Develop a model that, given a library of songs, after one song from the library finishes playing, selects the next song in a way that ensures a “seamless” transition between the two. The beginning of the second song should closely match the ending of the first song, maintaining a smooth flow.
We intend to predict the next song base on the ending of the previous song using autoagressive model or generative model.
We will evaluate our AI using metrics that assess the similarity between the first song and the selected song in terms of tempo, mood, key, note progression, and other relevant features. Baselines will include random selection and an algorithm that selects songs based solely on whether they share the same key. The success of our model will be measured by calculating a success score based on the similarity of the feature vectors of two songs, and comparing this score to those generated by the baselines.
A toy example will involve a small set of songs with clearly defined features, allowing us to observe if the model generates smooth transitions between songs. For sanity checks, we will test edge cases where songs differ significantly and ensure the model does not link them together. Our ambitious goal is to create an API that enables users to input a library of songs and receive a recommended order that ensures smooth transitions between all songs in the library.
The earlist meeting date will be Jan 28th 2025.