Speech Acoustic Model Using Deep Learning

The article "Discriminative Keyword Spotting For Limited-Data Applications" written by the Technion Electrical Engineering faculty members, is describing a system for word classification for cases where there is a small labeled dataset (and a large unlabeled dataset) using GMM as a statistic model for the language. Following this article, an idea was raised that using DNN instead of GMM as a statistic model for the language might provide better results. The idea came from the fact that DNN is currently the leading model for machine learning. Also, the DNN model has shown to provide the best results for speech recognition comparing to other models.
The goal of this project was to build an alternative model taken from an existing neural network, built according to the article "Sequence-discriminative training of deep neural networks" and using it instead of the existing GMM model.
The results showed that there was not statistical advantage for any system over the other.

The article "Discriminative Keyword Spotting For Limited-Data Applications" written by the Technion Electrical Engineering faculty members, is describing a system for word classification for cases where there is a small labeled dataset (and a large unlabeled dataset) using GMM as a statistic model for the language. Following this article, an idea was raised that using DNN instead of GMM as a statistic model for the language might provide better results. The idea came from the fact that DNN is currently the leading model for machine learning. Also, the DNN model has shown to provide the best results for speech recognition comparing to other models. The goal of this project was to build an alternative model taken from an existing neural network, built according to the article "Sequence-discriminative training of deep neural networks" and using it instead of the existing GMM model. The results showed that there was not statistical advantage for any system over the other.
The article “Discriminative Keyword Spotting For Limited-Data Applications” written by the Technion Electrical Engineering faculty members, is describing a system for word classification for cases where there is a small labeled dataset (and a large unlabeled dataset) using GMM as a statistic model for the language. Following this article, an idea was raised that using DNN instead of GMM as a statistic model for the language might provide better results. The idea came from the fact that DNN is currently the leading model for machine learning. Also, the DNN model has shown to provide the best results for speech recognition comparing to other models.
The goal of this project was to build an alternative model taken from an existing neural network, built according to the article “Sequence-discriminative training of deep neural networks” and using it instead of the existing GMM model.
The results showed that there was not statistical advantage for any system over the other.