Department of Engineering

Matt Shannon

Research

Acoustic modelling for statistical speech synthesis. For the first part of my PhD I was part of the EMIME (Effective Multilingual Interaction in Mobile Environments) project for personalized speech synthesis. I am currently part of the EPSRC Natural Speech Technology project.

Software

• armspeech
  Flexible python framework for probabilistic modelling of speech with a focus on autoregressive models. Allows the use of more complicated autoregressive-style output distributions than the HTS implementation below. Includes example experiments.
• Autoregressive HMM for HTS
  An implementation of the autoregressive HMM built on top of HTS (HMM-based Speech Synthesis System). It provides the ability to do embedded re-estimation, decision tree clustering and synthesis using the autoregressive HMM. An example of how to use this implementation is available (autoregressive HMM version of the HTS demo).

Journal papers

• M. Shannon, H. Zen and W. Byrne (2013)
  Autoregressive models for statistical parametric speech synthesis
  IEEE Transactions on Audio, Speech, and Language Processing (vol. 21, no. 3, pp. 587–597)
  [ postprint | publisher ]
  

Conference papers

• M. Shannon and W. Byrne (2013)
  Fast, low-artifact speech synthesis considering global variance
  Proc. ICASSP 2013
  [ postprint ]
  
• M. Shannon, H. Zen and W. Byrne (2011)
  The effect of using normalized models in statistical speech synthesis
  Proc. Interspeech 2011 (pp. 121–124)
  [ postprint | publisher | slides ]
  
• M. Shannon and W. Byrne (2010)
  Autoregressive clustering for HMM speech synthesis
  Proc. Interspeech 2010 (pp. 829–832)
  [ postprint | publisher | poster ]
  
• M. Shannon and W. Byrne (2009)
  Autoregressive HMMs for speech synthesis
  Proc. Interspeech 2009 (pp. 400–403)
  [ postprint | publisher | slides ]
  

Technical reports

• M. Shannon and W. Byrne (2013)
  Partially analytic speech parameter generation considering global variance
  Department of Engineering, University of Cambridge, UK
  Technical Report CUED/F-INFENG/TR.682
  (to appear; a preliminary version can be obtained by emailing me)
  
• M. Shannon and W. Byrne (2012)
  Viewing the trajectory HMM as a generalized autoregressive HMM
  Department of Engineering, University of Cambridge, UK
  Technical Report CUED/F-INFENG/TR.677
  (to appear; a preliminary version can be obtained by emailing me)
  
• M. Shannon and W. Byrne (2009)
  A formulation of the autoregressive HMM for speech synthesis
  Department of Engineering, University of Cambridge, UK
  Technical Report CUED/F-INFENG/TR.629
  [ pdf ]
  

Theses

• M. Shannon and M.J.F. Gales (2008)
  Sampling methods for instantaneous speaker adaptation
  MPhil thesis, University of Cambridge, UK
  [ final with corrections | final | slides ]
  

Other talks and posters

• M. Shannon (April 2013)
  Towards better probabilistic models of speech: why sampled trajectories sound bad and how to fix them – a discussion
  Speech group seminar, University of Cambridge, UK
  (led an informal discussion)
  
• M. Shannon (work with W. Byrne) (Jan 2013)
  Fast, low-artifact speech synthesis considering GV
  NST project meeting, University of Cambridge, UK
  [ slides ]
  
• M. Shannon (work with W. Byrne) (Dec 2012)
  An analysis of parameter generation considering global variance
  UKSpeech two-day conference, University of Birmingham, UK
  [ poster ]
  
• M. Shannon (Mar 2011)
  The effect of normalization – a case study in speech synthesis
  Machine Learning RCC, University of Cambridge, UK
  [ slides | talks.cam ]
  
• M. Shannon and H. Zen (Jan 2011)
  Modelling trajectories in statistical speech synthesis
  Cambridge statistical speech synthesis (SSS) seminar series
  University of Cambridge, UK
  [ my slides | Heiga's slides | talks.cam ]
  
• M. Shannon and S. Bratières (May 2010)
  Topics in statistical machine translation
  Machine Learning RCC, University of Cambridge, UK
  [ talks.cam ]
  
• M. Shannon (work with W. Byrne) (May 2010)
  Autoregressive HMMs for speech synthesis
  EMIME workshop, Cambridge, UK
  [ slides ]
  
• M. Shannon (work by Yee Whye Teh and others) (Jan 2009)
  A hierarchical Bayesian language model based on Pitman-Yor processes
  Machine Learning RCC, University of Cambridge, UK
  [ talks.cam ]
  

Teaching

• demonstrator for the CSTIT MPhil course 2008-9 and 2009-10 and the Advanced Computer Science MPhil course 2012-13

Contact information

Baker Building, BE5-02
Engineering Department
Trumpington Street, Cambridge
CB2 1PZ, United Kingdom