Using Exact Locality Sensitive Mapping to Group and Detect Audio-Based Cover Songs

Local summarization and multi-level LSH for retrieving multi-variant audio tracks

“…Regarding the first issue, the aim is to improve the accuracy of multi-variant audio track detection and the different proposals rely on pitch [9,10], Mel-Frequency Cepstral Coefficients (MFCC) [11,12,13] or Chroma [2,14]. With regard to the latter research issue, the goal is to accelerate the retrieval by similarity and the existing proposals include tree structures [7,15,16], other hierarchical structures [17], LSH [4,11,18], Exact Euclidean LSH (E 2 LSH) [3,4] and other variants of LSH [6,11]. It is however clear that the two research issues are not independent, since more accurate detection requires more elaborate representations of audio content, with a negative impact on scalability.…”

“…A summary is generated from a feature sequence, by using multivariable regression and Principal Component Analysis (PCA). In [6], weights are assigned to frequently employed features like MFCC, Chroma, Mel-magnitudes, based on a principle of spectral similarity invariance. A long audio feature sequence is summarized as a compact single Feature Union (FU).…”

“…One way to reduce the computation cost is to summarize the audio signals. In [6] audio signals are represented by statistics of a set of frequently used features. Although summarization effectively reduces the volume of data, a large part of the temporal information is lost and the accuracy decreases when the summaries are used for retrieval.…”

“…The effective and efficient detection of multi-variant audio tracks is thus a motivating topic in Music Information Retrieval (MIR), related to musical audio signal processing [1,2], audio content indexing [3,4], sequence matching [1,4,5], music retrieval evaluation [6], etc. It has already attracted significant research interest, see e.g.…”

“…[1,2,3,4]. The previous proposals rely on comparing either sequences of audio frame descriptions (feature sequences) [1,2] or on extracting very compact descriptions of entire audio sequences [6,7]. The use of feature sequences (denoted by {ri,j}) leads to more accurate retrieval but does not scale well to large audio databases because matching long sequences is expensive.…”