Unsupervised Acoustic Unit Discovery by Leveraging a Language-Independent Subword Discriminative Feature Representation

This paper tackles automatically discovering phone-like acoustic units (AUD)\nfrom unlabeled speech data. Past studies usually proposed single-step\napproaches. We propose a two-stage approach: the first stage learns a\nsubword-discriminative feature representation and the second stage applies\nclustering to the learned representation and obtains phone-like clusters as the\ndiscovered acoustic units. In the first stage, a recently proposed method in\nthe task of unsupervised subword modeling is improved by replacing a\nmonolingual out-of-domain (OOD) ASR system with a multilingual one to create a\nsubword-discriminative representation that is more language-independent. In the\nsecond stage, segment-level k-means is adopted, and two methods to represent\nthe variable-length speech segments as fixed-dimension feature vectors are\ncompared. Experiments on a very low-resource Mboshi language corpus show that\nour approach outperforms state-of-the-art AUD in both normalized mutual\ninformation (NMI) and F-score. The multilingual ASR improved upon the\nmonolingual ASR in providing OOD phone labels and in estimating the phone\nboundaries. A comparison of our systems with and without knowing the\nground-truth phone boundaries showed a 16% NMI performance gap, suggesting that\nthe current approach can significantly benefit from improved phone boundary\nestimation.\n