Inception Model for Automatic Arabic Speech Recognition

Abstract

Reproducing basic human abilities has always been the main purpose for Artificial Intelligence (AI) systems. Since speech is essential to people’s communication, AI was applied to this major field to achieve Automatic Speech Recognition (ASR). In this paper, we focus on the inception model as a solution for Arabic speech recognition, due to its remarkable results on image classification tasks. We adapted this model for ASR problems and tried it on a dataset of spoken Arabic digits collected from social media apps and published corpora which resulted in more than 54000 utterances. A comparison between the proposed model and a traditional Convolutional Neural Network (CNN) shows the superiority of the inception model in ASR tasks. The inception model achieved 99.70% accuracy on the training dataset which is far better than the traditional CNN that achieved 87.46% on the same set, it did also great performance on the test subset with 88.96% accuracy compared to the traditional model with 84.78% recognition rate.

Keywords

• Inception model
• Convolution neural network
• Artificial intelligence
• GoogLeNet

References

1. an, W., Zhang, Z., Zhang, Y., Yu, J., Chiu, C.-C., Qin, J., Gulati, A., Pang, R., & Wu, Y. (2020). ContextNet: improving ocnvolutional neural networks for automatic speech recognition with global context. arXiv.
2. Hourri, S., Nikolov, N. S., & Kharroubi, J. (2021). Convolutional neural network vectors for speaker recognition. International Journal of Speech Technology, 24(2), 389–400.
3. Kiranyaz, S., Avci, O., Abdeljaber, O., Ince, T., Gabbouj, M., & Inman, D. J. (2021). 1D convolutional neural networks and applications: A survey. Mechanical Systems and Signal Processing, 151, 107398.