40 likes | 119 Views
Explore improved GMM computation methods in s3.4, such as advanced search structures, word-end pruning, and phoneme look-ahead, to optimize speed and efficiency in your system. Plan for training and deployment enhancements included.
E N D
Speed-up Facilities in s3.3 GMM Computation Seach Lexicon Structure Tree. Pruning Standard Heuristic Search Speed-up Not Implemented Frame-Level Not implemented Senone-Level Not implemented Gaussian-Level SVQ-based GMM Selection Sub-vector constrained to 3 Component-Level SVQ code removed
Summary ofSpeed-up Facilities in s3.4 GMM Computation Seach Lexicon Structure Tree Pruning (New) Improved Word-end Pruning Heuristic Search Speed-up (New) Phoneme-Look-ahead Frame-Level (New) Naïve Down-Sampling (New) Conditional Down-Sampling Senone-Level (New) CI-based GMM Selection Gaussian-Level (New) VQ-based GMM Selection (New) Unconstrained no. of sub-vectors in SVQ-based GMM Selection Component-Level (New) SVQ code enabled
Near Term Improvement of Decoder • Improve LM facilities (Avail at Mar 31) • Improve speed-up techniques (Avail at Mar 31) • Complete phoneme look-ahead research • Complete machine optimization in Intel platform • Enable speed-up in live-mode recognition. (Avail at Mar 31) • Improved search structure • Modify code to use lexical tree copies (Apr 15) • Modify code to handle cross-word triphones (Apr 30)
Training Plan • Text-Processing (Avail at Mar 31) • First Pass of Acoustic/Language Modeling (Avail at Apr 15) • With the help of the new 4 cpus machine. • Training using standard recipe • CD + CI mode first pass models. • Trigram models. • Second Pass of Acoustic/Language Modeling • Improved training. • Decide what we should do after we get the results. • AM/LM Adaptation? (Don’t know yet)