Implementation of Shazam Entertainment's algorithm for audio matching.
Requires the rust compiler
git clone https://github.com/owenbrooks/atlas.git
cd atlas
Add a directory (tracks/ in this case) of .wav files to the database (must be 16 bit samples, mono or stereo):
cargo run --release -- -i tracks/ add
Run matching:
cargo run --release -- -i sample.wav match
- Create a spectrogram image of the audio file using a Fast Fourier Transform
- Perform a maximum filtering operation on the image
- Compare the filtered with the original image to find locations where the intensity doesn't change - these must be local peaks in the signal
- Compute hashes on nearby pairs of peaks
- These are either stored in the database, or matched against existing hashes in the database
- If many of the hashes in the sample match to a particular point in time of a particular song in the database, it is a succesful match
| Original Spectrogram | Filtered Spectrogram |
|---|---|
 |
 |
| Peak Locations |
|---|
 |
The track info and hashes are stored in a sqlite database on disk.
See https://www.ee.columbia.edu/~dpwe/papers/Wang03-shazam.pdf and https://www.cameronmacleod.com/blog/how-does-shazam-work for further explanations of the algorithm.