A Speech Coding Method using an Anthropomorphic and Acoustic Approach

1
A Speech Coding Method using an Anthropomorphic
and Acoustic Approach
Nguyen Viet Son Eric Castelli Speech Processing
Group
Oriental-COCOSDA 2007, Hanoi
2
Introduction

• Most automatic speech synthesis systems are
  based on signal processing techniques such as
  unit concatenation
• The quality is rather good.
• The greatest disadvantage difficult to change
  the voice.
• Another approach modeling the mechanical and
  aero-dynamical behavior of the vocal cords and
  modeling the wave propagation of the signal waves
• Very high quality and possible to change the
  voice but difficult to control it
• In order to reduce this difficulty a new
  approach consisting in coupling the analog model
  to the DRM (Distinctive Region Model) model
• For telecommunications and data transmission, our
  anthropomorphic approach can be an interesting
  way to reduce transmission rates while keeping a
  good quality for speech.

2
3
Content

• I. Anthropomorphic synthesis.
• II. Validation.
• III. Use for coding.
• IV. Conclusion.

3
4
Acoustic model

• The vocal cords
• Assume to be symmetric.
• Vibrate and act as an oscillating valve.
• The air is interrupted into a series of pulses as
  serves as the excitation source for the vocal
  tract.
• The vocal tract
• Dividing into elementary tubes.
• The relationship of the flow volume at each
  junction is following

Schematic diagram of the vocal cord and vocal
tract system. (J.L. Flanagan and al. 1975)
4
5
Distinctive region model (DRM)

• The DRM is structured in regions
• by the zero-crossings of the sensitivity function
  computed on the uniform closed-open tube.
• The sensitivity function specifies how each
  formant responds to a local perturbation of the
  area along a tube.
• In speech synthesis, the vocal tract is
  considered as a closed-open tube, or a
  closed-closed tube
• Initial area function 4 cm2, and can vary from
  0.5 to 16 cm2.
• The total length 18 cm

• Deforming the shape can change (increase/decrease)
  efficiently (small deformation lead to large
  formant variations) F1, F2, F3 or both of F1
  and/or F2 and/or F3, etc.
• At each step, the sensitivity function is
  recomputed and a new deformation is performed
  base on it.

5
6
Distinctive region model (DRM)
Distinctive regions
R4
R3
R2
R1

• The DRM is
• a 2-region model when only F1 is controlled,
• a 4-region model when (F1,F2) are controlled,
• a 8-region model when (F1, F2, F3) are
  controlled.
• The limits of those 8 regions correspond to the
  zero-crossing of the sensitivity function.

6
7
Distinctive region model (DRM)

• The places of articulation are fixed
• R8 corresponding to the lips.
• R7 to the teeth.
• R3, R4, R5, R6 to the tongue.
• R1 to the larynx.
• For vowel production, the DRM uses (F1, F2) with
• The front constriction (R5, R6) and back
  constriction (R3, R4) for the closed-open tube.
• The central constriction (R4, R5) for
  closed-closed tube.
• For consonant production, the DRM uses (F1, F2,
  F3) with
• R8 labial
• R6 coronal.
• R5 velar.

7
8
DRM model commands the anthropomorphic synthesizer
Vocal tract
Lungs
Cord vocal

• Our new approach consists in replacing the
  articulatory model by the DRM
• (R1-R8) are use the determination of the vocal
  tract shape.
• (Q-Ag0) to control the fundamental frequency
  (F0).
• Ps for the energy.

8
9
Validation (1)

• Produce vowels /a/, /i/, /u/ and their
  transitions
• Correct in the cases of the transition /a-u/ and
  /a-i/.
• Not really the same in the transition /i-u/ but
  their directions are more or less close.
• Produce some simple sentences with some
  consonants in Vietnamese language.
• Capable to synthesize the vowels and some
  consonants

9
10
Validation (2)
11
Use for coding

• In order to control the speech system, we need to
  use 12 parameters, but from the point of view of
  coding (more efficient in the control of the
  synthesizer/coder)
• Consider R1-R2 change very small ? keep their
  geometry constant.
• Only use 10 parameters Ps, (Q, Ag0), (R3-R8) and
  velum position parameter
• Vocalic and consonant gestures are modeled as
  linear or logarithmic geometry changes
• For coding application
• Not need to refresh command parameters at every
  sample.
• Using simple interpolations to determine
  intermediary values.
• A small transmission rate
• 10 parameters x 8bits x 40Hz 3.2Kbits

11
12
Conclusion

• A new approach to command an anthropomorphic
  speech synthesizer with several advantages
• Allow to better understand phenomena involved in
  the speech production.
• The coupling between acoustic model and the DRM
  is simple.
• Need only 10 - 12 parameters to control the whole
  system.
• Possible to change the quality of the voice and
  produce different types of voices.
• For coding, need only a 3.2kbits transmission
  rate to be controlled.
• Possible to realize an embedded version of the
  coder on DSP processor.

12
13
References

• C. C. Wai. 2003. Speech coding algorithms.
  Foundation and Evolution of Standardized Coders.
  Ed. Wiley Sons.
• E. Castelli. 1999. Modélisation anthropomorphique
  de la parole. Thèse HDR, INP Grenoble, France.
• J. L. Flanagan, K. Ishizaka, and K. L. Shipley.
  1975. Synthesis of speech from a dynamic model of
  the vocal cord and vocal tract. The Bell System
  Technical Journal, Vol.54, No.3, pages 485-506.
• K. Ishizaka, and J. L. Flanagan. 1972. Synthesis
  of voiced sounds from a two-mass model of the
  vocal cords. The Bell System Technical Journal
  Vol.51, No.6, pages 1233-1268.
• M. Mrayati, R. Carré and B. Guerin. 1988.
  Distinctive regions and modes A new theory of
  speech production. Speech Communication 7, pages
  257-286.
• R. Carré. 1996. Prediction of vowel systems using
  a deductive approach. Proceeding of Int. Cong.on
  Speech and Language Processing, Philadelphia,
  pages 434-437.
• R. Carré and Maria Mody. 1997. Prediction of
  vowel and consonant place of articulation.
  Proceeding of the Third Meeting of the ACL
  Special Interest Group in Computational
  Phonology, SIGPHON 97, Madrid.
• R. Carré. 2003. From an acoustic tube to speech
  production. Speech Communication 42, pages
  227-240.
• S. Maeda. 1982. A digital simulation method of
  the vocal tract system. Speech Communication 1,
  pages 199-299.

13
14

• Thanks for your attention !!!

14