An Investigation in Bahraini Dialect Text to Speech Synthesis Models and Datasets

Abstract

Abstract : In the past few years, there has been an accelerated development in the field of Artificial Intelligence (AI) and Machine Learning (ML). Various Text To Speech (TTS) modules have been developed, and successfully implemented in many languages. However, implemen-tations of such systems for the Arabic language has been minimal, with implementations mostly relying on makeshift solutions such as using diacritics to transliterate the Arabic text to English. This study presents the development of a Text To Speech (TTS) system for the Bahraini dialect. The proposed approach uses the Forward Tacotron model integrated with HiFi-GAN vocoder, and Optuna for hyperparameter optimization. Our approach pro-vides a thorough understanding of the requirements for synthesizing natural and expressive Bahraini Arabic speech, and investigates the possibility of training such a system with-out the use of diacritics or phonetization tools. For the purpose of this study we created a dataset in the Bahraini dialect, containing 9 hours of recordings collected from Sanaa Alsaad's YouTube channel. However, due to the small dataset size, data augmentation techniques were used expand the dataset to 36 hours. Another version of the dataset was created spanning 8 hours of augmented data and involving only higher quality recordings. The smaller dataset was then AI enhanced using Adobe Podcasts. The sample rate of the recorded data is 22050 Hz, as per the model's requirements. Our system was first trained on the 36 hours dataset, and then transfer learning was used to fine tune the training on the 8 hours dataset. The system was assessed using metrics such as Attention Score, Mel Loss, Duration Loss, Pitch Loss, Word Error Rate (WER), and Mean Opinion Score (MOS). The model demonstrated promising capabilities in replicating the local dialect with an MOS of 3.4, and a WER of 34.