Digit recognition from turkish sound signals with deep learning

In today's rapidly advancing technological landscape, speech-based recognition systems play a crucial role in various fields. Sound, beyond being a fundamental means of communication among individuals, serves as a critical factor in applications such as automation, security, and user experience. The effective utilization of sound in digital environments is made possible, particularly through the development of speech recognition technologies. These technologies have the capability to analyze sound signals, comprehend spoken language, and perform various tasks. Digital digit classification, particularly, constitutes a significant application area for these speech recognition technologies. Digital digit classification involves the development of systems that can accurately recognize and distinguish digital digits obtained from sound signals. This has a wide range of potential applications, from telecommunication systems to voice command systems and from speech-based security applications to various industrial and commercial applications. In this context, the importance of speech-based digital digit classification spans from everyday life to industrial applications. This study aims to contribute to technological advancements in this field by evaluating different machine learning models used in the process of digital digit classification from sound signals. SVM, LSTM, and CNN models were assessed for digital digit classification from sound signals, with the CNN model achieving the highest success rate at 81.94% in the 80-20 training-test ratio. The CNN model demonstrated high performance, particularly achieving a 98.2% success rate for the digit "Six (6)." Different success rates were observed among other digits, with high performance for "One (1)" and "Nine (9)" but lower success rates for "Three (3)," "Four (4)," and "Eight (8)." In the scope of the study, evaluations conducted under different training-test ratios revealed that the LSTM model exhibited the highest success at a 50-50 training-test ratio. SVM achieved its highest success rate at an 80-20 training-test ratio. However, overall, deep learning models, specifically LSTM and CNN, outperformed SVM, indicating that these numerical results highlight the effective use of deep learning models, especially in sound-based digital recognition applications. Key Words: CNN, CWT, LSTM, Sound digit classification