LAUSR

Automatic detection of vowels plays a significant role in the analysis and synthesis of speech signal. Detecting vowels within a speech utterance in noisy environment and varied contexts is a very challenging task. In this work, a robust technique based on non-local means (NLM) estimation is proposed for the detection of vowels in noisy speech signals. In the NLM algorithm, the signal value at each sample point is estimated as the weighted sum of signal values at other sample points within a search neighborhood. The weight value is computed by finding square of the difference between the signal values belonging to two different segments. During the estimation, one segment is kept as fixed, while other segment is slid over the search neighborhood. For any particular sample point, the sum of those weight values is significantly less when the segments under consideration are higher in magnitude. In a given speech signal, the vowels are regions of high energy. This will be true even under noisy conditions. In this work, the sum of weight values (SWV), computed at each time instant is used as a discriminating feature for detecting the vowels in a given speech signal. In the proposed approach, the regions where the SWV exhibits significant transitions and attain lower values for a considerable duration of time compared to the preceding and succeeding regions are hypothesized as the vowels. This hypothesis is statistically validated for detecting vowels under clean as well as noisy test conditions. For proper comparison, a three-class statistical classifier (vowel, non-vowel and silence) is developed for detecting the vowels in a given speech signal. For developing the said classifier, the mel-frequency cepstral coefficients are used as the acoustic feature vectors, while deep neural network (DNN)-hidden Markov model (HMM) is employed for acoustic modeling. The proposed vowel detection method is observed to outperform the DNN-HMM-based statistical classifiers as well as existing signal processing approaches under both clean and noisy test conditions.