VOICE CONVERSION METHODS FOR VOCAL TRACT AND PITCH CONTOUR MODIFICATION

1
EUROSPEECH 2003 (Interspeech 2003), 8th European
Conference on Speech Communication and
Technology, September 1-4, 2003, Geneva,
Switzerland
VOICE CONVERSION METHODS FOR VOCAL TRACT AND
PITCH CONTOUR MODIFICATION
Oytun Türk Levent M. Arslan RD Dept.,
SESTEK Inc., and EE Eng. Dept., Bogaziçi
University, Istanbul, Turkey oytun_at_sestek.com.tr
arslanle_at_boun.edu.tr
Bogaziçi University
SESTEK Inc.

• SELECTIVE PRE-EMPHASIS ANALYSIS
• Bandpass filtering frame-by-frame processing
• LP analysis on each subband component
• Fullband spectrum estimated as
• k1 lower cut-off of (i1)th filter
• k2 higher cut-off of ith filter

4

• SELECTIVE PRE-EMPHASIS TRAINING
• Same utterances from source target speaker
• Automatic alignment by Sentence-HMMs and/or
  manual alignment
• Selective pre-emphasis
• Generate LSF codebooks for each sub-band
  component

• SELECTIVE PRE-EMPHASIS TRANSFORMATION
• Selective pre-emphasis applied on source
  utterance
• Full-band excitation spectrum modified separately
• Weighted average of corresponding target sub-band
  codebook entries
• Full-band vocal tract spectrum estimated using
  selective pre-emphasis based synthesis

7
8
EXAMPLE LP vs. Selective pre-emphasis based
spectral estimation

• SELECTIVE PRE-EMPHASIS SYNTHESIS
• Use synthesis vocal tract and excitation spectra
• Inverse Fourier Transform
• Perform overlap-add synthesis

6
5