| EPSRC Reference: |
GR/K25380/01 |
| Title: |
ROBUST VERY LARGE VOCABULARY CONTINUOUS SPEECH RECOGNITION |
| Principal Investigator: |
Woodland, Professor PC |
| Other Investigators: |
|
| Researcher Co-Investigators: |
|
| Project Partners: |
|
| Department: |
Engineering |
| Organisation: |
University of Cambridge |
| Scheme: |
Standard Research (Pre-FEC) |
| Starts: |
09 August 1994 |
Ends: |
08 February 1997 |
Value (£): |
167,344
|
| EPSRC Research Topic Classifications: |
| Human Communication in ICT |
|
|
| EPSRC Industrial Sector Classifications: |
|
| Related Grants: |
|
| Panel History: |
|
|
Summary on Grant Application Form |
|
To improve the performance of an HMM-based large vocabulary speaker independent continuous speech recognition (CSR) system and in particular: Increase the recogniser vocabulary from 20,000 words to 65,000 words and above. Develop speaker adaptation techniques in the context of a large vocabulary system, concentrating on techniques for rapid speaker adaptation. Develop and integrate techniques for noise and channel compensation. Develop and integrate improved statistical language models. To continue to participate in the ARPA-sponsored CSR evaluations to demonstrate the utility of the techniques developed and compare with the state of the art.Progress:We have been working towards all the above objectives since the start of the project. The main acoustic modelling component of the system has also been improved by work on EPSRC/MOD Project GR/J10204. The vocabulary of the existing HTK Large Vocabulary Continuous Speech Recognition System has been extended from 20,000 words to 65,000 words. This involved developing a suitable pronunciation lexicon for this increased vocabulary as well as building statistical N-gram language models from a very large corpus (237 million words) of text training data. Software for language modelling that can work with this very large data set was developed which allowed rapid experimentation with both standard trigram and improved four-gram language models. If the same speaker uses the system for a number of utterances, performance can be improved by using speaker adaptation. A key issue is adapting a very large number of parameters with a small amount of speaker-specific data. Furthermore the adaptation should ideally be unsupervised (the correct transcription is unknown) and incremental so that the system improves with continued use. We have incorporated a technique that we had previously developed, called maximum likelihood linear regression (MLLR), into the large vocabulary system. We have found that MLLR adaptation is effective with limited adaptation data and can operate in incremental unsupervised mode. We have started work on improving the robustness of our large vocabulary system using the parallel model combination (PMC) technique developed in the Cambridge Speech Group. This work was evaluated in the relevant test of the November 1994 ARPA-sponsored CSR evaluation, and PMC was found to greatly enhance tolerance to additive noise. The current version of our 20k static recogniser returned the lowest error rate of any system in the November 1994 ARPA CSR H1-C1 test, and the developments reported above in extending vocabulary, language modelling and speaker adaptation resulted in a system with the lowest error rate in the November 1994 H1-P0 test [1],[2]. These results demonstrate that the systems developed define the state of the art in large vocabulary speaker independent continuous speech recognition. [1] Woodland P.C., Leggetter C.J., Odell J.J., Valtchev V. Young S.J. The 1994 HTK Large Vocabulary Speech Recognition System. Proc ICASSP, Detroit, 1995. [2] Woodland P.C., Leggetter C.J., Odell J.J., Valtchev V., Young S.J. The Development of the 1994 HTK Large Vocabulary Speech Recognition System. Proc. ARPA Spoken Language Technology Workshop, Barton Creek, 1995.
|
|
Key Findings |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Potential use in non-academic contexts |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
|
Impacts |
|
| Description |
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk |
| Summary |
|
| Date Materialised |
|
|
|
|
Sectors submitted by the Researcher |
|
This information can now be found on Gateway to Research (GtR) http://gtr.rcuk.ac.uk
|
| Project URL: |
|
| Further Information: |
|
| Organisation Website: |
http://www.cam.ac.uk |