Acoustic-Space Formulation of Target Function

[James B]

Taylor 16.4: For the target cost function, instead of using a feature representation (Hunt & Black), an ‘acoustic representation’ can be generated by ‘partial synthesis’. However, this does not produce an actual waveform, but an ‘approximate waveform’ – apparently still abstract, but able to be compared (in ‘perceptual space’) with real candidate units.

What is this representation? (Taylor mentions cepstra: would this be represented as MFCCs?) IF this method has managed to derive some acoustic MFCC-like representation from data, has it not done most of the work (except actual waveform synthesis) of a parametric synthesiser?

[Lars H]

Related question: how does this representation relate to HMMs? Taylor mentions that the acoustic representation of a target can match to a distribution of units, but how do we get the target specification for additional states? Do we interpolate the specification to neighbouring targets?

[Author]

Posts