COMPARISON OF MACHINE LEARNING METHODS AND DEEP LEARNING-BASED MODELS FOR EMOTION RECOGNITION IN TURKISH SPEECH

Zekeriya Anıl Güven

Makale detayları

COMPARISON OF MACHINE LEARNING METHODS AND DEEP LEARNING-BASED MODELS FOR EMOTION RECOGNITION IN TURKISH SPEECH

Yazarlar:

Zekeriya Anıl Güven

Yayıncı:

Süleyman Demirel University

Ana Başlıklar:

Bilgisayar Yazılımı | Yazılım Mühendisliği (Diğer)

Dergi:

Mühendislik Bilimleri ve Tasarım Dergisi Vol.12 Issue 2, pp.285-297

Belge Tipi:

Makale

Yayın Tarihi:

6/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001321511

pISSN:

1308-6693

Anahtar kelimeler:

Özet:

 Recently, depending on the amount of data, analyzes are needed in many areas such as health, education and marketing. Sentiment analysis is a very popular area for analyzing peoples comments and extracting their emotions in these areas. This study aims to determine the emotions by using voice characteristics and spectrograms on the Turkish speech dataset containing angry, happy, calm and sad emotion labels. In the analysis phase, machine learning methods and deep neural networks were trained, and their success was measured with the numerical features extracted from Librosa library. In addition, feature reduction was applied with low variance filtering, backward feature elimination, chi-square and principal component analysis methods, and the changes in the success of machine learning methods were also investigated with the new features obtained. The spectrograms, which are visual data, were used to train EfficientNet, ResNet, MobileNet and DenseNet deep learning-based models. During the training phase of these models, fine-tuning process was applied to these models together with the dataset. As a result of the experimental studies, Extreme Gradient Boosting, one of the machine learning methods, gave an accuracy value of 87.03, while the ResNet model gave an accuracy value of 79.23.

 Recently, depending on the amount of data, analyzes are needed in many areas such as health, education and marketing. Sentiment analysis is a very popular area for analyzing peoples comments and extracting their emotions in these areas. This study aims to determine the emotions by using voice characteristics and spectrograms on the Turkish speech dataset containing angry, happy, calm and sad emotion labels. In the analysis phase, machine learning methods and deep neural networks were trained, and their success was measured with the numerical features extracted from Librosa library. In addition, feature reduction was applied with low variance filtering, backward feature elimination, chi-square and principal component analysis methods, and the changes in the success of machine learning methods were also investigated with the new features obtained. The spectrograms, which are visual data, were used to train EfficientNet, ResNet, MobileNet and DenseNet deep learning-based models. During the training phase of these models, fine-tuning process was applied to these models together with the dataset. As a result of the experimental studies, Extreme Gradient Boosting, one of the machine learning methods, gave an accuracy value of 87.03, while the ResNet model gave an accuracy value of 79.23.

 Recently, depending on the amount of data, analyzes are needed in many areas such as health, education and marketing. Sentiment analysis is a very popular area for analyzing peoples comments and extracting their emotions in these areas. This study aims to determine the emotions by using voice characteristics and spectrograms on the Turkish speech dataset containing angry, happy, calm and sad emotion labels. In the analysis phase, machine learning methods and deep neural networks were trained, and their success was measured with the numerical features extracted from Librosa library. In addition, feature reduction was applied with low variance filtering, backward feature elimination, chi-square and principal component analysis methods, and the changes in the success of machine learning methods were also investigated with the new features obtained. The spectrograms, which are visual data, were used to train EfficientNet, ResNet, MobileNet and DenseNet deep learning-based models. During the training phase of these models, fine-tuning process was applied to these models together with the dataset. As a result of the experimental studies, Extreme Gradient Boosting, one of the machine learning methods, gave an accuracy value of 87.03, while the ResNet model gave an accuracy value of 79.23.

 Son zamanlarda veri miktarına bağlı olarak sağlık, eğitim, pazarlama gibi birçok alanda analizlere ihtiyaç duyulmaktadır. Duygu analizi ise bu alanlarda kişilerin yorumlarını analiz etme, duygularını çıkarma için oldukça popüler bir alandır. Bu çalışmada kızgın, mutlu, sakin ve üzgün duygu etiketleri içeren Türkçe konuşma veri seti üzerinde, ses karakteristik özellikleri ve spektrogramlardan yararlanarak duyguların tespit edilmesi amaçlanmaktadır. Analiz aşamasında Librosa kütüphanesi ile çıkarılan sayısal özellikler ile makine öğrenme yöntemleri ve derin sinir ağları eğitilerek başarıları ölçülmüştür. Ayrıca düşük varyans filtreleme, geri yönlü özellik eleme, ki-kare ve temel bileşen analizi yöntemleri ile özellik azaltım işlemi uygulanarak elde edilen yeni özellikler ile makine öğrenme yöntemlerinin başarısındaki değişiklikler de araştırılmıştır. Görsel veri olan spektrogramlar ise EfficientNet, ResNet, MobileNet ve DenseNet derin öğrenme tabanlı modellerin eğitilmesi için kullanılmıştır. Modellerin eğitim aşamasında veri seti ile beraber modellere ince ayar işlemi uygulanmıştır. Deneysel çalışmaların sonucunda makine öğrenme yöntemlerinden Ekstrem Gradient Artırma 87.03 doğruluk değeri verirken, ResNet modeli ise 79.23 doğruluk değeri vermiştir.

Makale detayları

COMPARISON OF MACHINE LEARNING METHODS AND DEEP LEARNING-BASED MODELS FOR EMOTION RECOGNITION IN TURKISH SPEECH

Yazarlar:

Zekeriya Anıl Güven

Yayıncı:

Süleyman Demirel University

Ana Başlıklar:

Bilgisayar Yazılımı | Yazılım Mühendisliği (Diğer)

Dergi:

Mühendislik Bilimleri ve Tasarım Dergisi Vol.12 Issue 2, pp.285-297

Belge Tipi:

Makale

Yayın Tarihi:

6/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001321511

pISSN:

1308-6693

Sentiment analysis

Voice analysis

Machine learning

Deep learning models

Emotion RecognitionDuygu analizi

Ses analizi

Makine öğrenmesi

Derin öğrenme modelleri

Duygu tanıma

Sayı

Makale detayları

Yazarlar: Zekeriya Anıl Güven

Yayıncı: Süleyman Demirel University

Ana Başlıklar: Bilgisayar Yazılımı | Yazılım Mühendisliği (Diğer)

Dergi: Mühendislik Bilimleri ve Tasarım Dergisi Vol.12 Issue 2, pp.285-297

Belge Tipi: Makale

Yayın Tarihi: 6/2024

Belge Dili: Türkçe

Document ID (Org No.): 2001321511

pISSN: 1308-6693

Sayı

Yazarlar:

Zekeriya Anıl Güven

Yayıncı:

Süleyman Demirel University

Ana Başlıklar:

Bilgisayar Yazılımı | Yazılım Mühendisliği (Diğer)

Dergi:

Mühendislik Bilimleri ve Tasarım Dergisi Vol.12 Issue 2, pp.285-297

Belge Tipi:

Makale

Yayın Tarihi:

6/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001321511

pISSN:

1308-6693