This demo is a prototype of a system that tells you which popular artist you sound most like, based on a recorded sample of your singing voice. The implementation currently has been built with 4 Indian popular music artists - Arijit Singh, Atif Aslam, Lata Mangeshkar, Chinmayi.
We obtained some recordings of each artist, separated the vocals using Spleeter and obtained embeddings on 10-second chunks using Resemblyzer. These embeddings are 256-length vectors, that are supposed to "encode" the essential characteristics of the voice. We then used the supervised UMAP projection to learn a reduced 2-dimensional space for these embeddings.
Given a test sample, we first encode it, use the same projection to reduce its dimensionality and classify it using a simple k-nearest neighbors method.
The demo is provided in the form of a python notebook with instructions in place, wherever needed.
- To run it on Google Colab, click on the 'Open In Colab' button at the top.
- To run it locally, you can use the 'demo_local.ipnb', but be sure to install the dependencies listed in the requirements file (do also check out the dependencies that resemblyzer will install and note that pyaudio will require portaudio to be pre-installed)
This tool was built as a hack during Music Hack Day India in December 2019, the first hackathon organised by MusicTechCommunity India.
- [Athithyan Paramasivan] ()
- Rohit M A
- Add scripts used for pre-processing and generating 2d embedding clusters