Young, S. J., Adda-Decker, M., Aubert, X., Dugast, C., Gauvain, J.L., Kershaw, D. J., Lamel, L., Leeuwen, D. A., Pye, D., Robinson, A. J., Steeneken, H. J. M., and Woodland, P. C., (1997). Multilingual Large Vocabulary Speech Recognition: The European SQALE Project. Computer, Speech, and Language 11, 73-89.  Home/Search   Document Not in Database   Summary   Related Articles   Check

....triphone models (about 9000 tied states distributed over 5000 models) In the third decoding pass unsupervised acoustic model adaptation based on MLLR [10] is carried out using the hypotheses generated in the second pass. DESIGN CONSIDERATIONS Previous work on LVCSR in French recognition [6, 8, 13] has highlighted the importance of increasing lexical coverage as an important issue in recognizer development. The link between coverage and language modeling is investigated more deeply here. Lexical Coverage The problem of lexical coverage has been addressed along different axes: word list ....

S.J. Young et al., "Multilingual large vocabulary speech recognition: the European SQALE project," Computer Speech & Language, 11(1), pp. 73-89, Jan. 1997.

....of lexical variety in a language. Dutch is characterized by processes of inflection, derivation and mainly compounding. Moreover, compounds form orthographic units like in German. The influence of lexical variety on lexical coverage for different languages is summarized in table 1, adapted from [4]. Lexical coverage was measured on newspaper text, respectively on the Wall Street Journal (37.2M words) for English, Le Monde (37.7M words) for French and the Frankfurter Rundschau (31.5M words) for German. We added results for No. words English French German Dutch Table 1. Lexical ....

S.J. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.-L. Gauvain, D.J. Kershaw, L. Lamel, D.A. Leeuwen, D. Pye, A.J. Robinson, H.J.M. Steeneken, and P.C. Woodland, "Multilingual large vocabulary speech recognition: the European SQUALE project," Computer Speech and Language, vol. 11, no. 1, pp. 73--89, 1997.

....language. A lexicon for German language speech recognition needs to contain 3 4 times as many forms as that for English language speech recognition in order to attain the same coverage. A French language lexicon achieves the same coverage as the English language lexicon with only a doubled size [27]. Sub word language model units alleviate the problem of the bloated lexicon [13] 20] But as the base units of the language model become fewer and smaller, the language model becomes less constrained. Base units must be optimized to achieve maximum coverage and minimum ill effects of decreased ....

S.J. Young, M. Adda-Dekker, X. Aubert, C. Dugast, J.-L. Gauvin, D.J. Kershaw, L. Lamel, D.A. Leeuwen, D. Pye, A.J. Robinson, H.J.M. Steeneken, P.C. Woodland, Multilingual large vocabulary speech recognition the European SQUALE project, Computer Speech and Language 11, 1997.

.... on speaker independent isolated word databases are around 1 for 100 words vocabulary, 3 for 600 words and 10 for 8000 words [DER98] For a speaker independent continuous speech recognition database, the error rates are around 15 with a trigram language model and for a 65000 words vocabulary [YOU97]. 2.2 The Speech Recognition Process The Speech Recognition process can be divided in many different components illustrated in figure 4. Speech Feature Extraction Probability Estimation Decoding Language Models Recognized Sentences Fig. 4 The speech recognition process. Note that the ....

Young S., Adda-Dekker M., Aubert X, Dugast C., Gauvain J.L., Kershaw D. J., Lamel L., Leeuwen D. A., Pye D., Robinson A. J., Steeneken H. J. M., Woodland P. C., Multilingual large vocabulary speech recognition : the European SQALE project, Computer Speech and Language, Vol 11, pp 73-89, 1997

.... much of recent speech recognition research for American English has been supported by ARPA and has been based on text materials which were processed to remove case distinction and compound words [11] Case is generally kept as a distinctive feature in French and more importantly in German [8, 14]. The large amounts of training texts required for lexical and statistical language model design need to be cleaned and normalized before use. We compare different types of normalization of a source text containing 185 million words of the French newspaper Le Monde. The lexical coverages and ....

S.J. Young et al., "Multilingual large vocabulary speech recognition: the European SQALE project," Computer Speech & Language, 11(1), pp. 73-89, Jan. 1997. Eurospeech'97, Adda-Adda-Decker-Gauvain-Lamel 59

....project. 1. INTRODUCTION For languages like English many large vocabulary continuous speech recognition engines have been evaluated on several different tasks. Recently the interest increased on LVCSR systems in Asian languages like Chinese, Japanese and Korean. Furthermore, projects like SQALE [1] focus on transferring the evaluation paradigms and training methods to languages spoken in Europe like French, German, Spanish etc. However, so far there have been no attemps for Turkish Large Vocabulary Continuous Speech Recognition (LVCSR) This has several reasons: First, there is a lack of ....

S.J. Young et al.: Multilingual large vocabulary speech recognition: the European SQALE project, Computer Speech and Language, vol 11, pp. 73-89, 1997.

....to handling so called compounding languages like German, Dutch, Norwegian, Danish, Swedish and Greek, in which complex concepts are expressed as single words. In order to achieve the same OOV rates as non compounding languages, compounding languages must use a much larger recognition lexicon [1][2] Even if its size has been stepped up considerably, an orthographic word based speech recognition lexicon lacks the flexibility to cover new compounds or compounds coined by the speaker on the fly. The generally accepted remedy is that for compounding languages, orthographic words must be ....

....word based speech recognition lexicon lacks the flexibility to cover new compounds or compounds coined by the speaker on the fly. The generally accepted remedy is that for compounding languages, orthographic words must be split into linguistically meaningful sub units in the recognition lexicon [1][2] 3] Splitting lexicon entries, however, robs them of their inherent contextual content and exacerbates acoustic confusability. In fact, many authors have taken exactly the opposite tactic and have reported WER improvement gained from combining orthographic words into phrases in the recognition ....

Young, S.M., Adda-Decker, M., Aubert, X., Dugast, C., Gauvain, J-L., Kershaw, D.J., Lamel, L., Leeuwen, D.A., Pye, D., Robinson, A.J., Steeneken, H.J.M. and Woodland, P.C. "Multilingual large vocabulary speech recognition: the European SQALE project", Computer Speech and Language, Vol. 11, pp. 73-89, 1998.

....languages like homophones or compound words and other characteristics affecting the recognition process. For these characteristics, algorithms have to be found that can cope with these new problems. The system is then trained with data of the new language. This approach can be found for example in [2, 3, 11]. Another approach follows the same application goal as the approach above with the only difference, that there is not sufficient training material available in the new language. Thus, for cross lingual recognition methods must be found to use training material of another language for a rough ....

S. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.-L. Gauvain, D. Kershaw, L. Lamel, D. Leeuwen, D. Pye, A. Robinson, H. Steeneken, and P. Woodland. Multilingual large vocabulary speech recognition: the European SQUALE project. Computer Speech & Language, 11:73--89, 1997.

....mostly from the area of South Tyrol and show a big variety of dialects especially in the German language. Further on we call the variation of a non native speaker accent and the one of a native speaker dialect. There are different approaches to multi and bilingual speech recognition, for example [3, 4, 11, 12]. One approach consists of the development of a system that recognizes one language at a time, but with language independent algorithms that cover language specific aspects like homophones or coarticulation effects. A second approach is the portation of a recognizer to another language with as ....

S. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.-L. Gauvain, D. Kershaw, L. Lamel, D. Leeuwen, D. Pye, A. Robinson, H. Steeneken, and P. Woodland. Multilingual large vocabulary speech recognition: the European SQUALE project. Computer Speech & Language, 11:73--89, 1997.

....probabilities used by the HMMs. The hybrid HMM ANN system has already been successfully applied in American English and British English. In this paper, this system is tested on a French continuous speech database : BREF 80. Some good results on this particular database have already been published [10] [5] using continuous density HMMs and context dependent models. We will compare those results with our baseline context independent hybrid system on this particular task. 2. BREF 80 BREF 80 [7] is a large read speech corpus from 80 speakers. The text material was selected from the French ....

Young S.J. and al, "Multilingual large vocabulary speech recognition : the European SQALE project", Computer Speech and Language 0-73-890, 1997.

.... and ELRA (European Language Resources Association) which is the European equivalent to LDC (Linguistic Data Consortium) The LRE Sqale project aimed to adapt the ARPA LVCSR (Large Vocabulary Continuous Speech Recognition) evaluation paradigm (Pallett et al. 1995) to a multilingual context (Young et al. 1997). The project EAGLES (Expert Advisory Group on Language Engineering Standards) has created a manual for the development of linguistic resources and the evaluation of spoken language systems (Gibbon et al. 1997) Recently, AUPELF UREF (Association des Universites Partiellement ou Entierement de ....

S. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.L. Gauvain, D.J. Kershaw, L. Lamel, D.A. Leeuwen, D. Pye, A.J. Robinson, H.J.M. Steeneken, and P.C. Woodland (1997), "Multilingual large vocabulary speech recognition: the European SQALE project," Computer Speech and Language, 11:73-89.

....l Agriculture) 2 known as hybrid HMM ANN method. HMM ANN system has already been successfully applied in American English and English. In this paper, this system is tested on a continuous speech French database: BREF 80. Some good results on this particular database have already been published [1] using continuous density HMMs and context dependent models. We will compare those results with our hybrid system on this particular task. II. BREF 80 BREF 80 [3] is a large speech corpus read by 80 speakers. The text material was selected from the French newspaper Le Monde so as to provide a ....

Young, S. J. and al, Multilingual large vocabulary speech recognition : the European SQALE project, Computer Speech and Language 0-73-89-0, 1997.

No context found.

S. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.L. Gauvain, D.J. Kershaw, L. Lamel, D.A. van Leeuwen, D. Pye, A.J. Robinson, H.J.M. Steeneken, P.C. Woodland, "Multilingual Large Vocabulary Speech Recognition: The European SQALE Project," Computer Speech and Language, 11(1): 73-89, Jan. 1997.

.... Audio Segmentierung, akustische Modellierung, Spachmodellierung 1 Introduction Over the last 5 years significant advances have been made in large vocabulary, continuous speech recognition, which has been a focal area of research, serving as a test bed to evaluate models and algorithms [5, 6, 45]. However, these tasks remain relatively artificial as they mainly make use of laboratory read speech data. In this paper we report on moving toward real world speech data in order to build a system for transcribing radio and television broadcast news [6, 7, 8, 9] While this paper focuses on our ....

S.J. Young, M. Adda-Decker, X. Aubert, C. Dugast, J.L. Gauvain, D.J. Kershaw, L. Lamel, D.A. Leeuwen, D. Pye, A.J. Robinson, H.J.M. Steeneken, P.C. Woodland, "Multilingual large vocabulary speech recognition: the European SQALE," Proc. Computer, Speech and Language, 11(1), pp. 73-89, Jan. 1997.

No context found.

Young, S. J., Adda-Decker, M., Aubert, X., Dugast, C., Gauvain, J.L., Kershaw, D. J., Lamel, L., Leeuwen, D. A., Pye, D., Robinson, A. J., Steeneken, H. J. M., and Woodland, P. C., (1997). Multilingual Large Vocabulary Speech Recognition: The European SQALE Project. Computer, Speech, and Language 11, 73-89.

No context found.

S.J. Young et al.: Multilingual large vocabulary speech recognition: the European SQALE project, Computer Speech and Language, 1997, vol 11.

No context found.

S.J. Young et al., "Multilingual large vocabulary speech recognition: the European SQALE project," Computer Speech & Language, 11(1), pp. 73-89, Jan. 1997.

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