On audio recognition performance via robust hashing

Abstract

In recent years, automated audio recognition is arising as a relevant issue. One of the most successful approaches has proved to be the Philips method [1], by Haitsma et al.. Our work exploits this framework in order to allow for a better tuning of the algorithm. In particular, we show that more accurate hashes do not always lead to a better performance, which can then be attained with an inferior computational cost. A simple and efficient search method is also proposed.Index Terms-Audio fingerprint, content based audio identification, music information retrieval.

Reliable Automatic Recognition for Pitch-Shifted Audio

“…More than one threshold would likewise correspond to different degrees of reliability. Heuristic considerations [7] suggest a single T = 0.35, which we verified [8] can be often regarded as appropriate. A high T will more easily provide an answer, but it will also lead to an increased false positive rate, which is unacceptable in some scenarios.…”

“…Details and insights regarding our current implementation for FA can be found in [8] and will be summarized thanks to Fig. 2, where major blocks are depicted.…”

“…The choice of sub-bands is discussed in III-A, while we fixed b = 16 in light of [8]. A vector of b bits is then computed according to the rule:…”

“…Our previous investigation [8] showed that improvements could indeed be attained by means of a careful tuning of the parameters. Through this algorithm and a few enhancements we successfully addressed the problem of recognizing pitch-shifted audio, which is ordinarily a demanding task.…”

A Framework for Robust Audio Fingerprinting

“…A possible work-around, which scans anyway every database entry, is described in [30]. There, we compute the distance only on promising regions, whose locations are determined by the peaks of the cross-correlation between X and each possible R. The cross-correlation is performed along the time dimension and averaging along the frequency axis.…”