Convolutional Neural Network and Language Embeddings for End-to-End Dialect Recognition

Abstract

Dialect identification (DID) is a special case of general language identification (LID), but a more challenging problem due to the linguistic similarity between dialects. In this paper, we propose an end-to-end DID system and a Siamese neural network to extract language embeddings. We use both acoustic and linguistic features for the DID task on the Arabic dialectal speech dataset: Multi-Genre Broadcast 3 (MGB-3). The endto-end DID system was trained using three kinds of acoustic features: Mel-Frequency Cepstral Coefficients (MFCCs), log Mel-scale Filter Bank energies (FBANK) and spectrogram energies. We also investigated a dataset augmentation approach to achieve robust performance with limited data resources. Our linguistic feature research focused on learning similarities and dissimilarities between dialects using the Siamese network, so that we can reduce feature dimensionality as well as improve DID performance. The best system using a single feature set achieves 73% accuracy, while a fusion system using multiple features yields 78% on the MGB-3 dialect test set consisting of 5 dialects. The experimental results indicate that FBANK features achieve slightly better results than MFCCs. Dataset augmentation via speed perturbation appears to add significant robustness to the system. Although the Siamese network with language embeddings did not achieve as good a result as the endto-end DID system, the two approaches had good synergy when combined together in a fused system. 1

Arabic natural language processing: An overview

“…Many works are oriented on Arabic identification (Ali et al, 2016;El Haj et al, 2017;Shon et al, 2018;Tachicart et al, 2017). These works were respectively presented by Ali et., El Haj et al, Shon et al and Tachicart et al They dealt with the Multi dialectal identification as well as the MSA one where the last one proposes an identification system distinguishing between the Moroccan dialect and MSA.…”

“…More recently Ali (2018) proposed a character-level convolution neural network model for distinguishing between MSA and multi dialects. The authors proposed a CNN model including five layers Shon et al (2018) proposed an end-to-end Dialect identification system and a Siamese neural network to extract language embeddings. The authors used acoustic and linguistic features on the Arabic dialectal speech dataset.…”

“…Much work targeted different aspects related to how this language and its dialects are processed, such as: Morphological analysis, resource building, Machine translation, etc. In order to present the characteristics of this language and to classify the works handling it, different surveys were proposed (Habash, 2010;Farghaly and Shaalan, 2009;Shoufan and Alameri, 2015). Habash gives a detailed overview of Arabic and its dialects by paying particular attention to different orthographic, morphological and syntactic aspects of this language (Habash, 2010).…”

“…Farghaly and Shaalan (2009) described some challenges of Arabic NLP (ANLP) and presented solutions that would guide practitioners in this field. In the most recent ANLP survey of Shoufan and Alameri (2015), the authors proposed a general classification of achievements for Arabic and its dialects which can be grouped into four categories: (1) Basic Language Analyses (BLA). (2) Building Resources (BR), (3) Language Identification (LI) and (4) Semantic-Level Analysis (SemA).…”

Domain Attentive Fusion for End-to-end Dialect Identification with Unknown Target Domain

“…In this work, we adopt the end-to-end dialect identification system proposed in [17]. This system has a stack of convolutional neural network (CNN) layers, followed by a global pooling layer that aggregates frame level representations to produce utterance level representations.…”