Breaking the bounds:introducing informed spectral analysis

Informed Sound Analysis is a novel approach consisting in adding just missing data needed to otain a fixed level of quality for estimations.

Applications: Informed Sound Analysis combinated with Watermarking technique will make possible to create augmented content allowing high-quality sound application like effect application, music transcription or sound source separation.

In fallowing experiment, we consider a synthetic sound sampled at 44100Hz composed by 5 sinusoids with a time and spectral envelope of piano. Sound samples are frame-by-frame analyzed using reassignment method with Hann (size window N=511). Sinusoid frequencies are estimated first, then resynthetized with additional information. SNR per partial is computed before white noise addition in order to estimate required information per partial for each time. A blind reassignment estimation is available for listenning to make auditory comparison.

Original sound Sound + Noise
Sound blind estimated Sound Informed estimated
Global SNRSNR per partial
Global bitratebitrate for partial 1
bitrate for partial 2bitrate for partial 3
bitrate for partial 4bitrate for partial 5
Used global information:10625 bits
Global maximal information (16bits) :14960 (16 x 5 x 187 frames) bits
Signal size (2.16sec, 44.1Khz, 16bits):1524096 bits (2.16 x 16 x 44100)
Global information per partialOverall information utilized
Partial 1 frequency trajectoryPartial 2 frequency trajectory
Partial 3 frequency trajectoryPartial 4 frequency trajectory
Partial 5 frequency trajectory