Short-term sound stream characterization for reliable, real-time occurrence monitoring of given sound-prints

Abstract

The identification in real-time streaming data using plain template-matching algorithms is difficult due to the undefined frame position in the on-line data compared to the frame-based features' position of the templates. We have investigated the use of the short-time Fourier-spectrum, the short-time Walsh-rpectrum and the short-time signal energy and the ratio of tbe succeeding features from the frame-shift point of view. The last two features aim to improve calculation s p e d in larger set of records. For further simplifying opcrrtions in the comparison stage, a quantization step was applied to the spectrum values which resulted in temary-bgic time-frequency maps. This is also useful for eliminating the effects of non-extraordinary spectral-shape dltortions by utilizing the prominent parts of the spectrum. An rlgorithm was developed for selecting the most suitabk segment combination of the sound records to be monitored where tbe differences between a11 segment pairs are the l a m The method was applied to identify advertisements am t&e RealAudio broadcast of the Hungarian Radio.Znder terms-d t i m e detection, feature selection, sound stream m o~i t o r i 4

A Review of Audio Fingerprinting

“…Some other transforms have been proposed: the Discrete Cosine Transform (DCT), the Haar Transform or the Walsh-Hadamard Transform [2]. Richly et al did a comparison of the DFT and the Walsh-Hadamard Transform that revealed that the DFT is generally less sensitive to shifting [20]. The Modulated Complex Transform (MCLT) used by Mihçak et al [5] and also by Burges et al [21] exhibits approximate shift invariance properties [5].…”

“…It is quite common to apply a very low resolution quantization to the features: ternary [20] or binary [4,19]. The purpose of quantization is to gain robustness against distortions [4,19], normalize [20], ease hardware implementations, reduce the memory requirements and for convenience in subsequent parts of the system.…”

“…The purpose of quantization is to gain robustness against distortions [4,19], normalize [20], ease hardware implementations, reduce the memory requirements and for convenience in subsequent parts of the system. Binary sequences are required to extract error correcting words utilized in [5,19].…”

“…In [8], the classification is Nearest Neighbor using a cross entropy estimation. In the systems where the vector feature sequences are quantized, a Manhattan distance (or Hamming when the quantization is binary) is common [4,20]. Mihçak et al [5] suggest that another error metric, which they call "Exponential Pseudo Norm" (EPN), could be more appropriate to better distinguish between close and distant values with an emphasis stronger than linear.…”