Aside from subjective listening tests, are there other ways of measuring the 'naturalness' of speech, especially synthetic speech?

Are there specific features of the speech signal that correlate well with naturalness scores? Alternatively, can a neural net be tuned to a dataset using speech as the input, and naturalness scores as the dependent variable/classification, using feature engineering?

  • No idea about the first question, but as for the second question, a neural net can be trained to decide whether the input is "natural" - or, more precisely, whether the input was from the original training data or a synthetic output. That's the basic premise of GAN (generative adversarial network). – jick Jul 27 '20 at 17:51

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.