Navajo Speech Recognition Using Low-Resource Language Models

The paper describes the development of a speech recognition system to classify Navajo (Diné) words using Low Resource Language (LRL) datasets. There are presently no known high-quality open-sourced datasets for the Diné language that are needed to train models for speech Recognition. A small dataset was designed to train several models. To overcome the scarcity of the LRL dataset, the audio recordings were augmented to account for time-stretching, amplitude variations, time shifts, and small amounts of white Gaussian noise. Several models were trained using different optimization methods. The models included a Recurrent Neural Network (RNN), a Convolutional Neural Network (CNN), and a Long Short-Term Memory (LSTM) model. The results compare nine methods: SGD, Momentum, Nesterov, AdaGrad, RMSProp, Adam, Adamax, Nadam, and AdamW. For the best model, we report a family of different training/validation curves. The results demonstrate excellent classification performance on LRL models for Navajo Speech Recognition.