Yazar "Tufekci, Zekeriya" seçeneğine göre listele

Listeleniyor 1 - 2 / 2
  • Küçük Resim Yok
    Öğe
    Feature-based hybrid strategies for gradient descent optimization in end-to-end speech recognition
    (Springer, 2022) Dokuz, Yesim; Tufekci, Zekeriya
    With the increasing popularity of deep learning, deep learning architectures are being utilized in speech recognition. Deep learning based speech recognition became the state-of-the-art method for speech recognition tasks due to their outstanding performance over other methods. Generally, deep learning architectures are trained with a variant of gradient descent optimization. Mini-batch gradient descent is a variant of gradient descent optimization which updates network parameters after traversing a number of training instances. One limitation of mini-batch gradient descent is the random selection of mini-batch samples from training set. This situation is not preferred in speech recognition which requires training features to collapse all possible variations in speech databases. In this study, to overcome this limitation, hybrid mini-batch sample selection strategies are proposed. The proposed hybrid strategies use gender and accent features of speech databases in a hybrid way to select mini-batch samples when training deep learning architectures. Experimental results justify that using hybrid of gender and accent features is more successful in terms of speech recognition performance than using only one feature. The proposed hybrid mini-batch sample selection strategies would benefit other application areas that have metadata information, including image recognition and machine vision.
  • Küçük Resim Yok
    Öğe
    Mini-batch sample selection strategies for deep learning based speech recognition
    (Elsevier Sci Ltd, 2021) Dokuz, Yesim; Tufekci, Zekeriya
    With the use of deep learning technologies, speech recognition systems gained more success and human-computer interactions became more prevalent. Deep learning based speech recognition systems are getting more attention and are having tremendous success in all areas of speech recognition, such as voice search, mobile communication, and personal digital assistance. However, speech recognition is still challenging due to hardness of adapting new languages, difficulty in handling variations in speech datasets, and overcoming distorting factors. Deep learning systems have the ability to overcome these challenges using high-level abstractions in the datasets by using a deep graph with multiple processing layers using training algorithms, such as gradient descent optimization. In this study, a variant of gradient descent optimization, mini-batch gradient descent is used. We proposed four strategies for selecting mini-batch samples to represent variations of each feature in the dataset for speech recognition tasks to increase model performance of deep learning based speech recognition. For this purpose, gender and accent adjusted strategies are proposed for selecting mini-batch samples. The experiments show that proposed strategies perform better in comparison with standard mini-batch sample selection strategy. (C) 2020 Elsevier Ltd. All rights reserved.