Bebek Ağlamalarının Makine Öğrenmesi Yöntemleriyle Sınıflandırılması


Creative Commons License

Durak B. Ş., Bayındır L.

Avrupa Bilim ve Teknoloji Dergisi, sa.27, ss.784-791, 2021 (Hakemli Dergi)

Özet

Bu çalışmada bebek ağlama seslerinden yararlanılarak, bebeklerin ihtiyaçlarını belirleyebilmek ve herhangi bir hastalığa maruz olup olmadıklarını öğrenmek için çeşitli makine öğrenmesi yöntemleri kullanılmıştır. Bebek ağlama sesleri için iki farklı veri kümesinden yararlanılmıştır. Veri kümeleri eğitim, öğrenme ve test aşamalarından geçtikten sonra ağlama türünün bulunması hedeflenmiştir. Ağlama seslerinin özniteliklerini çıkarmak için Mel Frekans Kepstrum Katsayıları (MFCC) ve Doğrusal Öngörü Kepstrum Katsayıları (LPCC) yöntemleri karşılaştırılmış ve MFCC yönteminin doğruluğu arttırmada LPCC’ye göre daha etkili olduğu bulunmuştur. Öznitelikleri çıkarılan ses sinyallerinin sınıflandırılma aşamasında makine öğrenme algoritmalarından k-En Yakın Komşuluk (k-NN) algoritması, Çok Katmanlı Algılayıcılar, Karar Ağacı ve Rastgele Orman algoritmaları kullanılmış ve başarı oranları karşılaştırılmıştır. Sonuçta MFCC ve Çok Katmanlı Algılayıcılar yöntemi kullanılarak %93 başarı oranı bulunmuştur. 

In this study, various machine learning methods were used to determine the needs of babies and to find out whether they were exposed to any disease by making use of baby crying sounds. Two different data sets were used for baby crying sounds. It is aimed to find the type of crying after the data sets go through the training, learning and testing stages. Mel Frequency Cepstral Coefficients (MFCC) and Linear Prediction Cepstral Coefficients (LPCC) methods were compared to extract the attributes of crying sounds, and the MFCC method was found to be more effective than LPCC in increasing accuracy. In the classification phase of the audio signals whose attributes were extracted, k-Nearest Neighborhood (k-NN) algorithm, Multilayer Perceptron, Decision Tree and Random Forest algorithms from machine learning algorithms were used and success rates were compared. As a result, a 93% success rate was found using MFCC and Multilayer Perceptron.