Speech recognition system and a method of using dynamic bayesian network models

A computer-implemented system for speech recognition, said system comprising:

• an input device for registering an electrical signal representing speech,
• a converting module for converting said registered electrical signal representing speech to a converted signal in a frequency or time-frequency domain,
• an analysis module based on a dynamic Bayesian network, configured to analyze said converted signal and to generate hypotheses of words (W) and their probabilities on the basis of observed acoustic (OA) signal features or acoustic (OA) and visual (OV) signal features,
• a recognition module for recognition of text corresponding to said electrical signal representing speech on the basis of said generated hypotheses of words (W) and their probabilities,
• wherein the system further comprises:
• at least two signal parameterization modules for determining, during the recognition process, for said analysis module at least two observed acoustic (OA) signal features or acoustic (OA) and visual (OV) signal features in at least two parallel signal processing lines, wherein at least one of said signal processing lines is configured to receive said converted signal segmented in a different manner than the other signal processing lines such that at least some frames of at least one of said processing lines can be shared between different states of said Dynamic Bayesian Network,
• wherein said analysis module is configured to analyze dependencies between said observed signal features for at least two distinct time segments.

A computing device program product for speech recognition using a computing device, the computing device program product comprising:

• a non-transitory computer readable medium;
• first programmatic instructions for registering, by means of an input device, an electrical signal representing speech and converting said signal to frequency or time-frequency domain to obtain a converted signal,
• second programmatic instructions for analyzing said converted signal in an analysis module based on Dynamic Bayesian Network, configured to generate hypotheses of words (W) and their probabilities on the basis of observed acoustic (OA) signal features or acoustic (OA) and visual (OV) signal features,
• third programmatic instructions for recognizing, on the basis of certain word (W) hypotheses and their probabilities, a text corresponding to said electrical signal representing speech,
• fourth programmatic instructions for, during the recognition process, inputting to said analysis module said observed acoustic (OA) signal features or acoustic (OA) and visual (OV) signal features which are determined for said converted signal in frequency or time-frequency domain in at least two parallel signal processing lines, wherein at least one of said signal processing lines is configured to receive said converted signal segmented in a different manner than the other signal processing lines such that at least some frames of at least one line of said signal processing lines can be shared between different states of said Dynamic Bayesian Network,
• fifth programmatic instructions, for analyzing, in said analysis module, relations between said observed signal features for at least two distinct time segments.