Pashto isolated digits recognition using deep convolutional neural network.

Area of Science:

• Computer Science
• Artificial Intelligence
• Signal Processing

Background:

• Human-computer interaction is increasingly reliant on speech recognition technologies.
• There is a growing demand for automatic speech recognition (ASR) systems tailored for local languages.
• Pashto, a widely spoken language, currently lacks robust ASR solutions for numerical recognition.

Purpose of the Study:

• To develop and evaluate a deep convolutional neural network (CNN) based ASR system for isolated Pashto digits (0-9).
• To assess the performance of the proposed system against existing methods for Pashto digit recognition.

Main Methods:

• A database of 50 utterances for each Pashto digit (0-9) was curated.
• Twenty Mel-frequency cepstral coefficients (MFCCs) were extracted as features for each digit.
• A deep CNN architecture with 4 convolutional layers, ReLU activation, and max-pooling layers was employed.
• The network was trained on 50% of the data and tested on the remaining 50%.

Main Results:

• The deep CNN model achieved an average accuracy of 84.17% for isolated Pashto digit recognition.
• This performance represents a 7.32% improvement compared to existing similar works in the field.
• The MFCC features effectively represented the Pashto digits for the CNN.

Conclusions:

• The developed deep CNN model demonstrates a viable and effective approach for Pashto isolated digit recognition.
• The system's improved accuracy highlights the potential of deep learning for low-resource language ASR.
• Further research can explore more complex speech tasks and larger datasets for Pashto.