The 'FreddieMeter' is an app that was launched in late 2019 by the band Queen to let fans 'measure' how closely their voice matched that of Freddie Mercury's. We set out to build our own version of it as a hack for the hackday organised by MusicTechCommunity India in December 2019.
The idea was to encode the voices of different popular artists in Indian film music and build a reduced dimensional space, in which the encodings ("embeddings") formed well-separatred clusters corresponding to each artist. Then, given a test voice sample, we would similarly encode it and find the closest artist cluster.
With the limited time we had, we worked with the audios of 4 singers - Arijit Singh, Atif Aslam, Lata Mangeshkar, Chinmayi.Â
Using Spleeter we separated out their vocals from tracks containing all kinds of other instruments, used a deep-learning based voice analyzer model called Resemblyzer to find the embeddings, and then used UMAP to project them down to a 2-D space.
An interactive python-notebook demo of the final prototype: https://github.com/rohitma38/sing-alike-interactive