Fusion of live audio recordings for blind noise reduction

The prevalence of digital cameras and video-capable mobile phones enables the common practice of audiences capturing recordings of live music performances. It is now increasingly common to find some of these personal recordings online, uploaded to popular video hosting websites. Recognizing the desire of music fans to obtain a recording of good audio quality, we offer a time-domain fusion technique for combining these samples to achieve higher signal-to-noise ratio (SNR) than the single best original sample. When no composite output can improve SNR, the best original sample recording is returned. The scenario is modeled as a single blind source with multiple diverse receivers. As every live performance is unique, we assume no prior knowledge of a reference signal, and no knowledge of the original recordings' SNRs. Using statistical characteristics among the samples, we infer relative SNR, rank samples by quality, and determine whether a composite delivers improvement. The technique can be applied in a variety of contexts where multiple receivers have opportunity to capture audio, speech, or other signals.