| Summary: | Multilingual based voice activated human computer interaction systems are currently in high demand. The Spoken Language Identification Unit (SPLID) is an inevitable front end unit of such a multilingual system. These systems will be a great boon to a country like India where around 24 official languages are spoken. Deep learning architectures for spoken language identification have progressed to the point that they can now perform well, even in the presence of various background noises. However, a strong phonetic relationship across various Indian languages leads to increased confusion in the SPLID unit. Therefore, the goal of this study is to propose a synthetic voice data augmentation method based on speech synthesis to improve the spoken Indian language identification system. Here the research attempts to determine how well pre-trained computer vision models recognize spoken languages in synthetic and classical audio augmentation environments. The accuracy of the models was compared using bottleneck features extracted from three different pre-trained models VGG16, RESNET50, and Inception-v3 while using an Artificial Neural Network (ANN), Support Vector Machine (SVM), Logistic Regression (LR), Random Forest (RF), Naive Bayes (NB), Decision Tree (DT) and KNN (K-Nearest Neighbors) as classifiers. The proposed system was tested on three Indian language datasets - two comprising seven Indian languages (Hindi, Malayalam, Tamil, Telugu, Marathi, Kannada and Bengali), one containing five Indian languages (Tamil, Hindi, Malayalam, Oria and Assamese), and on a foreign language dataset. It was found that the addition of synthetic audio samples improved the accuracy by 17%. Among the pre-trained models, VGG16 and Inception-v3 combined with PCA and ANN were found to have the maximum accuracy of 97%.
|
|---|