Unsupervised Discovery of Recurring Spoken Terms Using Diagonal Patterns

P. Sudhakar, K. Sreenivasa Rao, Pabitra Mitra

Abstract: Spoken term discovery is a challenging task when a lot of spoken content is generated without annotation. The spoken term discovery task accomplished by pattern matching techniques resolves the challenge by directly capturing the resemblance of the spoken terms at the acoustic feature level. Despite feasibility, the pattern-matching approach generates more false alarms during the discovery task due to fluctuations that arise in natural speech; hence degradation in the performance was observed. In the proposed approach, the challenge that arises due to the variability is addressed in two stages. In the first stage, the RASTA-PLP spectrogram was used as an acoustic feature representation that reduces the variabilities among similar spoken contents. In the second stage, the novel Diagonal Pattern Search method unconstrainedly computes the pattern resemblance between the identical spoken terms at the segmental level. The proposed approach was evaluated using the IITKGP-SDUC speech corpus and inferred that a 10.11% improvement in the accuracy was achieved compared to other state-of-the-art systems in the spoken term discovery task.