Duration and spectral based stress token generation for HMM speech recognition under stress

In this paper, we address the problem of isolated word recognition of speech under various stressed speaking conditions. The main objective is to formulate an alternate training algorithm for hidden Markov model recognition, which better characterizes actual speech production under stressed speaking styles such as slow, loud and Lombard (1911) effect. Without the need for collecting such stressed speech data. The novel approach is to first construct a previously suggested source generator model of word production employing knowledge of the statistical nature of duration and spectral variation of speech under stress. This is used in turn to produce simulated stressed speech training tokens from neutral tokens, and thus replace neutral data used in the recognizer training phase. The token generation training method is shown to improve isolated word recognition by 8% for slow speaking style, 14% for loud speaking style, and 24% for speech under Lombard effect when compared to neutral trained isolated word recognition