K. Torkolla (1993). An efficient way to learn english grapheme-to-phoneme rules automatically, ICASSP93 Proceedings. ERRATA Sabine DELIGNE, Francois YVON and Fr'ed'eric BIMBOT vol. 3, pp. 2243-2246 It turned out that some of the results that we report are slightly over-estimated. This only affects per phoneme accuracies.  Home/Search   Document Details and Download   Summary   Related Articles   Check

....that our approach is fully unsupervised, our first results are encouraging. However they seem to reach a ceiling around 95 phoneme accuracy, which is quite below the scores obtained using alternative self learning algorithms: for example, experiments which we carried out on the same data with DEC [8] outperform our per phoneme results by about 3 points. The careful examination of a sample of joint segmentations (see Table 4) produced during the learning stage can improve our understanding of the model behavior. Graphemic Phonemic (verb) al) vRb) al) heurt) er) oeRt) e) l egal) it e) ....

K. Torkolla (1993). An efficient way to learn english grapheme-to-phoneme rules automatically, ICASSP93 Proceedings. ERRATA Sabine DELIGNE, Francois YVON and Fr'ed'eric BIMBOT vol. 3, pp. 2243-2246 It turned out that some of the results that we report are slightly over-estimated. This only affects per phoneme accuracies.

....in a transcription is computed on the basis of the string to string edit distance with the target pronunciation. Table 2 reports only the best results obtained in our complete evaluation of each model, including an additionnal reference point obtained using a decision tree learning technique, DEC [6]. The main outcome of this evaluation is that the Model words phonemes (5; 5) Joint Multigram 64.5 92.3 (3; 2) Joint Multigram 71.3 93.7 Overlapping Model 86.3 95.4 Decision Tree (DEC) 86.7 97.9 Table 2. Comparative evaluation of the joint multigram model and of the overlapping model on a ....

Kari Torkolla. An efficient way to learn English grapheme-to-phoneme rules automatically. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), volume 2, Minneapolis, Apr. 1993.

....group. As can be seen, on these lexicons 4 , SMPA systematically outperforms PRONOUNCE, all the differences being statistically significant at a 0.01 confidence level in a two tailed test of mean comparison on paired samples. Additional comparison sources are provided by the evaluations of DEC [18], a decision tree learning algorithm. We have also reimplemented this algorithm, and performed an evaluation on the same data bases. The decision tree approach enables to achieve accuracy scores that are in between those of PRONOUNCE and SMPA: very close to PRONOUNCE s performance on nettalk and ....

Kari Torkolla. An efficient way to learn English grapheme-to-phoneme rules automatically. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), volume 2, pages 199--202, Minneapolis, apr. 1993.

....tree corresponds to objects belonging to the same class, or until no more attribute is available. Several successful attempts to use ID3 for graphemeto phoneme conversion were reported for example in [27, 15, 14] Simplified version of ID3 were also tested, with again very satisfying results (see [43], or [45] where a multilingual experiment is carried out) These simplified versions differ from ID3 mostly in that they fix a priori the list of clustering features, and their hierarchy. When applied in the conversion task, where objects (graphemes) are described in terms of their surrounding ....

....SELEGRAPH, another IBL software developed within the framework of ONOMASTICA by Ove Andersen and Paul Dalsgaard [2, 3] As we shall see, the main difference between SELEGRAPH and DEC is that the latter allows varying context size. 3 DEC yet another IBL algorithm DEC (Dynamic Expanding Context) [24, 42, 43] is yet another instance based learning algorithm. The main difference with other such algorithms (e.g SELEGRAPH) is that is does not fix a priori the context size, and rather let the algorithm decide, for each grapheme, the appropriate length of context needed to achieve a correct prediction. To ....

Kari Torkolla. An efficient way to learn english grapheme-to-phoneme rules automatically. In Proceedings of the International Conference on Acoustics, Speech and Signal Processing (ICASSP), volume 2, pages 199-- 202, Minneapolis, apr. 1993.

No context found.

Torkkola, Kari. 1993. An efficient way to learn English grapheme-to-phoneme rules automatically. IEEE.

No context found.

Kari Torkkola. 1993. An efficient way to learn English grapheme-to-phoneme rules automatically.

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