Whither the Artificial Voice?

In order to explore the issue of artificiality, it would first seem to be necessary to specify what we mean by non-artificial speech.  For example, if we define natural speech as being the linguistic behaviour exhibited by a normal human vocal tract in a quiet non-reverberant environment, then we can start to distinguish different ways in which artificiality might be introduced.

For example, it is possible to itemise the following processes:

• speech transformation (processes which alter the characteristics of a speech signal);
• speech sampling (processes based on the selection of segments of pre-recorded speech);
• speech simulation (processes which mimic some physical aspect of the speech production process); and,
• speech synthesis (methods for creating a speech signal from a parametric specification).

Examples of speech transformation would be reverberant environments, communication channels, vocoders, pitch modifiers, voice morphers, etc.  An example of speech sampling would be a concatenative/unit-selection text-to-speech system. Examples of speech simulators would be articulatory models and the Sheffield animatronic tongue/vocal tract (AnTon - http://staffwww.dcs.shef.ac.uk/people/R.Hofe/anton/anton.html).  Examples of speech synthesis would be formant synthesis or hidden-Markov model based synthesis.

Given the definition of natural speech, then, human speech emanating from a non-human source would also be classified as artificial!  I.e. even telephone speech could be viewed as artificial.

Roger Moore
Posted 20th September 2011 at 5:19 PM