VOCAL SEPARATION USING SINGER-VOWEL PRIORS OBTAINED FROM POLYPHONIC AUDIO

Introduction


This page presents some results and media related to the article "VOCAL SEPARATION USING SINGER-VOWEL PRIORS OBTAINED FROM POLYPHONIC AUDIO", Shrikant Venkataramani, Nagesh Nayak, Preeti Rao, Rajbabu Velmurugan, to be published in the proc. of the conference, Int. Soc. for Music Information Retrieval(ISMIR) 2014, to be held in Taipei, Taiwan Oct 2014.

Summary of the work

Soft masks derived from a dictionary of singer-vowel spectra are used to improve upon the vocal-instrumental music separation achieved by harmonic sinusoidal modeling for polyphonic music of the particular singer. The main contribution of this work is an NMF based framework that exploits the amply available original polyphonic audios of the singer as training data for learning the dictionary of singer spectral envelopes. Appropriate constraints are introduced in the NMF optimization for training and test contexts. The availability of lyric-aligned audio (and therefore phone labels) helps to improve the homogeneity of the training data and have a better model with fewer basis vectors. Significant improvements in reconstructed signal quality are obtained over binary masking. Further it is demonstrated that a vowel-dependent soft mask obtained from clean data of an available singer is not as good as the singer-vowel dependent soft mask even if the latter is extracted from polyphonic audio.

Separation Results

In this section, we present some separation results. These example phoneme mixtures were obtained from different songs sung at average pitches of 200Hz and 300Hz respectively. The singer-specific dictionaries were obtained using a modified NMPCF algorithm, for all phonemes, at both the pitch ranges.
Mixture Ideal sources Separation using Binary Mask Separation using Soft Mask
Chookar_i_p200_01
Vocals
Accompaniments
Chookar_i_p200_02
Vocals
Accompaniments
Chookar_i_p300_01
Vocals
Accompaniments
Maanaa_i_p300_01
Vocals
Accompaniments
Maanaa_o_p300_01
Vocals
Accompaniments
OMere_o_p300_01
Vocals
Accompaniments

Paper Links

  • ISMIR pdf + tex
  • ISMIR pdf


    • Shrikant Venkataramani
    Last modified: Tue July 8 16:40:28 IST 2014