HMM Decoding in Noisy Enviroments using Uncertain Observations

J.A. Arrowood and M.A. Clements (USA)


Robust Speech Recognition, Uncertain Observations


This paper proposes a new technique for adapting Hidden Markov Model (HMM) speech recognition systems to ad ditive environmental noise based on an alternative imple mentation of the Parallel Model Combination (PMC) algo rithm [1]. While PMC is an effective technique for creat ing new acoustic models that match well to a noisy envi ronment not present in the training set, the computational requirements in both processing time and memory prevent its use in continuously adaptive systems, especially in en vironments where the background noise changes over time. This paper presents a novel approach by reformulating the model compensation problem to associate the uncertainty due to the background noise with the input signal, leaving the model unchanged. A model of the background environ ment is generated as in PMC, which is used to estimate pa rameters of a pdf describing the original clean signal, given a noisy observation frame. The HMM decoding algorithm is extended 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.

Important Links:

Go Back