Uday Jain, Connected Digit Recognition over Long Distance Telephone Linesusing the SPHINX-II System MS report, CMU, May 1996 Abstract This report documents the performance of the SPHINX II system on two connected digit databases. The two databases, MALL88 and MALL91 were collected on long distance telephone lines at two sites, and they have the identical vocabulary of the ten digits + "OH". We describe the performance of the system using existing models developed for large vocabulary speech recognition. To further improve recognition accuracy two training procedures (bootstrapped training and data-driven training) are described and compared. It was found that further improvements can be obtained by untying parameters trained and increasing the model size. This is possible due to the small vocabulary size. To facilitate the untying of the parameters a new phone set was defined and trained, which separated distributions that were shared in previous training procedures. The new phones also ensure that training data is not shared between words. Results for all three training procedures are presented. Application of the procedures described in this report reduced the word error rate from 18 percent (using systems that had been trained for large vocabulary recognition) to 1.6 percent. Finally the use of channel and test-set normalization was also explored. It was found that the CDCN algorithm (which compensates for unknown additive noise and unknown linear filtering) did not provide further improvements to recognition accuracy for these data.