SUMMARY OF SESSION 10 - Continous Speech Recognition II 
George Doddington, chairman 
Algorithms and techniques to improve the robustness of speech recognition were the 
principal theme in session 10. 
The session opened with two papers which addressed the compensation for noise and 
microphone characteristics. Significant performance improvement was demonstrated. 
Michael Picheny made an important comment in response to these papers, namely that these 
techniques have been developed before. Why are they not incorporated as a standard part 
of speech preprocessing for all speech recognition systems? 
Ostendorf presented work on the modeling of temporal correlation and discussed some 
interesting techniques. Although this work is in an early stage, the issues and ideas treated 
may be an area of great important to progress in speech recognition. Although the work was 
contrasted as being distinct from HMM, the general ideas could and should be applied using 
HMM's as well. 
The session concluded with a paper on the use of decision trees in speech recognition and 
one on the use of parallel algorithms for improved speech recognition. 
309 
