Around ten years ago, music streaming services were competing heavily for the best music recommendation system. Clearly, a flawless recommendation system would provide the user with the exact piece of music that optimally satisfies their needs, every time. However, some people view recommendation systems as transitional technology. Ultimately, no matter the size of your music catalog, there can’t be a perfect fit available for each possible user request.
Modern generative AI systems could potentially solve this problem by generating music that is (robot) hand-tailored to each request. On the other hand, these generative systems are still not producing high-quality music, have tremendous computational costs, and are subject to complex ethical and legal concerns.
Therefore, this article aims to compare the benefits & limitations of search-based music retrieval and music generation to find out whether we should expect generative systems to fully replace, augment, or not even affect the current solutions. Before we start, let’s define what we mean by a “search algorithm” and a “generative model”.
A search algorithm is a solution to a search problem. A search problem exists when a user wants to retrieve a piece of information or an object like a video or a song from a database. Let’s call the user’s request the query and the result of the search the response. The goal of a search algorithm is to find that piece of information that optimally satisfies the user’s needs, i.e. provides an optimal response for the given query.
However, there is also a time constraint on the search problem. Most of the time, we would prefer to receive the second-best response after 10 seconds over the absolute best response after 10 hours. Therefore, a search algorithm should find a response that is qualitatively satisfactory, within a reasonable time.
A generative model is a solution to a prediction problem. Based on a set of input parameters (query), the model generates a prediction for what the optimal…