| The restoration of musical recordings has been greatly improved by denoising techniques developed in the last few decades. However, these techniques rely on the important assumption that the noise level in each frequency band of the spectrum---also called noise floor---is known. Before noise reduction can be applied, the noise floor must be estimated from sections of the audio that contain exclusively noise (that is, no musical signal). This has been typically done by manually selecting a portion of noise-only audio, from which the noise floor is estimated in a so-called "noise learning" step. However, in many scenarios it is impossible, ineffective or undesirable to do this. Our dissertation research has focused on algorithms that eliminate the need for the manual noise-learning procedure. We present here algorithms that automatically extract the noise floor from an audio signal, even when there is music mixed with the noise, and no noise-only sections are available. This makes it possible to denoise an audio signal "on the fly", without the need to perform any initial steps. For example, it allows for a very large number of noisy sound files to be cleaned up without any user intervention (batch processing). Denoising a signal corrupted by noise that changes over time also becomes possible. We propose a new probabilistic model that simultaneously provides an optimal solution for two problems: noise floor extraction and frequency-dependent signal detection. Overcoming typical limitations in current technology, the proposed new algorithms allow for automatic noise reduction, eliminating the need for manual noise-learning steps before denoising can be applied. |
