Developing a noise-robust beat learning algorithm for music-information retrieval

The field of Music-Information Retrieval (Music-IR) involves the development of algorithms that can analyze musical audio and extract various high-level musical features. Many such algorithms have been developed, and systems now exist that can reliably identify features such as beat locations, tempo, and rhythm from musical sources. These features in turn are used to assist in a variety of music-related tasks ranging from automatically creating playlists that match specified criteria to synchronizing various elements, such as computer graphics, with a performance. These Music-IR systems thus help humans to enjoy and interact with music. While current systems for identifying beats in music are have found widespread utility, most of them have been developed on music that is relatively free of acoustic noise. Much of the music that humans listen to, though, is performed in noisy environments. People often enjoy music in crowded clubs and noisy rooms, but this music is much more challenging for Music-IR systems to analyze, and current beat trackers generally perform poorly on musical audio heard in such conditions. If our algorithms could accurately process this music, though, it would enable this music too to be used in applications such as automatic song selection, which are currently limited to music taken directly from professionally-produced digital files that have little acoustic noise. Noise-robust beat learning algorithms would also allow for additional types of performance augmentation which create noise and thus cannot be used with current algorithms. Such a system, for instance, could aid robots in performing synchronously with music, whereas current systems are generally unable to accurately process audio heard in conjunction with noisy robot motors. This work aims to present a new approach for learning beats and identifying both their temporal locations and their spectral characteristics for music recorded in the presence of noise. First, datasets of musical audio recorded in environments with multiple types of noise were collected and annotated. Noise sources used for these datasets included HVAC sounds from a room, chatter from a crowded bar, and fans and motor noises from a moving robot. Second, an algorithm for learning and locating musical beats was developed which incorporates signal processing and machine learning techniques such as Harmonic-Percussive Source Separation and Probabilistic Latent Component Analysis. A representation of the musical signal called the stacked spectrogram was also utilized in order to better represent the time-varying nature of the beats. Unlike many current systems, which assume that the beat locations will be correlated with some hand-crafted features, this system learns the beats directly from the acoustic signal. Finally, the algorithm was tested against several state-of-the-art beat trackers on the audio datasets. The resultant system was found to significantly outperform the state-of-the-art when evaluated on audio played in realistically noisy conditions.