One of the first hurdles in most deep learning applications is obtaining high quality labelled data, which is usually time consuming and expensive. There are strong models in NLP (BERT), computer vision (DINOv3) and speech recognition (wav2vec), that circumvent this problem using self-supervised learning. Our goal will be to apply/adapt these methods this to anomalous sound detection (ASD) with the purpose of automatic machine failure detection.

The data consists of the normal/anomalous operating sounds of several types of real/toy machines. There are two concrete datasets which can be found at:
-https://zenodo.org/records/3384388
-https://zenodo.org/records/6355122
These include some baseline models which we can use to benchmark our proposals against. We will only use the normal operating sounds to train our model as that is what would be available to us in a real-life scenario. Performance will be measured using a subset of normal together with all anomalous operating sounds.

Music recommendation algorithms by companies such as Spotify optimize for continuous engagement, usually recommending albums and songs that conform closely to what the listener is already familiar with. This limits possible vectors of music exploration. Using datasets of full-text music reviews, this project uses natural language processing to create a music recommendation engine that uses *semantic similarity* between reviews of different albums as a source of musical discovery. Rather than suggesting music of the same genre(s) a person has been listening to, the model will propose albums that reviewers have used similar semantic approaches to describe. For the curious music lover interested in deep listening and willing to venture a bit further astray than their typical taste profile, this model offers critic-informed suggestions that tap into cultural and aesthetic connections that behavioral data and audio features alone can’t capture.