SPHINX-II For WSJ CSR Task.pdf

Key advancements and adaptations include:

1. System Enhancements:
- Incorporation of additional dynamic and speaker-normalized features.
- Replacement of discrete models with sex-dependent semi-continuous hidden Markov models (SCHMMs).
- Augmentation of within-word triphones with between-word triphones.
- Extension of generalized triphone models to shared-distribution models.

2. Addressing Task-Specific Challenges:
- Extension of the Viterbi beam-search to handle the larger vocabulary and more general task domain of WSJ CSR.
- Development of a practical, time- and memory-efficient between-word co-articulation modeling using left context-dependent between-word triphones.
- Significant reduction in memory requirements through algorithmic changes and data structure optimizations.

3. SPHINX-II System Review:
- SPHINX-II reduces word error rate by more than 50% compared to SPHINX through between-word coarticulation modeling, high-order dynamics, sex-dependent SCHMMs, and shared-distribution models.
- The system uses LPC cepstral coefficients, differenced LPC cepstral coefficients, power, and relative power as speech features, vector-quantized into four independent codebooks.
- Training involves two stages: generating shared-distribution mapping tables and training 4-codebook models with sex-dependent SCHMMs.
- Recognition uses a Viterbi beam search with activated models based on the automatically determined sex of the input utterance.

4. Funding and Disclaimer:
- The research is supported by DARPA and monitored by the Space and Naval Warfare Systems Center under contract N00039-91-C-0158.
- Views and conclusions are those of the authors and do not represent official policies, either expressed or implied, of DARPA or the U.S. Government.