National Repository of Grey Literature 2 records found  Search took 0.00 seconds. 
Methods of Domain Adaptation for Speech Recognition
Kratochvíl, Jonáš ; Bojar, Ondřej (advisor) ; Dušek, Ondřej (referee)
The goal of this thesis is to develop a complete pipeline of Automatic Speech recognition for the Czech language with a particular focus on effective adap- tation of the model across a variety of diverse domains. Due to the scarcity of training data, we introduce two approaches for data preparation. First, we segment a portion of our audio files in a fully unsupervised way and use them to train our baseline acoustic model. We then use this model for further refinement of the segments. With our data pipeline, we prepare over 1500 hours of training data for the Czech language, from which 444 hours are made available to the public under a non-restrictive license. For our experiments, we use the hybrid acoustic model that combines the Gaus- sian Mixture Model and Hidden Markov Model with Neural Network-based methods. We also present our approach to language modeling in which we hier- archically combine interpolated n-gram models and a recurrent neural network model used to re-score the output lattices. Experiments with acoustic adap- tation, which finetune the neural network to a small amount of target domain audios, are presented as well. Lastly, we introduce an efficient implementation of a model for sentence embeddings, which we use to query an extensive cor- pus database and condition the search on a...
Efficiency of Prague Stock Exchange Market using Markov Chains
Kratochvíl, Jonáš ; Červinka, Michal (advisor) ; Hausenblas, Václav (referee)
The main intention of this thesis is to analyze the weak form efficiency of Prague Stock Exchange. We conduct our empirical analysis on daily, weekly and monthly return data of the PX index collected in time period 1994-2017. The theory of Markov chains is employed to decide whether the index returns follow a random walk, the evidence of weak form efficiency. Bayesian Informa- tion Criterion is used to establish the optimal order of the Markov chain, which is in turn tested against the order 0 by Likelihood ratio criterion. The model assumptions of time homogeneity, irreducibility and aperiodicity of transition probability matrix are validated. We reject the weak form efficiency for daily index returns and establish its optimal Markov chain order to be 1. The weak form efficiency is not rejected for weekly and monthly index returns so is the as- sumption of time homogeneity for the whole time period 1994-2017. We propose further analysis of daily returns for time period 2006-2017, which exploits the fact of the weak form inefficiency. Discussion of results and related literature is provided as well as the presentation of all contemplated methods. 1

Interested in being notified about new results for this query?
Subscribe to the RSS feed.