C. Wooters and A. Stolcke. Multiple-pronunciation lexical modeling in a speaker indepen- dent speech understanding system. In Proceedings of the 3rd Int'l Conference on Spoken Language Processing (ICSLP-94), 1994.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....by unsupervised adaptation of the acoustic models, but rather that the achieved WER reductions are additive. 1. TRODUCTION In recent years, numerous attempts have been made to introduce multiple pronunciations or graph based probabilistic pronunciation rules into HMM based speech recognizers [1, 3, 4, 11, 13, 9]. This work has been motivated by the observation that there is a strong mismatch between canonical dictionary pronunciations and the actually observed pronunciations, in particular for spontaneous speech [1] Unfortunately, incorporating additional pronunciation variants or general pronunciation ....

C. Wooters, A. Stolcke, "Multiple-Pronunciation Lexical Modeling in a Speaker Independent Speech Understanding System", ICSLP'94, pp. 1363-1366.

....Other generic modeling approaches, not incorporating a tree predictor are also found in the literature. Paliwal [50] describes a bigram model for P ( Phi j W ) In the work of Cohen [16] probabilities of handwritten phonological rules are learned from training data. In the work of Wooters et al. [71], the probabilities of a predetermined set of pronunciations is derived from the training data by means of a forced alignment. The probability augmented pronunciations of a word are then merged so as to maximize the posterior probability of the word model. The third class of models use criterion ....

C. Wooters and A. Stolcke, "Multiple-pronunciation lexical modeling in a speaker independent speech understanding system," In Proceedings of the International Conference on Spoken Language Processing, pp. 1363-1366, 1994. 137

....and potential baseforms is to augment the existing lexicon with all potential baseforms and to perform a forced alignment of the reference transcription to the training set using the expanded lexicon. The baseforms most frequently selected by the alignment process are favoured in this case [212, 196, 162, 208]. A useful feature of this evaluation method is that a prior probability estimate can be easily computed for a baseform given the relative frequencies of occurrence of a word s candidate baseforms in the forced alignment. The assignment of prior probabilities to baseforms in a lexicon has been ....

C. Wooters and A. Stolke. Multiple-pronunciation lexical modeling in a speaker independent speech understanding system. In Proceedings of the International Conference on Spoken Language Processing, pages 1363--1366, 1994.

....the identity of a phonetic segment when a pronunciation change occurs, this Pronunciation modeling by sharing Gaussians 143 technique is able to model partial changes. The method presented in this paper shares this characteristic of modeling partial pronunciation changes. A technique proposed by Wooters and Stolcke (1994) for generating a phone graph for a word from empirically observed pronunciations has been extended by Eide (1999) to generating a HMM state graph for a context dependent phoneme. Wakita, Singer and Sagisaka (1999) also use empirically observed HMM state sequences to infer alternate word ....

Wooters, C. & Stolcke, A. (1994). Multiple pronunciation lexical modeling in a speaker independent speech understanding system. Proceedings of the International Conference on Spoken Language Processing (ICSLP), pp. 1363--1366.

....in future work to address a number of shortcomings of these experiments, for example including some spontaneous speech corpora, and looking at a wider variety of rules. In addition, we have extended our algorithm to induce new pronunciations which generalize over pronunciations seen in the corpus (Wooters Stolcke 1994). We now plan to augment our probability estimation to use the pronunciations from this new HMM induction based generalization step. This will require extending our tag based probability estimation step to parse the phone strings from the forcedViterbi. In other current work we have also been ....

Wooters, Chuck, & Andreas Stolcke. 1994. Multiple-pronunciation lexical modeling in a speaker-independent speech understanding system. In ICSLP-94 .

....and Mercer, 1984; Bahl et al. 1981#, or by generalizing phone recognizer output to broad phonetic categories #such as stops and fricatives # #Schmid et al. 1987#. One can also generalize over the set of pronunciations learned by these techniques, using techniques such as HMM Generalization #Wooters and Stolcke, 1994#, which allows induction of new baseforms not seen in the training data by #nding common variations among pronunciations seen during training. Eide #1999# used a similar technique in their Broadcast News speech recognizer, #nding improvements in word error rate for spontaneous and fast speech. ....

C. Wooters and A. Stolcke. Multiple-pronunciation lexical modeling in a speaker independentspeech understanding system. In Proceedings of the 3rd Int'l ConferenceonSpoken Language Processing #ICSLP-94#, 1994.

.... and Mercer, 1984; Bahl et al. 1981] or by generalizing phone recognizer output to broad phonetic categories (such as stops and fricatives) Schmid et al. 1987] One can also generalize over the set of pronunciations learned by these techniques, using techniques such as HMM Generalization [Wooters and Stolcke, 1994], which allows induction of new baseforms not seen in the training data by finding common variations among pronunciations seen during training. Eide [1999] used a similar technique in their Broadcast News speech recognizer, finding improvements in word error rate for spontaneous and fast speech. ....

C. Wooters and A. Stolcke. Multiple-pronunciation lexical modeling in a speaker independent speech understanding system. In Proceedings of the 3rd Int'l Conference on Spoken Language Processing (ICSLP-94), 1994.

.... a decision tree to generate alternate pronunciations [4, 7, 14] Use a baseform representation and apply a sequence of pronunciation rules to generate the possible pronunciation alternatives [16, 17, 18, 26, 23, 24] Merge multiple linear pronunciations together to form a pronunciation graph [20 22]. One should note that listing the pronunciations for a word may be adequate for isolated word recognition but this technique cannot account for any dependence of the pronunciation realization on the neighboring words. The technique of applying rules or decision trees to a pronunciation baseform ....

C. Wooters, and A. Stolcke, "Multiple-Pronunciation Lexical Modeling in a Speaker Independent Speech Understanding System," 1994 ICSLP.

....final real test for the phonetic set hashing scheme would be to check performance on real data. Conventional HMM based recognition systems use a single transcription per word as provided by the lexicon. Some researchers have recently explored data driven multiple pronunciation lexical modeling [1]. How useful is the single transcription of each word for recognition using hashing schemes In order to examine this, the lexicon transcriptions of the TIMIT dr1 test and train sentences were hashed in as before. However, the real test set now consisted of the phonetician labeled test set data ....

Chuck Wooters, and Andreas Stolcke, "Multiple-Pronunciation Lexical Modeling in a Speaker Independent Speech System Understanding System", to appear in Proc. of ICSLP-94.

....labor intensive to be done by hand. Problems are compounded by the need to include dialectical and prosodic variants. These issues have spawned several responses, including the use of rules that model phonetic variability [1] or hidden Markov models that learn the patterns of phonetic variability [2,3]. These additional models add considerable complexity to the decoding process. They also fail to address the practical problem of how to find bad phonetic transcriptions in existing large dictionaries. We propose and test a practical means of finding poor pronunciations and discovering missing ....

Wooters, C. and Stolcke, A., Multiple-pronunciation lexical modeling in a speaker independent speech understanding system. ICSLP-1994, pp. 1363-1366.

No context found.

WOOTERS,CHUCK,&ANDREAS STOLCKE. 1994. Multiple-pronunciation lexical modeling in a speakerindependent speech understanding system. In Proceedings International Conferenceon SpokenLanguage Processing, Yokohama.

....an e#cient solution to this problem, which can be applied in a varietyof domains and which has some degree of cognitive plausibility. The model merging approach has since been applied successfully to Bayesian inference in the domain of #nite state probabilistic grammars #Stolcke Omohundro 1993; Wooters Stolcke 1994#, and the current approach for L 0 is a direct extension of that work. 1 In this framework, informative priors, or linguistic theories, can be seen as very specialized codes that provide short descriptions for all theory conforming grammars, leaving the long descriptions for the nonconforming ....

Wooters, Chuck,&Andreas Stolcke. 1994. Multiple-pronunciation lexical modeling in a speaker-independent speech understanding system. In Proceedings International ConferenceonSpoken Language Processing,volume 3, 1363#1366, Yokohama.

....efficient solution to this problem, which can be applied in a variety of domains and which has some degree of cognitive plausibility. The model merging approach has since been applied successfully to Bayesian inference in the domain of finite state probabilistic grammars (Stolcke Omohundro 1993; Wooters Stolcke 1994), and the current approach for L 0 is a direct extension of that work. 1 In this framework, informative priors, or linguistic theories, can be seen as very specialized codes that provide short descriptions for all theory conforming grammars, leaving the long descriptions for the nonconforming ....

Wooters, Chuck, & Andreas Stolcke. 1994. Multiple-pronunciation lexical modeling in a speaker-independent speech understanding system. In Proceedings International Conference on Spoken Language Processing , volume 3, 1363--1366, Yokohama.

....use of about twice as many transitions as the more compact merged HMMs, which would have a serious impact on potential applications of such models in speech recognition. Finally, the HMM merging algorithm was integrated into the training of a medium scale spoken language understanding system (Wooters Stolcke 1994). Here, the algorithm also serves the purpose of inducing multi pronunciation word models from speech data, but it is now coupled with a separate process that estimates the acoustic emission likelihoods for the HMM states. The goal of this setup was to improve the system s performance over a ....

WOOTERS, CHUCK, & ANDREAS STOLCKE. 1994. Multiple-pronunciation lexical modeling in a speaker-independent speech understanding system. In Proceedings International Conference on Spoken Language Processing, Yokohama.

....exactly the set of input pronunciations, constructed by stringing together each observed phone sequence between a start and end state. Next, this initial model is simplified and generalized by repeatedly merging states until we reach a model with (locally) maximum posterior probability. See [14] for more details. Our second method augments this bottom up approach with top down information. Following [5] we manually develop phonological rules and automatically apply them to a set of baseform or dictionary pronunciations, to automatically generate multiple pronunciations for any word in ....

Chuck Wooters and Andreas Stolcke. Multiple-pronunciation lexical modeling in a speaker-independent speech understanding system. In ICSLP-94, 1994. To appear.

No context found.

C. Wooters and A. Stolcke. Multiple-pronunciation lexical modeling in a speaker indepen- dent speech understanding system. In Proceedings of the 3rd Int'l Conference on Spoken Language Processing (ICSLP-94), 1994.

No context found.

C. Wooters and A. Stolcke. Multiple-pronunciation lexical modeling in a speaker indepen- dent speech understanding system. In Proceedings of the 3rd Int'l Conference on Spoken Language Processing (ICSLP-94), 1994.

No context found.

C. Wooters, A. Stolcke, "Multiple-Pronunciation Lexical Modeling in a Speaker Independent Speech Understanding System", ICSLP'94, pp. 1363--1366.

No context found.

C. Wooters and A. Stolcke : "Multiple-pronunciation lexical modeling in a speaker independent speech understanding system," Proc. ICSLP-94, pp. 1363-- 1366, 1994.

No context found.

Chuck Wooters, and Andreas Stolcke, "Multiple-Pronunciation Lexical Modeling in a Speaker Independent Speech System Understanding System", to appear in Proc. of ICSLP-94.

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