Likability: Speaker Likability Database Sample Clauses

Likability: Speaker Likability Database. The Speaker Likability Database (SLD) (Xxxxxxxxx et al., 2010) was the official corpus of the Likability Sub- Challenge in the INTERSPEECH 2012 Speaker Trait Chal- lenge (STC) (Xxxxxxxx et al., 2012a). The speech is recorded over fixed and mobile telephone lines at a sample rate of 8 kHz. An age and gender balanced set of 800 speakers is selected. For each speaker, the longest sentence (con- sisting of a command embedded in a free sentence) was selected. Likability rating was executed by 32 labellers ac- cording to how well they personally liked the voices. They were asked not to take into account the linguistic content or the transmission quality. The rating was done on a seven point Likert scale ([-0, -0, -0, 0, 1, 2, 3]). To establish a co- herent consensus from the highly individual likability rat- ings, the evaluator weighted estimator (EWE) (Xxxxx and Xxxxxxxx, 2005) was used in the challenge. It uses higher weights for more agreeable labellers. Based on the median EWE rating of all stimuli in the SLD, the data was discre- tised at the threshold of 0.108 into the classes—‘likable’ (L, EWE > 0.108) and ‘non-likable’ (NL, EWE < 0.108). The final challenge set of 800 instances is partitioned as follows: training set (L, 189; XX, 000), development set (L, 92; NL, 86), and test set (L, 119; NL, 109). Fig. 1 (b) shows the instance distribution along with human agreement lev- els. Here, we slightly modify Kappa as follows: