Audio Identification System

Coursework 2 : Music Informatics (2021) Queen Mary University of London

An implementation and evaluation of an audio fingerprinting and identification system using librosa and 2D-array peak picking algorithm.

This system by default uses the GTZAN datasets for recorded audio (for fingerprinting) 1 and query audio (for identification) 2.

Requirements

• python3
• librosa
• numpy
• skimage
• scipy
• hashlib
• matplotlib

Runing

To run both the fingerprinter and the identification against the two default datasets:

$ make print-and-id

The program is also callable from a python environment as follows:

> fingerprintBuilder(/path/to/database/,/path/to/fingerprints/)
> audioIdentification(/path/to/queryset/,/path/to/fingerprints/,/path/to/output.txt)

To only invoke the fingerprint builder:

$ make fingerprinter

Or

> fingerprintBuilder(/path/to/database/,/path/to/fingerprints/)

To only invoke the audio identification program:

$ make shazam

Or

> audioIdentification(/path/to/database/,/path/to/fingerprints/)

NOTE: The default location for the database recordings is './database_recordings' and the default location for query recordings is'./query_recordings' niether of which is included in this repo.

Evaluation

To evaluate a single run you must have your estimates and your ground truth sequences in local files and the evaluator can be evoked by running:

$ make eval-one BEAT_ESTIMATE=<path-to-estimate-sequence> BEAT_GROUND_TRUTH=<path-to-ground-truth>   

To evaluate the outputs

$ make evaluate

Or

> evaluate.py

References

[1] [2] [3]