Two Vocoder Techniques for Neutral to Emotional Timbre Conversion

In this paper, we describe the application of two vocoder techniques for an experiment of spectral envelope transformation. We processed speech data in a neutral standard reading style in order to reproduce the spectral shapes of two emotional speaking styles: happy and sad. This was achieved by means of conversion functions which operate in the frequency domain and are trained with aligned source-target pairs of spectral features. The first vocoder is based on the source-filter model of speech production and exploits the Mel Log Spectral Approximation filter, while the second is the Phase vocoder. Objective distance measures were calculated in order to evaluate the effectiveness of the conversion framework in predicting the target spectral envelopes. Subjective listening tests also provided interesting elements for the evaluation.