This special issue on Advances in Large Vocabulary Speech Recognition includes seven contributions solicited by the ISCA ITRW ASR'2000 Scientific Committee for publication in Computer Speech and Language. The focus of the ASR'2000 workshop, entitled "Automatic Speech Recognition: Challenges for the new Millennium", was on core speech recognition technology and in particular on recent advances and prospective work (http://www-tlp. limsi.fr/asr2000). Some of the cited challenges for speech recognition research were efforts to develop robust generic technology, low-cost development strategies, eased portability to new tasks and languages, and adaptability. Over 60 full paper submissions were received and reviewed by the Scientific Committee, and 32 were retained for oral presentation in eight sessions: Acoustic Model Training; Decoding; Noise Robustness (I, II); Language and Pronunciation Modeling; Acoustic Model Adaptation; Acoustic Modeling; Error Analysis, Confidence Measures and Metadata. The conference attendance was limited to 115 people to ensure an informal atmosphere and encourage technical exchange and open discussion. In addition to the reviewed papers, there were 5 keynote talks.
The workshop started with a keynote presentation by Phil Woodland (Cambridge University Engineering Department, Cambridge, United Kingdom) who presented recent work on using large scale discriminative training for large vocabulary speech recognition. The next two keynotes addressed decoding techniques for large vocabulary speech recognition. Xavier Aubert (Philips Research Laboratories, Aachen, Germany) gave an overview of some of the most efficient decoding techniques in current use; and Michael Riley (AT&T Labs Research, Floram Park, New Jersey) addressed the use of weighted finite-state transducers in speech recognition. These three keynotes have written contributions in this special issue. The fourth keynote talk entitled "Towards Super-Human Speech Recognition," by Michael Picheny (IBM T. J. Watson Research Center, Yorktown Heights, New York) highlighted some of the outstanding challenges in speech recognition, making reference to the large gap between the performance of machines and humans. In particular, he pointed out the inadequacy of current modeling techniques as evidenced by the high recognition error rates on spontaneous speech, and the lack of widespread use of the technology due to the need to tune a system to a particular task/environment. The final keynote, presented by Michael H. Cohen (Nuance Communications, Menlo Park, California) was entitled "Surfing the Voice Web: Issues in the design of a voice browser" and illustrated future potential voice-enabled, wireless web-based services. Such VoiceWeb access will evidently require sophisticated spoken language technology and poses challenges for the underlying technologies, infrastructure, and user interface paradigms. The workshop concluded with an afternoon session on national and international speech projects and future prospects, which included presentations by funding representatives from the European Commission and DARPA, as well as some short presentations of selected research projects. A final discussion period was led by the EC Coretex project on "Outstanding challenges and future directions" for speech recognition.