Search

End-to-end deep learning approach has greatly enhanced the performance of speech recognition systems. With deep learning techniques, the overfitting stills the main problem with a little data. Data augmentation is a suitable solution for the overfitting problem, which is adopted to improve the quantity of training data and enhance robustness of the models. In this paper, we investigate data augmentation method for enhancing Arabic automatic speech recognition (ASR) based on end-to-end deep learning. Data augmentation is applied on original corpus for increasing training data by applying noise adaptation, pitch-shifting, and speed transformation. An CNN-LSTM and attention-based encoder-decoder method are included in building the acoustic model and decoding phase. This method is considered as state-of-art in end-to-end deep learning, and to the best of our knowledge, there is no prior research employed data augmentation for CNN-LSTM and attention-based model in Arabic ASR systems. In addition, the language model is built using RNN-LM and LSTM-LM methods. The Standard Arabic Single Speaker Corpus (SASSC) without diacritics is used as an original corpus. Experimental results show that applying data augmentation improved word error rate (WER) when compared with the same approach without data augmentation. The achieved average reduction in WER is 4.55%.

Numerous researches on content-based video indexing and retrieval besides video search engines are tied to a large-scaled video dataset. Unfortunately, reduction in open-sourced datasets resulted in complications for novel approaches exploration. Although, video datasets that index video files located on public video streaming services have other purposes, such as annotation, learning, classification, and other computer vision areas, with little interest in indexing public video links for purpose of searching and retrieval. This paper introduces a novel large-scaled dataset based on YouTube video links to evaluate the proposed content-based video search engine, gathered 1088 videos, that represent more than 65 hours of video, 11,000 video shots, and 66,000 unmarked and marked keyframes, 80 different object names used for marking. Moreover, a state-of-the-art features vector, and combinational-based matching, beneficial to the accuracy, speed, and precision of the video retrieval process. Any video record in the dataset is represented by three features: temporal combination vector, object combination vector with shot annotations, and 6 keyframes, sideways with other metadata. Video classification for the dataset was also imposed to expand the efficiency of retrieval of video-based queries. A two-phased approach has been used based on object and event classification, storing video records in aggregations related to feature vectors extracted. While object aggregation stores video records with the maximal occurrence of extracted object/concept from all shots, event aggregation classify based on groups according to the number of shots per video. This study indexed 58 out of 80 different object/concept categories, each has 9 shot number groups.