I am trying to figure out speech features that can be inverted to recover the audio back with no loss or as less loss as possible.

I am currently using the power spectrum which is completely reversible.

I also read about MFCC and LPC as potential features but both of them seem to be lossy in terms of recovering the audio.

Also is there a way to compute the power spectrum in the Wavelet domain and is it any better a feature than the power spectrum computed using FFT?

If there are any other features it would be great to know about that.

Thanks for the help.

  • $\begingroup$ whether the wavelet spectrum is any better than the FFT depends on what you're trying to accomplish $\endgroup$ – Aaron May 31 '14 at 3:42
  • $\begingroup$ In general, you're going to have problems recovering, say, 800 samples of waveform from, say, 12 MFCC coefficients; you just don't have the degrees of freedom. Most often, the reason we use features (rather than the waveform) is to reduce dimensionality by discarding information irrelevant to the particular task (e.g. phoneme classification) we're concerned with. If you want both features and waveform, why not just store both the features and the waveform? $\endgroup$ – dpwe Jul 16 '14 at 19:34

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.