Sarcasm recognition in speech using a real-time approach
This study focuses on the recognition of sarcasm in speech, and attempts to address the problem of inaccuracy with regard to identifying this particular audio signal. This problem widens the gap between computers and humans since interactions are not completely understandable using sarcastic comment...
Saved in:
| Main Author: | |
|---|---|
| Format: | text |
| Language: | English |
| Published: |
Animo Repository
2015
|
| Subjects: | |
| Online Access: | https://animorepository.dlsu.edu.ph/etd_masteral/5073 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Institution: | De La Salle University |
| Language: | English |
| Summary: | This study focuses on the recognition of sarcasm in speech, and attempts to address the problem of inaccuracy with regard to identifying this particular audio signal. This problem widens the gap between computers and humans since interactions are not completely understandable using sarcastic comments. With this in mind, the goal is to create a model capable of identifying sarcasm which is generic enough to work not only on a specific set but also on any kind of sarcastic statements using audio signals. This was accomplished using machine learning and digital signal processing techniques appropriate for real time processing. Audio features like pitch, intensity, Mel Frequency Cepstral Coefficients (MFCC), and formants were experimented on using a new acted speech corpus that was annotated as sarcastic and non sarcastic by six participants which include the researcher. By using Support Vector Machine with polynomial kernel on a data set containing 0.4 second segments with 30% overlap, an accuracy and kappa of 69% and 0.39, respectively. The results suggest that pitch, intensity and certain MFCC and formant features are predictive of sarcasm. With only 10 features, SVM with polynomial kernel processes a single 0.4 second clip in 1.2 seconds making it suitable for real time processing. |
|---|