Speaker Clustering for Multilingual Synthesis
Alan W. Black, Tanja Schultz · 2006
Today, speech synthesizers in new languages are typically built by collecting several hours of well recorded speech in the target language. The time and effort involved in collection and correction can be prohibitive when lack of resources is common in addressing under-represented languages. An alternative method is to use acoustic data from an existing synthesizer in a different language and train adaptation models from a small corpus (20-50 sentences) in the target language. Following the work of GlobalPhone [Schultz, 2001], which uses multi-lingual databases and adapts acoustic models to target languages for speech recognition. This paper presents a method to build multi-lingual synthesizers using combined data from multiple languages, instead of building a large number of different synthesis voices. Using speaker clustering techniques we find suitable speakers for combined synthesizers.