Synthesising emotion with prosody control

The techniques for synthesizing emotion have followed closely the general developments in speech-synthesis algorithms. First-generation techniques include the work by Murray [324] and Calm [79], [80], who used formant synthesisers, and thus were able to vary every parameter of the synthesiser as desired. The experimental paradigm used 

As techniques become in general more data-driven, the issue of how to collect emotional data comes to the fore. There are basically two approaches. One can collect natural, perhaps spontaneous speech from a conversation and then label each sentence with regard to its emotional content. This produces the most-natural data, but has the difficulty that the speech then has to be labelled, difficulties which should not be lightly dismissed as we saw in Section 17.1.4. Databases collected in this way are described by Campbell [81] and Douglas-Cowie et al. [112]. An alternative strategy is to have actors produce sentences intentionally spoken with particular emotions. Since the type of emotion is an input to this process, no labelling is required. This approach has a different problem in that the resultant speech might not be particularly natural an actor s version of a particular emotion may be too hammed up . 

---