Subjective evaluation of an emotional speech database for Basque Aholab Signal Processing Laboratory – University of the Basque Country Authors: I. Sainz, I. Saratxaga, E. Navas, I. Hernáez, J. Sanchez, I. Luengo, I. Odriozola
Index Introduction Corpus Design and Recording Evaluation Process Results Conclusions & Work in Progress
Introduction Speech Synthesis Intelligibility Naturalness Variability Emotional Speech Prosody Acoustic Features Corpus Based Tecniques Implicit Acoustic modeling
Corpus Design and Recording Aim Emotional Corpus-based TTS Prosodic and acoustic analysis Recording emotional speech Spontaneous Elicited Acted 2 speakers, 702 sentences 1 hour per emotion
Corpus Design and Recording Big 6 Emotions
Evaluation Process Subjetive evaluation campaign 30 stimuli per actor Forced choice 20 Subjects 11 native speakers
Results Confusion Matrix Recognition average 76.6% Far above chance level 17%
Results Confusion Matrices(Actress & Actor) Average 75.83% Average 76.50%
Results Effects of Listeners Student’s t-test Determine if the hypothesis is true Groups (recognition rate) Women (72.78%) Vs Men (77.62%) Not significant (95% confidence interval) Native speakers (77.12%) Vs Non-native (75%) Not significant (95% confidence interval) 1 st Half of the test (72.64%) Vs 2 nd Half (79.7%) t=2.85 p= < 0.05 Significant! Recognition rate increase of 7% –Almost constant for all groups
Conclusions & Work in progress Valid resource of emotional speech for Basque language Emotions readily recognized for both actors Study and modeling of emotional speech Corpus based TTS with emotions Neutral already developed
Subjective evaluation of an emotional speech database for Basque Aholab Signal Processing Laboratory – University of the Basque Country Authors: I. Sainz, I. Saratxaga, E. Navas, I. Hernáez, J. Sanchez, I. Luengo, I. Odriozola