Sarah Creer wrote:
Hello, I've been trying to find a way to evaluate the voices that I've been building objectively.
Hi Sarah, objective evaluation: the holy grail of TTS research :-)As far as building HTS voices goes (e.g. tuning the decision tree clustering), then the likelihood of the training data (or a dev set) is probably the only objective measure people use.
Distance to natural speech is very problematic, because there is more than one 'good answer' in speech synthesis, and these good versions of a sentence may not cluster together in acoustic space (think about different prosodic renderings of a sentence). However, many people use this measure.
The speech coding community use various objective measures, and we are hoping to try some of them on past Blizzard data, and also in Blizzard 2008 - this will give us an idea of whether they correlate in any way with human judgements.
Simon