Unsupervised query-by-example spoken term detection based on DPHMM tokenizer
Cao Jiankai, Zhang Lianhai · 2017
This paper investigates the use of Dirichlet process hidden Markov model (DPHMM) tokenizer for the template matching based query-by-example spoken term detection (QbE-STD) task. DPHMM can be obtained following an unsupervised iterative procedure without any training transcriptions. The STD performance of the DPHMM tokenizer is evaluated on TIMIT Corpus. We construct three kinds of DPHMM based QbE-STD systems. Compared to GMM baseline system, posteriorgram based system has a better detection precision, acoustic unit has a better detection speed, and pseudo feedback based system has a good balance between precision and speed. Although nonparametric DPHMM is still worse than supervised language-matched phoneme recognizer, but it outperforms other unsupervised tokenizers and language-mismatched phoneme recognizer.