• Modern Methods for the Recognition of Strongly Distorted Based on Deep Neural Networks
  Nowadays, speech recognition systems work generally with high accuracy under standard conditions, however, the accuracy of speech recognition strongly declines when spech is strongly distorted (e.g. when the speech is recorded by a far-talk microphone, speech corrupted by high-level background noise of Lombard effect). The research related to this topic will be focused mainly on an analysis, selection, and optimization of suitable architecture of DNN-HMM based acoustic modelling (i.e. choice and optimization of DNN architecture as CNN, BLSTM, etc.) as well as on analysis of an impact of possible preprocessing of corrupted speech. The target system could be used for voice control of various devices or speech dictation in real environment of smart-home, running-car environment or for many other applications.

• Optimization of Language Modelling for the Recognition of Spontaneous and Informal Speech
  Language modelling is mandatory and key block of any continuous speech recognition system. Especially, when spontaneous speech should be recognized, currently used n-gram models reach their limits. Other modern methods of language modelling use classes, morphological information or techniques based on neural networks. Nowadyas LM based on deep neural networks start being used which may have potential especially for the language modelling of spontaneous speech. Optimalization of LM for therecognition of spontaneous or informal speech is the main general goal of this research. The second and more pratical goal is the usage of the on-line creation and updates of LM from the publicly available resources.

• Applications of Deep Neural Networks within Speech Enhancement Systems
  The research within this topic will be focused on possible applications and training of Deep Neural Networks (DNN) in the speech enhancement tasks. Neural networks are nowadays used mainly as a core of modern DNN-HMM-based speech recognition systems or for computation of aposteriory probabilities of chosen classes (e.g. phonemes, articulatory features, or information about voice activity detection. An analysis of possible usage of estimated aposteriory features within one-channel or multi-channel speech enhancement approaches will be the main purpose of the research in this field. Special attention will be paid to an enhancement of the speech distorted by high-level or strongly non-stationary background noise.

Important information about the rules and organization of PhD-study in English at CTU FEE can be found at official websites Doctoral Degree Studies and Essential Information about Doctoral Studies.

Realized and currently supervised PhD theses:

Ing. Michal Borský, PhD. - 2019 : Robust recognition of strongly distorted speech
Ing. Josef Rajnoha, PhD. - 2013 : Modelling of non-speech events in robust speech recognition (in Czech)
Ing. Jiří Tatarinov, PhD. - 2011 : Voice Activity Detection based on Hidden Markov Models (in Czech)
Ing. Hynek Bořil, PhD. - 2008 : Robust Speech Recognition: Analysis and Equalization of Lombard Effect in Czech Corpora
Ing. Petr Fousek, PhD. - 2007 : Extraction of Features for Automatic Recognition of Speech Based on Spectral Dynamics
Ing. Petr Opršal, PhD. - 2007 (TU Ostrava, 2nd supervisor specialist) : Methods of speech enhancement based on Wavelet transform (in Czech)
Ing. Josef Vopička, PhD. - 2004 (2nd supervisor specialist) : The task of keyword spotting in a continuous uttenrace (in Czech)
Ing. Jan Ingerle, PhD. - 2003 (2nd supervisor specialist) : Speech enhancement methods using a beamformer with post-fitering (in Czech)

Please, send me an email in the case of any problem with this page. E-mail: pollak with faculty address fel.cvut.cz Last change: Tue Jun 23 11:50:47 CEST 2015