Recently, we have proposed a detection-based speech recog-nizer which has two main components: a bank of phoneticfeature detectors implemented with hidden Markov models(HMMs), and an event merger. Each detector generates a scorethat pertains to some phonetic features, e.g. voicing. Themerger combines all these scores to generate phone labels. Theparameters of the detectors and the merger can be optimized ei-ther separately or jointly, and we showed that penalized logisticregression machine (PLRM) is a convenient tool for joint opti-mization. We validated our approach on a rescoring scheme. Inthis work, we tackle the phone classification problem and showthat high level phone accuracy can be achieved without a directmodeling of the phones when PLRM is used. We also showthat better results can be obtained by increasing the numberof phonetic features, and that our method outperforms phoneclassifiers trained either by maximum likelihood estimation, ormaximum mutual information.
展开▼
