The outliers in speech include three types:

  1. syntax errors,
  2. semantic errors, and
  3. errors in pragmatics.

I wonder if we can borrow the techniques from computer vision (CV) to tackle that, such as Fréchet Audio Distance inspired by the Fréchet Inception Distance in CV. In the paper on Fréchet Audio Distance, FID is applied to music, but the language would be much more complicated.

Any suggestions?

  • 1
    $\begingroup$ To me, it's not fully clear what is being asked here. What do you mean by "reference-free"? Moreover, the question in the title does not seem to be consistent with what you say in the body, which seems to suggest that you're concerned with how to deal with certain outliers and errors. So, please, clarify what your main specific question is and how it's related to what you're saying in the body of this post. $\endgroup$ – nbro Feb 9 at 12:08

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.