An offline signal processing technique for accurate localisation of stop release bursts in vowel-consonant-vowel utterances

Accurate temporal localisation of release bursts in stop consonants is of utmost importance in many applications. Because of the transient nature of release bursts, burst onset localisation needs to have temporal accuracies of the order of ±5 ms. This is a challenging problem for conventional speech processing approaches using frame lengths and frame shifts of the order of multiples of 10 ms. In this paper, we have evaluated the performance of burst localisation of an HMM based speech recogniser (M1) with a method based on differentiation of Gaussian modelled spectrogram (M2) using 300 vowel-consonant-vowel utterances. Only 42% of the release bursts were detected within ±5 ms of the manual labels for M1 compared to 90% for M2. This result indicates the potential use of the method M2 for localising stop release bursts in offline applications requiring accurate temporal localisation of stop release bursts.