Noise-compensated hidden Markov models

The technique of hidden Markov models has been established as one of the most successful methods applied to the problem of speech recognition. However, its performance is considerably degraded when the speech signal is contaminated by noise. This work presents a technique which improves the performance of hidden Markov models when these models are used in different noise conditions during the speech recognition process. The input speech signal enters unchanged to the recognition process, while the models used by the recognition system are compensated according to the affecting noise characteristics, power and spectral shape. Hence, the compensation stage is independent of the recognition stage, allowing the models to be continually adjusted. The models used in this work are from a continuous density hidden Markov algorithm, having cepstral coefficients derived from linear predictive analysis as state parameters. It is used only static features in the models in order to show that, when properly compensated for the noise, these static features contribute significantly to improve noisy speech recognition. It is observed from the results that the parameters kept their capability to discriminate among different classes of signals, indicating that, in the context of speech recognition, the use of autoregressive-derived parameters with noisy signals does not represent an impediment. A matrix-way of converting from autoregressive coefficients to normalized autocorrelation coefficients is presented. The affecting noise is assumed additive and statistically independent of the speech signal. Although the noise dealt with should also be stationary, good performance was achieved for nonstationary noise, such as operations room noise and factory environment noise. The concept of intra-word signal-to-noise ratio is presented and successfully applied. The resulting compensated models revealed to be less dependent on the training data set when compared to the trained hidden Markov models. Due to the computational simplicity, the time required to adjust a model is significantly shorter than the time to train it.