Not at all useless. Simply decode all possible sequences and rank them, ranking the most self-consistent interpretation highest. You may also have other sources of data to help correlate the interpretation (there was an article earlier this year about measuring sound using the video footage of a mylar potato chip bag's vibrations.) Even if the room is crowded, it might be possible to identify a few isolated words from the audio recording of the conversation.
The next thing you do is throw away those conversations that you're not interested in. Regardless of whether the conversation resulted in "You punched a fish" or "You munched a dish", neither is going to have value when you're searching for criminal activity. But if your streams could be "I bought the ammo so we can rob the bank" or "I mopped the jam up sorry can you mop the tank?" one of those could be valuable.
99.999% of conversations are inane drivel. If this technology is applied, the number of false positives is going to rapidly overwhelm a system. More discrimination and correlation is going to be needed to actually produce intelligence from this data. But never think that data is worthless or unusable.