New speech sound features for inference and feedback.

Using variational autoencoders we will build low dimensional latent space generative models of human vowel sounds which can be used to reconstruct vowels and measure their properties. We will determine the curves in latent space which correspond to meaningful linguistic variations (length, vocal tract size, height, backness, nasality, etc) guided by the input of an expert linguist. The model will be useful both for speech analysis and feedback, and for making automatic improvements to speech signal intelligability by transforming the raw speech signal. The methods will have applications in linguistic analysis, voice training, language learning, and speech signal processing.