A comparison of speech recognition techniques and models

Abstract

There has been an increase in speech recognition software used in many devices. Primarily it had been introduced to the public mostly on smartphones. As technology is improving speech recognition software is found on different devices and household appliances such as televisions and even refrigerators. As people are using speech to text software more frequently this form of computer interaction is becoming the norm and so inaccuracies, such as words misunderstood for others, may cause issues in the future. The project aims to make comparisons to understand which model is best adapted when tested against the obtained dataset. This study will present a better view of how these devices and appliances work since speech recognition software seems to be becoming an attractive form of computer interaction. The project delineates different topics such as sound and pre-processing sound recordings. It then focuses on feature extraction, which retrieves key characteristics of sound waves, to be used in the speech recognition stage. An example of this is known as the ‘Mel-frequency cepstral coefficients’ (MFCC) that can be presented graphically by using a spectrogram. Ultimately, the data from the feature extraction process is fed into the different models. Three models, the Hidden Markov Model, a Convolutional Neural Network and the Dynamic Time Warping Algorithm, were selected and compared using the same datasets, and the results were evaluated using accuracy and running time.