AI Learns to Decode Garbled Police Radio Without Human Help

Police radio is a mess of static, sirens, overlapping voices and shorthand codes. General-purpose AI transcribers like Whisper, which handle a clean podcast fine, stumble badly on this kind of audio. That's a problem for anyone trying to study how police actually make decisions, because you need a written record before you can analyze anything — and paying humans to type up thousands of hours of garbled audio is slow and expensive.

So the researchers tried a shortcut called "pseudo-labeling" — essentially letting the AI teach itself. The AI listens to raw audio and writes its own transcript, then trains on that transcript as if it were correct. The obvious risk: its mistakes get baked in and repeated. The team found that AI's built-in confidence scores are surprisingly useless at spotting bad transcripts. Their fix was clever — a second AI reads each transcript and asks, "does this actually make sense?" Anything that sounds contextually impossible gets thrown out before it poisons the training data.

They tested this on police radio archives from Baltimore and Chicago. The "does this make sense?" filter cut the error rate significantly compared with trusting the AI's own confidence scores. They also tried a cross-model trick — letting one AI learn from the other's transcripts — which they flag as the most promising direction for future work. Still, a gap remains compared with a perfect, human-checked filter.

Why does this matter beyond academia? Police radio and body camera audio sit at the center of debates about accountability, misconduct and public trust. Cheaper, more accurate transcription means watchdogs, journalists and investigators can review far more footage for far less money. The same trick could eventually help with 911 calls, customer service recordings or messy meeting notes. The catch: errors in a legal or disciplinary context aren't harmless typos, and this system still makes them.