New framework turns noisy body-cam audio into editable transcripts

A new arXiv paper (2607.24958) from Vishnoi, Khurana, Timmapur, Rai, and Mohanty introduces a speech intelligence framework purpose-built for the extreme acoustic conditions of body-worn cameras. These recordings suffer from high ambient noise, variable microphone placement, and overlapping speakers—challenges that break conventional ASR and diarization. The proposed architecture uses a dual-path pipeline: a diarization branch with DeepFilterNet for denoising, VAD, and NVIDIA's Multi-Scale Speaker Diarization Decoder (MSDD) with TitaNet embeddings, and a transcription branch using loudness normalization and WhisperX Large-v3 with forced alignment. Outputs are fused at the word level by assigning each recognized word to the speaker segment with the greatest temporal overlap.

The framework was tested on a curated dataset of U.S. and U.K. police body-worn camera recordings. Results show that task-specific acoustic conditioning and probability-guided speech segmentation significantly improve both speaker diarization and word-level attribution accuracy. The authors highlight the modular design as an extensible foundation for future operational conversational intelligence systems, such as real-time transcription for law enforcement, emergency services, or field journalism. This work bridges a gap between controlled lab speech systems and the chaotic realities of field audio.