# Is Speech Emotion Recognition Language-Independent? Analysis of English and Bangla Languages using Language-Independent Vocal Features

**Authors:** Fardin Saad, Hasan Mahmud, Md. Alamin Shaheen, Md. Kamrul Hasan, Paresha Farastu  
**arXiv:** [2111.10776](https://arxiv.org/abs/2111.10776)  
**Date:** 2021-11-21  
**Summary page:** https://www.summarizepaper.com/en/paper/2111.10776v1/

## Key points

- Researchers explored Speech Emotion Recognition (SER) to determine language-independence
- Emotions studied: happiness, anger, neutral, sadness, disgust, fear
- Three Emotional Speech Sets used: two by Bengali speakers in Bangla and English, one by English speakers from Canada
- Language-independent prosodic features analyzed with SVM model for classification
- Three experiments conducted to test hypothesis: evaluating individual speech sets' performance, combining for unified context classification rate measurement, training/testing with different sets for recognition rate assessment
- SER is predominantly language-independent but discrepancies noted in recognizing disgust and fear across languages
- Non-native speakers can convey emotions through speech similar to native expression
- Vocal feature extraction done using Praat software's signal processing techniques
- Prosodic features like pitch median/mean/standard deviation/intensity captured emotional nuances aiding analysis
- Study highlights interplay between language/emotion recognition in speech processing systems and cross-linguistic applicability in SER research

## Summary

In this study, researchers explored the realm of Speech Emotion Recognition (SER) to determine its language-independence. Utilizing Bangla and English languages, they assessed the feasibility of distinguishing emotions from speech regardless of linguistic differences. The emotions under scrutiny included happiness, anger, neutral, sadness, disgust, and fear. Three Emotional Speech Sets were employed for investigation: two crafted by native Bengali speakers in both languages and one developed by native English speakers from Canada. Meticulously selected language-independent prosodic features were analyzed using a Support Vector Machine (SVM) model for classification purposes. Three distinct experiments were conducted to test the hypothesis: evaluating individual speech sets' performance, combining them for unified context classification rate measurement, and training/testing with different sets for comprehensive recognition rate assessment. Findings suggested that SER is predominantly language-independent; however, discrepancies were noted in recognizing disgust and fear across Bangla and English languages. Non-native speakers are capable of conveying emotions through speech similar to their native tongue expression. Detailed insights into methodology revealed vocal feature extraction via Praat software's signal processing techniques. Language-independent prosodic features such as pitch median/mean/standard deviation/intensity captured emotional nuances within speech patterns significantly aiding analysis. This study sheds light on intricate interplay between language/emotion recognition in speech processing systems while underscoring potential cross-linguistic applicability in SER research endeavors.