G.S.MOZE COLLEGE OF ENGINNERING BALEWADI,PUNE -45.

1
G.S.MOZE COLLEGE OF ENGINNERINGBALEWADI,PUNE
-45.

•  
• A PRESENTATION ON
• Voice Morphing
• PROJECT GUIDE
• By
• Anil Mahadik
  Prof.
  Sonali Ghote

2
Content

• Title
• Introduction
• History
• Need of Vocal track area function
• Vocal track area function
• AR-HMM Analysis
• AR-HMM Diagram
• Re-synthesis of Converted voice

3

• Training Phase
• Conversion and morphing phase
• Application
• Conclusion
• References

4
Title

• The Project title is Voice Morphing.
• Give the information about Flexible Voice
  Morphing based on linear combination of
  multispeakers vocal tract area function.
• Voice morphing or voice conversion usually means
  transformation from a source speakers speech to
  a target speakers.

5
Introduction

• The main goal of the developed audio morphing
  methods is the smooth transformation from one
  sound to another.
• These techniques are considered to be a kind of
  point-to-point mapping in a feature space.
• There are many applications which may benefit
  from this sort of technology.
• Research on voice morphing aims to extend this
  restriction to area-to-area mapping by
  introducing multi-speakers .

6
History

• Voice morphing is a technology developed at the
  Los Alamos National Laboratory in New Mexico, USA
  by George Papcun and publicly demonstrated in
  1999.
• Voice morphing enables speech patterns to be
  cloned and an accurate copy of a person's voice
  be made which can then say anything the operator
  wishes it to say.

7
Need of Vocal track area function

• Since the 1990s, many techniques for voice
  conver-sion have been proposed 1-7.
• One successful technique is to use a statistical
  method for mapping a source speakers voice to a
  target speakers but a weakness of these methods
  is the discontinuity of formants.
• The proposed method employs an estimated vocal
  tract area function to avoid such weakness.

8
Vocal Tract area function(A)

• Interpolation in the vocal tract area domain is
  considered to provide reasonably continuous
  transition of formants.
• Estimation of the vocal tract area function
  implies simultaneous estimation of the voice
  source characteristics.

9
AR-HMM analysis

• For this purpose of Estimation of the vocal
  tract area function introduce Auto-Regressive
  Hidden Markov Model (AR-HMM) analysis of speech.
• The AR-HMM model represents the vocal tract
  characteristics by an AR model and the glottal
  source wave by an HMM.
• The AR-HMM analysis estimates the vocal tract
  resonance characteristics and vocal source waves
  in the sense of maximum likelihood estimation.

10
Diagram of AR-HMM
11
(No Transcript)
12
Re-synthesis of the converted voice

• There are two phases Training phase and
  Conversion Morphing phase.
• The procedure of each phase is as follow
• in Diagram.

13
Training phase

• AR-HMM analysis Speech samples with the same
  phonetic content from both source and target
  speaker are analyzed .
• Feature alignment The feature vectors obtained
  above are time-aligned using dynamic time warping
  (DTW) in order to compensate for any differences
  in duration between source and target utterances.
  
• Estimation of the conversion function The
  aligned vectors are used to train a joint GMM
  whose parameters are then used to construct a
  stochastic conversion function.

14
Training phase
15
Conversion and morphing phase

• AR-HMM analysis In this case only the source
  speakers utterances are used.
• Features Transformation The GMM-based
  transfor-mation function constructed during
  training is now used for converting every source
  log vocal tract area function and vocal cord
  cepstrum into its most likely target equivalent.
• Linear Interpolation ,Synthesis of the source
  wave and LPC synthesis.

16
Conversion and morphing phase
17
Application

• Applications as the creation of peculiar voices
  in animation films.
• Voice morphing has tremendous possibilities in
  military psychological warfare and subversion.
• Voice morphing is a powerful battlefield weapon
  which can be used to provide fake orders to the
  enemy's troops, appearing to come from their own
  commanders.

18
Conclusion

• This paper has presented a voice morphing method
  based on mappings in the vocal tract area space
  and glottal source wave spectrum that can each be
  independently mod-ified.
• These features have been realized using AR-HMM
  analysis of speech.
• In future, we will investigate how to improve the
  quality of voice conversion with interpolation
  techniques.

19
References

• 1 L.M. Arslan, D.Talkin, Voice conversion by
  codebook map-ping of line spectral frequencies
  and excitation spectrum, Proc. Eurospeech,
  pp.1347-1350, 1997.
• 2 Y.Stylianou, O.Cappe, A system voice
  conversion based on probabilistic classification
  and a harmonic plus noise mod-el, Proc.ICASSP,
  pp.281-284, 1998 .
• 3 A.Kain, Spectral voice conversion for
  text-to-speech syn-thesis, Proc.ICASSP
  pp.285-288, 1998.
• 4 H. Ye, S. Young, High Quality Voice
  Morphing, in Proc.IEEEICASSP, pp.9-12, 2004.

20
(No Transcript)