SPEECH RECOGNITION FOR ANALYSIS OF POLICE RADIO COMMUNICATION.

Area of Science:

• Speech Recognition
• Computational Linguistics
• Public Safety Communications

Background:

• Broadcast police communications (BPC) via two-way radio are vital for police coordination and emergency response.
• These audio BPC are largely untranscribed, limiting their analytical potential due to challenging naturalistic audio properties.
• Automatic transcription of BPC is difficult, hindering the extraction of valuable information.

Purpose of the Study:

• To evaluate the feasibility of automatic speech recognition (ASR) for broadcast police communications (BPC).
• To assess the performance of various ASR models, including off-the-shelf, fine-tuned, and customized end-to-end models.
• To establish a benchmark for ASR performance in the police radio domain and identify areas for future research.

Main Methods:

• Collected a corpus of approximately 62,000 manually transcribed police radio transmissions (around 46 hours of audio).
• Evaluated the performance of large, off-the-shelf ASR models.
• Trained and evaluated ASR models fine-tuned specifically on the BPC corpus and developed customized end-to-end models.

Main Results:

• Both human and machine transcription of BPC proved challenging.
• Large, off-the-shelf ASR models exhibited poor performance on BPC data.
• ASR models fine-tuned on BPC data achieved performance comparable to human transcription levels.

Conclusions:

• Fine-tuning ASR models on domain-specific data significantly improves transcription accuracy for broadcast police communications.
• Further research is needed to address challenges like short utterances and potential miscommunication in police radio.
• The released corpus and annotation pipeline will facilitate future research in police communication analysis.