Collaborative Audio Enhancement: Crowdsourced Audio Recording

This paper presents a collaborative audio enhancement system that aims to re-cover a high-quality recording from multiple low-quality recordings provided by the crowd attending the same event. We see this procedure as a crowdsourcing ex-ample, because neither an automated system nor a set of crowdsourced recordings cannot easily replace a professionally processed manual audio recording, which is expensive and not always available. We do the job in the context where each recording is uniquely corrupted by different frequency responses of microphones, audio coding algorithms, interferences, noise, etc. To this end, we adopt a method of simultaneous probabilistic topic modeling on synchronized inputs, called Prob-abilistic Latent Component Sharing (PLCS). In PLCS, some of the parameters are fixed to be same during and after the learning process to capture the common audio content while the rest model unwanted recording-specific interferences and artifacts. The main contribution of the paper is that we speed up the EM-based PLCS parameter estimation process by incorporating Winner-Takes-All (WTA) hashing so that the update can be performed with fewer selective observations rather than the whole. Therefore, we can deal with a bigger set of recordings more efficiently. Experiments on music signals with various artifacts show that the pro-posed method provides with sensible speed-up with no degrade of audio quality compared with the comprehensive PLCS model.