Audio Deconvolution Tool
Reverb removal is a challenging problem within audio DSP where one attempts to extract a “dry” original signal from a reverberant audio recording (often for the purposes of audio restoration or improving speech intelligibility). In order to investigate this problem within the scope of a single semester, we focused on the more constrained problem of audio deconvolution: an approach which assumes that the user has an impulse response of the room they are attempting to eliminate the effects of. Just as a dry audio signal can be convolved with an impulse response of a room to simulate the reverberant properties of that room (convolution reverb), a reverberant signal can be “deconvolved” by an impulse response to approximate the unknown dry signal.
Deconvolution algorithm results for a Cello sample
Recording test audio and impulse responses
In order to test this method, we first collected our own test set of impulse responses and reverberant recordings in four rooms with contrasting acoustic properties. However, we soon encountered a challenge with this method: due to deconvolution being equivalent to division in the frequency domain, subtle background noise in the recordings is often amplified into loud unpleasant ringing artifacts. Our final software program resolves this issue by combining audio deconvolution with a Wiener filter (a method which excels at removing consistent noise, i.e. the ringing artifacts), resulting in outputs that were significantly clearer than either method could produce on its own. For a more detailed description of the process, check out the project website, available here.