Chapter 3 Uncertainty Propagation Ramón

While it is often fairly straightforward to estimate the reliability of speech features in the time-frequency domain, this may not be true in other domains more amenable to speech recognition, such as e.g. for RASTA-PLP features or those obtained with the ETSI advanced front end. In such cases, one useful approach is to estimate the uncertainties in the domain where noise-reduction preprocessing is carried out, and to subsequently transform the uncertainties, along with the actual features, to the recognition domain. In order to develop suitable approaches, we will first give a short overview of relevant strategies for propagating probability distributions through nonlinearities. Secondly, for some feature domains suitable for robust recognition, we will show possible implementations and sensible approximations of uncertainty propagation and discuss the associated error margins and trade-offs. 3.1 Uncertainty Propagation In automatic speech recognition (ASR) an incoming speech signal is transformed into a set of observation vectors x = x1 · · ·xL which are input to the recognition model. ASR systems which work with uncertain input data replace this observation set by a distribution of possible observed values conditioned on the available information I, p(x1 · · ·xL|I). This uncertainty represents the lost information that causes the mismatch between the observed speech and the trained ASR model. By combining this distribution with observation uncertainty techniques [23], superior recognition robustness can be attained. There are multiple sources from which this Ramón Fernandez Astudillo Electronics and Medical Signal Processing Group, TU Berlin, 10587 Berlin, e-mail: ramon@ astudillo.com Dorothea Kolossa Electronics and Medical Signal Processing Group, TU Berlin, 10587 Berlin, e-mail: dorothea. [email protected]