Abstract
Many speaker language detection systems depend on deep learning (DL) approaches, and utilize long recorded audio periods to achieve satisfactory accuracy. This study aims to extract features from short recording audio files that are convenient in order to detect the spoken languages under test successfully. This detection process is based on audio files of (1 or 2) seconds whereas most of the previous languages Classification systems were based on much longer time frames (from 3 to 10 seconds). This research defined and implemented many low-level features using Mel Frequency Cepstral Coefficients (MFCCs), the dataset compiled by the researcher containing speech files in three languages (Arabic, English. Kurdish), which is called M2L_dataset is the source of data used in this paper.
A Bidirectional Long Short-Term Memory (BiLSTM) algorithm applied in this paper for detection speaker language and the result was perfect, binary language detection had a test accuracy of 100%, and three languages detection had a test accuracy of 99.19%.
)
