Browsing by Author "Kuanyshbay D.N."

Now showing 1 - 1 of 1
  • Thumbnail Image
    ItemOpen Access
    Development of methods, algorithms of machine learning for Kazakh speech recognition
    (Suleyman Demirel University, 2021) Kuanyshbay D.N.
    Automatic speech recognition system is on its way to reach its full potential mainly because of the significance of the field and rapid growth of the speech data. However, the growth of speech recognition for low-resourced languages like Kazakh is not observed, since there is not enough data and there is no automated tool for data collection. All existing speech recognition models for Kazakh language are based on datasets that were collected locally and manually, which makes these datasets private and inaccessible for other researchers to use them. Manual speech data collection requires a lot of time and effort, which may not match the quality requirements. Low quality speech data may ultimately affect the performance of speech recognition model considerably. This research work focuses on two important steps of building reliable automatic speech recognition system for Kazakh language – 1) construction of automated speech data collection system, 2) application of transfer learning to recurrent neural network. Automatic data collection system is based on a website, which perfectly segments the audio data and separates these audio files with corresponding transcriptions by speakers. As a result system produced around 100 hours of speech data, which in terms of structure is suitable for neural network to train. Transfer learning technique is based on Russian speech recognition model, which transfers all weights to the neural network that built for Kazakh language. Using transfer learning multilingual Automatic speech recognition model was obtained, which outperforms the simple LSTM based model by 32% and BiLSTM model by 24% (in terms of Label Error Rate).