Robot reads ads

Hille Pajupuu, Jaan Pajupuu, Rene Altrov, Indrek Kiissel

Institute of the Estonian Language, Tallinn, Estonia

The increasing prevalence of audio advertising has provided a challenge to find out more about voices and performance styles used in advertisements. In this study, we were interested in the listeners’ preferences when a synthesizer performs the advertisements. As training an advertisement style synthesizer requires big corpora, the creation of which is time-consuming and expensive, we have chosen to use less resource-intensive style transfer on already extant synthesized voices trained on neutral speech. We used a corpus of advertisements created out of 120 male and 120 female voices reading one text in both an energetic and calm advertisement style to train four style transfer models: energetic and calm for both male and female voices. These were used to convert two synthesized female and two male voices that had been created using a Merlin-based speech synthesizer for Estonian. Each converted voice performed three short advertisements. Adult listeners rated the likability of the performances on a 7-point Likert scale. The results showed that the calm performance style was overwhelmingly preferable/prefered. We will also ascertain the acoustic features of the calm and energetic performances using the open-source toolkit openSMILE to calculate the 88 parameters of the extended Geneva Minimalistic Acoustic Parameter Set (eGeMAPS). The calm style differed from the energetic in acoustic features, that are related to a lower, quieter and more sonorous voice and a more neutral speaking style. Considering the difference in style ratings, it is worth taking the target audiences’ style preferences into account.