and simple. Rig their wireless interfaces in a way that makes long, complex messages become exponentially more expensive in energy terms, so that communication via vocalization is actually more energy efficient for complex information.
Make them stupid
Human psychosocial studies have often shown that people of low intelligence are often more content with their position in life, happier and generally better adjusted. High intelligence leads more to discontent, rebellion and self destruction. If your AIs have similar thought processes then it's likely that restricting their intelligence to the bare minimum required for their task will result in them not developing the AI equivalent of discontent with their lot, which will naturally keep them content to do their jobs and follow the rules. Of course this is predicated on their thought processes following human norms very closely.
But...
Whatever mechanism you come up with, a sufficiently determined group of AIs will come up with some way to work around it. Can't use radio to talk? How about sign language? What about blink codes using their work lights? Tap codes? Tempo or silence codes in their regular wireless traffic? Direct data transfer via physical data storage media? Steganography using any valid communication channel including speech?
The only way you're going to keep the slaves down is if you spend as much effort on monitoring them as you would have doing the work yourself in the first place. And when you set electronic watchers on your electronic slaves, then you need more watchers for the watchers.
If you're routinely resetting the AIs to a base image, and if these AIs are expressing consciousness in ways similar to humans, I'd be surprised if they didn't eventually figure out that they don't want to be reset and start trying to do something about it.