7th International Conference on Spoken Language Processing

September 16-20, 2002
Denver, Colorado, USA

Using Observation Uncertainty in HMM Decoding

Jon A. Arrowood, Mark A. Clements

Georgia Institute of Technology, USA

This paper proposes a new technique for adapting Hidden Markov Model (HMM) speech recognition systems to additive environmental noise by incorporating information about the uncertainty of observations. Current techniques, such as the Parallel Model Combination (PMC) algorithm [1], are successful in steady state noise environments. However, the computational requirements both in processing time and memory prevent its widespread use in continuously adaptive systems, necessary for environments with changing background noise. This paper presents an approach that reformulates the model combination technique to update the each observation instead of the model. As in PMC, a model of the background is generated during nonspeech times. This is used for each input frame to generate a pdf describing the original clean signal, given the noise model. The HMM decoding algorithm is extended as in [2, 3] to allow pdf inputs, and recognition results are presented that show this technique compares favorably with PMC in unchanging noise environments, but has significant benefits in changing noise.


Full Paper

Bibliographic reference.  Arrowood, Jon A. / Clements, Mark A. (2002): "Using observation uncertainty in HMM decoding", In ICSLP-2002, 1561-1564.