An Urdu speech corpus for emotion recognition.

Area of Science:

• Computational Linguistics
• Speech Processing
• Machine Learning

Background:

• Emotion recognition from acoustic signals is crucial for human-computer interaction (HCI) and behavior analysis.
• Urdu language lacks a dedicated emotional speech database, hindering research in this domain.
• Speech interfaces enhance natural human-machine communication.

Purpose of the Study:

• To develop the first emotional speech database for the Urdu language.
• To create a system for classifying five distinct emotions (sadness, happiness, neutral, disgust, anger) in Urdu speech.
• To evaluate the effectiveness of various machine learning algorithms and speech descriptors for emotion recognition.

Main Methods:

• Collected an emotional speech corpus from 20 Urdu speakers.
• Extracted speech features including Mel Frequency Cepstrum Coefficients (MFCC), Linear Prediction Coefficients (LPC), energy, spectral flux, spectral centroid, spectral roll-off, and zero-crossing.
• Employed machine learning algorithms, including K-nearest neighbors, for emotion classification and conducted subjective listening tests for evaluation.

Main Results:

• The K-nearest neighbors algorithm achieved 66.5% accuracy in classifying five emotions within the Urdu emotional speech corpus.
• Disgust emotion exhibited a lower recognition rate compared to other emotions.
• Excluding disgust significantly improved the overall classifier performance to 76.5%.

Conclusions:

• The developed Urdu emotional speech database and classification system represent a significant contribution to speech emotion recognition research.
• Machine learning models show promise for Urdu speech emotion analysis, though challenges remain, particularly with recognizing disgust.
• Further research can refine models to enhance the recognition of subtle emotions and improve overall accuracy in Urdu speech.