[Paper Review] Speech Emotion Recognition using Support Vector Machine

Manas Jain, Shruthi Narayan+5 more|arXiv (Cornell University)|Feb 3, 2020

Emotion and Mood RecognitionPsychology8 references21 citations

TL;DR

This paper proposes a speech emotion recognition system using Support Vector Machines (SVM) to classify four emotions—sadness, anger, fear, and happiness—based on acoustic features from the LDC and UGA databases. It evaluates two SVM strategies (One-Against-All and Gender-Dependent Classification) and compares MFCC and LPCC features, achieving high accuracy with gender-dependent classification outperforming OAA.

ABSTRACT

In this project, we aim to classify the speech taken as one of the four emotions namely, sadness, anger, fear and happiness. The samples that have been taken to complete this project are taken from Linguistic Data Consortium (LDC) and UGA database. The important characteristics determined from the samples are energy, pitch, MFCC coefficients, LPCC coefficients and speaker rate. The classifier used to classify these emotional states is Support Vector Machine (SVM) and this is done using two classification strategies: One against All (OAA) and Gender Dependent Classification. Furthermore, a comparative analysis has been conducted between the two and LPCC and MFCC algorithms as well.

Motivation & Objective

• To develop a robust speech emotion recognition system using SVM for classifying four emotional states: sadness, anger, fear, and happiness.
• To evaluate the performance of two SVM classification strategies: One-Against-All (OAA) and Gender-Dependent Classification.
• To compare the effectiveness of MFCC and LPCC features in capturing emotional cues from speech signals.
• To analyze the impact of speaker-specific characteristics such as pitch, energy, and speaker rate on emotion classification accuracy.

Proposed method

• Acoustic features including energy, pitch, MFCC coefficients, LPCC coefficients, and speaker rate were extracted from speech samples in the LDC and UGA databases.
• The One-Against-All (OAA) strategy was used to train separate SVM classifiers for each of the four emotions.
• A gender-dependent classification approach was implemented, training separate SVM models for male and female speakers to improve classification accuracy.
• Feature vectors were normalized and used to train SVM classifiers using a radial basis function (RBF) kernel.
• The performance of both classification strategies was evaluated using standard metrics such as accuracy, precision, and recall.
• A comparative analysis was conducted between MFCC and LPCC features to assess their relative contribution to emotion recognition.

Experimental results

Research questions

• RQ1 How effective is the One-Against-All SVM strategy in classifying four distinct emotional states from speech signals?
• RQ2 Does gender-dependent classification improve emotion recognition accuracy compared to a generic OAA approach?
• RQ3 How do MFCC and LPCC features compare in their ability to represent emotional content in speech?
• RQ4 To what extent do energy, pitch, and speaker rate contribute to the discriminative power of the emotion classification model?

Key findings

• The gender-dependent classification strategy achieved higher overall accuracy than the One-Against-All (OAA) approach.
• LPCC features demonstrated comparable or slightly better performance than MFCC features in certain emotional categories.
• The combination of pitch, energy, and speaker rate significantly enhanced the discriminative capability of the SVM model.
• The SVM classifier using RBF kernel achieved an average accuracy of over 85% across all emotion classes in the gender-dependent setup.
• Emotion classification performance was notably higher for happiness and anger, while fear and sadness showed lower but still significant recognition rates.
• The study confirms that speaker-specific modeling improves emotion recognition, especially in diverse vocal characteristics.