Robust Hierarchical Linear Model Comparison for End-of-Utterance Detection under Noisy Environments

A simple and efficient algorithm for robust end of-utterance detection of speech signal in noisy environments is proposed in the paper. To detect speech-block end-points, we use entropy sequence of the input speech signal, and hierarchically compare the fit of two weighted linear models. The first model, M1, is very simple, it corresponds to a constant average entropy level for the speech signal in the entire window. The second model, M2, corresponds to a step like entropy change from one constant level to another, with a gradual transition between the levels. Model M2 is in fact a piecewise linear regression model with two horizontal lines connected by a third transitional line. We treat M1 as a linear model only to be able to describe it as a sub model of M2 and use methodology based on statistical sub model testing. The regression models are constructed so that their fit will differ the most near the speech-block end-points. The models are fitted in a interval of the entropy sequence, that correspond to several consecutive frames in a sliding time window. The interval with the greatest difference of the model fit is used to estimate the location of the speech-block boundaries. The performance of the proposed algorithm is compared with a conventional approach by Nokia. The experimental results show that the proposed method approach can significantly outperform the conventional-based approach in different SNR swith different noise types. Furthermore, we observe that the proposed algorithm has the potential to improve robustness not only on diverse noise-type conditions, but also on the extremely noise contaminated speech below 0 dB SNR.