Z. Rivlin, M. Cohen, V. Abrash, and C. T. A phonedependent confidence measure for utterance rejection. In Processing (ICASSP), 1996.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....the assumption of equal priors. One can try to learn explicitly the anti model for every word, but this is possible only in very limited applications [4, 9, 12] For a more general task, implicit alternate models can be derived from other models in the system or from the competing hypotheses [3, 4, 11]. It should be noted that some experimental studies (e.g. 3] have shown that an implicit model can outperform an explicit alternate model in performance and in computational cost. In this study we will focus on the confidence measures based on the use of implicit anti models. However, since ....

....However, since different letters may be variously modeled in the Handwriting Recognition engine, we investigated the use of letter level confidence measures, that may be combined to compute word level confidence measures. A comparable scheme has already been used in the speech recognition field [2, 11] and has shown interesting results. In the following, we first present how a word level confidence measure may be derived from letter level confidence measures (x3.1) Then, we define some anti models that we used to compute letter level confidence measures (x3.2) 3.1. Word Confidence Measure ....

Z. Rivlin, M. Cohen, V. Abrash, and C. T. A phonedependent confidence measure for utterance rejection. In Processing (ICASSP), 1996.

....1 Wessel et al. 207] compute an estimate of the unconditional probability of the acoustics over an interval P (X ne ns ) spanned by a decoding hypothesis by summing the likelihoods of all hypotheses spanning that interval contained in an n best word lattice. Keyword Spotting Rivlin et al. [165] were the first to use approach (2) for computing P (x n ; for the subsequent estimation of posterior phone probabilities on a per frame basis. These posterior probabilities were averaged in the log domain over the interval spanned by a phone hypothesis to form a phone level confidence ....

Z. Rivlin, M. Cohen, V. Abrash, and T. Chung. A phone-dependent confidence measure for utterance rejection. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing, pages 515--518. IEEE, 1996.

.... Y n P (q k n jx) P (q k n jW ) P (q k ) P (W ) 1) Hybrid HMM ANN systems like Abbot have been shown to produce useful estimates of the posterior phone probability given acoustic data [8] these posterior phone probabilities can then be used in calculating acoustic confidence measures [6], 11] Gethin Williams and Steve Renals at the University of Sheffield report using acoustic based confidence measures derived from the posterior phone probabilities of the Abbot HMM ANN speech recogniser to verify speech recognition hypotheses [9] 10] They define the duration normalised ....

Rivlin, Z, M. Cohen, V. Abrash, and T. Chung. "A Phone-Dependent Confidence Measure for Utterance Rejection." Proceedings of the IEEE, 1996.

....likelihood of the Viterbi path for the decoding hypothesis in question by the likelihood of the best Viterbi path for the period spanned by the decoding hypothesis. This confidence measure was not found to perform as well as one derived using an explicit alternate model, however. Rivlin et al. [49] describe using the posterior probability of a given context independent phone class as a phone level confidence measure, where the probability was computed using Bayes Theorem and the likelihoods and prior probabilities for all phone classes. Word level confidence estimates could then be found ....

Z. Rivlin, M. Cohen, V. Abrash, and T. Chung. A phone-dependent confidence measure for utterance rejection. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing, pages 515--518. IEEE, 1996.

....This may be motivated by ease of computation (simply subtracting the Viterbi scores at the start and the end of the word) It will be shown that the whole word approach gives much worse performance than hierarchical averaging for the corpora and recognition methods used in this thesis. Rivlin, Cohen, Abrash, and Chung (1996) show that normalizing by phone durations improves performance. They argue that to get the best recognition match, these [incorrect ] phones will have minimal duration in the Viterbi backtrace. Furthermore, since these recognized phones are incorrect, they [typically] have very poor ....

....which measures the number of times the proposed word occurs in a set of alternative hypotheses. Both of these approaches make explicit use of language models and is beyond the scope of this present research which is limited to acoustic based information only. 2.4. 3 Confidence Work at SRI Rivlin, Cohen, Abrash, and Chung (1996) shows that normalizing by phone durations improves performance. Weintraub, Beaufays, Rivlin, Konig, and Stolcke (1997) develops confidence metrics based on numerous features combined by an ANN. Some of these features are similar or identical in nature to those used in the hierarchical averaging ....

[Article contains additional citation context not shown here]

Rivlin, Z., M. Cohen, V. Abrash, and T. Chung (1996). A Phone-Dependent Confidence Measure for Utterance Rejection. In Proceedings of the 21st International Conference on Acoustics, Speech, and Signal Processing (ICASSP-96), Atlanta, Volume 1, pp. 515--517.

....research has been devoted to the development of confidence scores associated with the outputs of automatic speech recognition (ASR) systems. These scores were used mostly to help spot keywords in spontaneous or read texts, and to provide a basis for the rejection of out of vocabulary words (e.g. [4 11]) Many other ASR applications could also benefit from knowing the level of confidence in correct recognition. For example, text dependent speaker recognition systems could put more emphasis on words recognized with higher confidence; unsupervised adaptation algorithms could adapt the acoustic ....

....The acoustic features we implemented measure the normalized log likelihood (LL) of the acoustic realization of each word. They differ by the models used to normalize the LLs and by the way the frame level LLs are combined. Normalization was done either with context independent HMMs (CI HMMs) [4] or with a Gaussian mixture model (GMM) The frame level LLs were combined at the word level, phone level, or phone state level. This gives six possible features of which five were implemented. For example, the phone averaged, GMM normalized feature is defined as: 1 N N X j=1 i 1 Nf j N ....

Z. Rivlin, M. Cohen, V. Abrash, Th. Chung, "A PhoneDependent Confidence Measure for Utterance Rejection ", it ICASSP'96, vol. I, pp. 515-519.

Online articles have much greater impact   More about CiteSeer.IST   Add search form to your site   Submit documents   Feedback  

CiteSeer.IST - Copyright Penn State and NEC