The effect of normalization -- a case study in speech synthesis

Add to your list(s) Download to your calendar using vCal

• Matt Shannon (University of Cambridge)
• Thursday 03 March 2011, 14:00-15:30
• Engineering Department, CBL Room 438.

If you have a question about this talk, please contact Shakir Mohamed.

Undirected graphical models are ubiquitous in application domains of machine learning. However the normalization constants in these models are often difficult to compute, and as a result are frequently dropped altogether. In this talk we’ll look at the qualitative effect of this lack of normalization in the domain of statistical speech synthesis.

Specifically we’ll compare the predictive distributions of the standard unnormalized speech synthesis model, its globally-normalized undirected counterpart, and a more tractable directed graphical model. Along the way, we’ll highlight some of the general issues surrounding the choice between undirected and directed graphical models for sequence data.

The introduction to speech synthesis I will give will be aimed entirely at machine learners with no background in modelling speech, and will I hope be realistic (close to state-of-the-art), self-contained and framed in terms of probabilistic modelling.

This talk is part of the Machine Learning Reading Group @ CUED series.

This talk is included in these lists:

• All Talks (aka the CURE list)
• Cambridge University Engineering Department Talks
• Centre for Smart Infrastructure & Construction
• Engineering Department, CBL Room 438
• Featured lists
• Inference Group Journal Clubs
• Inference Group Summary
• Machine Learning Reading Group @ CUED
• Machine Learning Summary
• Quantum Matter Journal Club
• Required lists for MLG
• School of Technology
• TQS Journal Clubs
• engineering

Note that ex-directory lists are not shown.