Re: [espeak-ng:master] new issue: Add a custom voice #github
Reece H. Dunn
On Wed, Oct 11, 2017 at 12:55 pm, Josh Kennedy wrote:
Neural network based speech synthesis is trained around a given language, so it is harder to get it to properly articulate sounds from another language. ESpeak supports around 100 languages, some of which have sounds that are hard to get the neural network synthesizers to pronounce.
That said, some of my longer-term goals are to support different voices at the phoneme data level, to provide support for better quality voice data (including klatt-based voices), and to provide better tools to develop and experiment with espeak-based voices. This includes supporting the mbrola voices at the phoneme data level, making it easier to use those voices in a different language (e.g. the German voices speaking English, or the Spanish voice speaking Italian). I also want to use this to provide higher-quality voices.
Right now the espeak code is making it complex to do that, as it is making several assumptions about how the voices are structured in order to try to be as compact as possible. As a result, big changes like this will take time.