Cross-correlation of beat-synchronous representations for music similarity

Abstract

Systems to predict human judgments of music similarity directly from the audio have generally been based on the global statistics of spectral feature vectors i.e. collapsing any large-scale temporal structure in the data. Based on our work in identifying alternative ("cover") versions of pieces, we investigate using direct correlation of beat-synchronous representations of music audio to find segments that are similar not only in feature statistics, but in the relative positioning of those features in tempo-normalized time. Given a large enough search database, good matches by this metric should have very high perceived similarity to query items. We evaluate our system through a listening test in which subjects rated system-generated matches as similar or not similar, and compared results to a more conventional timbral and rhythmic similarity baseline, and to random selections.

Audio Cover Song Identification and Similarity: Background, Approaches, Evaluation, and Beyond

“…These notes can be unique for each time slot (a melody) or can be played jointly with others (chord or harmonic progressions). From a MIR point of view, clear evidence about the importance of tonal sequences for music similarity and retrieval exists [9,22,34]. In fact, almost all cover song identification algorithms exploit tonal sequence representations extracted from the raw audio signals: they either estimate the main melody, the chord sequence, or the harmonic progression.…”

Error Tolerant Multimedia Stream Processing: There's Plenty of Room at the Top (of the System Stack)

“…Within the MSP frameworks of a music matching system [54] and an MPEG-7 metadata calculation from audio streams, it was demonstrated that this leads to virtually no effect on the applications' precision [6]. Similarly, other experiments [7] show that the peak performance of the generic matrix multiplication (GEMM) routine of BLAS achieved on a multicore processor can be increased by up to 92% in comparison to the state-of-the-art double-precision GEMM routine of the Goto library [68].…”

“…Application examples utilizing such kernels are: document clustering [53], multimedia retrieval engines (such as images/video/music/forensic-indices/metadata-based retrieval [54]), webpage ranking systems [13], etc. Given the prevalence of such systems, significant emphasis has been placed on their efficient parallelization of their computationally-intensive elements (kernels) in computer clusters or GPUs [55].…”

Precision-energy-throughput scaling of generic matrix multiplication and discrete convolution kernels via linear projections

“…and aims for minimum mean-squared error (MSE) or maximum learning, recognition or matching rate against ground-truth or training data, rather than performance bounds for individual inputs. Examples of such error-tolerant (ET) systems include: lossy image/video/audio compression [2], [3], computer graphics [4], [5], webpage indexing and retrieval [6], object and face recognition in video [7], [8], image/video/music matching [9]- [12], etc. For instance, all face recognition and webpage ranking algorithms optimize for the expected recall percentage against ground-truth results and not for the worst-case.…”