Informed Spectral Analysis for Isolated Audio Source Parameters Estimation


Informed Spectral Analysis is a two step approach for spectral analysis that consists in combining a classic blind estimator with the extra-information necessary to reach a target quality.

Motivation:
  • music content creator have access to separated tracks before the mixing process
  • Blind analysis method have theoretical limitation for the best reachable accuracy
    		•

    Proposal of solution:
    a complete method in order to embeds inaudibly into the mixture the minimal necessary extra-information required by the chosen estimator to obtain the desired quality.

    Coder:

    Decoder:

    Sounds results 1: For this experiment we consider a single-channel 44.1kHz-sampled music signal with 2 sources. During the preliminary analysis step, sinusoidal parameters of the target source signal are first blind estimated with the reassignment method using a Hann analysis window of length N=1023 with 50% overlap.
    Original Mixture
    (voice + guitar)    		 Watermarked mixture Voice semi-Blind estimation Guitar semi-Blind estimation
    wav wav wav (SNR=-2.37dB) wav (SNR=-9.05dB)

    VoiceGuitar
    SNR extra-info bit rate Sound file SNR extra-info bit rate Sound file
    61.8dB 60.93kb/s (instead of 125.67kb/s without estimator) wav 70dB 40.16kbits/s (instead of 68.21kb/s without estimator) wav


    Sounds results 2: For this experiment we consider a single-channel 44.1kHz-sampled music signal with 5 sources : voice, guitar, bass, drum and ambiance. The voice and guitar track are respectively estimated using Informed Spectral Analysis.
    Original Mixture
    (5 sources)    		 Watermarked mixture Voice semi-Blind estimation Guitar semi-Blind estimation
    wav wav wav (SNR=0.54dB) wav (SNR=0.46dB)

    VoiceGuitar
    SNR extra-info bit rate Sound file SNR extra-info bit rate Sound file
    58dB 72.55kb/s (instead of 131kb/s without estimator) wav 52dB 59.52kb/s (instead of 98.75kb/s without estimator) wav