Output details

Integrating Articulatory Features Into HMM-Based Parametric Speech Synthesis

<20> Originality: Statistical methods now dominate speech synthesis, whereby machine-learning algorithms automatically tune millions of parameters. These provide flexibility and performance, but are difficult to control explicitly and parsimoniously. This paper highlights this and is first to offer a general solution to controlling complex synthesis systems with parsimonious "meta" parameters.

Significance: IEEE Signal Processing Society's Young Author Best Paper award (2010).

Rigour: Both objective (RMS error) and subjective (human perceptual tests) evaluations conducted, with statistical significance testing. Listening tests, under controlled lab conditions, involved 3 forced-choice tests (40 listeners each) and a type-in word identification test (24 listeners).