Developed new model representation and parameter sharing approach that increases recognition speed by cutting the amount of Gaussian computation by a factor of 2, while significantly improving accuracy and decreasing memory requirements for current state-of-the-art speech recognition systems
Developed new adaptive training algorithm that results in a factor of 3 reduction in amount of parameters necessary to store models, and a significant improvement in accuracy.
Designed and implemented the Gaussian Merging Splitting (GMS) training algorithm that largely automates the determination of model size given limited data, thereby resulting in robust training.
Developed and implemented several new lattice algorithms for multi-pass recognition, increasing the lattice recognition speed by more than a factor of 10, and decreasing lattice size by a factor of 6.
Applied acoustic adaptation algorithms developed under this project to train models for specific acoustic environments with small amount of training data. Developed acoustic clustering and cluster-specific adaptation algorithms to give significant improvements in recognition performance. These techniques have been used in MARVEL, a system for speech-based information archival and retrieval jointly developed under this project and the DARPA-sponsored SRI Spoken Language Systems project.
Developed automatic segmentation algorithm to break long speech streams into sentence-like fragments, greatly reducing the memory and computation requirements for recognizing long speech streams in real sources such as broadcast news.
Developed algorithms for MARVEL, a speech-based information archival and retrieval system jointly developed with the DARPA-sponsored SRI Spoken Language Systems porject. Innovative techniques include the use of statistical language models to rank the archived stories in order of relevance to the spoken query, and an algorithm for automatically generating a list of most salient words for the task.
Developed novel HMM state clustering algorithm that allows clustering of states both within and across phone boundaries by the use of broad phone classes. Resulted in a significant performance improvement in initial experiments.
Developed many new language modeling technologies including topic-specific language models, and methods to retrieve data from alternate data sources to augment the language model training data.