Amazigh Speech Recognition via Parallel CNN Transformer-Encoder Model

Speech recognition technologies serve as a cornerstone in advancing the capabilities of artificial intelligence systems, providing a direct and efficient means of human communication. In the pursuit of fostering a more seamless and intuitive interaction between humans and automated systems, the incorporation of the Amazigh language into Automatic Speech Recognition technology holds profound implications. This integration not only contributes to the preservation of linguistic diversity but also facilitates enhanced accessibility while mitigating the complexities inherent in cross-language ASR transfer. In our research endeavor, we employ a novel approach combining convolutional neural networks with a Transformer encoder network to effectively capture spatial and temporal features for audio classification. Leveraging a dataset comprising 11,644 audio files representing 300 classes, encompassing 200 isolated words and 100 short sentences, recorded from native speakers of the Amazigh Tarifit language, our proposed model demonstrates promising results. Specifically, our model achieves a precision of 84.45% on the test set, underscoring its efficacy in accurately classifying Amazigh speech inputs.