Panako : a scalable audio search system

Panako solves the problem of finding short audio fragments in large digital audio archives. The content based audio search algorithm implemented in Panako is able to identify a short audio query in large database of thousands of hours of audio using an acoustic fingerprinting technique. The dataflow in Panako closely resembles the flow depicted in fig 1. Audio is transformed to features which are grouped in recognizable fingerprints. The fingerprints are compared with a database of reference fingerprints. If a match is found, it is reported or, in case of a true negative, the system reports that the audio is not present in the database. The properties of the acoustic fingerprinting system mainly depend on the selection of features, the information captured by the fingerprints and the performance of the matching step. There are three algorithms implemented in Panako. The first is based on pitch class histograms. The other two are based on peaks in a spectral representation: a baseline algorithm and the main Panako algorithm. The main algorithm in Panako finds matches between a short query fragment and the audio in the database even if the query is time-stretched, pitch-shifted, slowed down or sped up while maintaining other desirable features such as scalability, robustness and reliability. This is important since changes in replay speed do occur commonly, they are either introduced by accident during an analog to digital conversion or are introduced deliberately. Accidental audio speed changes are often the result of varying or uncalibrated recording or playback speeds of analogue physical media such as wax cylinders, wire recordings, magnetic tapes or gramophone records. To identify duplicates in a digitized archive, a music search algorithm should be robust against replay speed. Deliberate speed manipulations are sometimes introduced during radio broadcasts: occasionally songs are played a bit faster to fit into a time-slot. During a DJ-set speed changes are almost always present. To correctly identify audio in these cases as well, a music search algorithm must be robust against pitch shifting, time stretching and speed changes