1 / 6

Sphinx Recognizer Progress Q2 2004

Sphinx Recognizer Progress Q2 2004. Speed. Combination of fast GMM computation techniques with various types of pruning 0.48xRT in 2k task (Communicator), 0.6xRT in 5k task (WSJ) Phoneme Look-ahead research completed. 15-20% gain when fast GMM computation techniques and pruning was applied.

bill
Download Presentation

Sphinx Recognizer Progress Q2 2004

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Sphinx Recognizer Progress Q2 2004

  2. Speed • Combination of fast GMM computation techniques with various types of pruning • 0.48xRT in 2k task (Communicator), 0.6xRT in 5k task (WSJ) • Phoneme Look-ahead research completed. • 15-20% gain when fast GMM computation techniques and pruning was applied. • Detail can be found in paper in ICSLP 2004 (Chan et.al) • Compilation Optimization (~8% gain).

  3. Accuracy • S3.4 is much better than S2 • In Communicator task • S2: 17% WERR ~0.3xRT • S3.4: (32 mix, not tuned for speed) 14% WERR 1.1xRT • S3.4: (64 mix, not tuned for speed) 12% WERR 1.6xRT • Continuous HMM performs much better than Semi-Continuous HMM. • ~30% improvement

  4. Interface • Sphinx 3.4 Release Candidate II is distributed since Jun 10. • http://cmusphinx.sourceforge.net/downloads/sphinx3-0.4-rc2.tgz • New feature list: • Fast Gaussian Mixture Model (GMM) computation; the following techniques are supported:a. Down-sampling. b. CI-based GMM selection.c. VQ-based and SVQ-based Gaussian selection.d. Arbitrary configuration of Sub-vectors in sub-vector quantization. • Fast Match based on phoneme look-ahead. • Class-based LMs and dynamic LM selection. • Command-line configuration of many front-end parameters and the feature type. • Bug fixes to make "make test" work in wider range of platforms. • Bug fixes in live mode recognition. • Instructions for compilation using Intel compilers, if desired. • Better compilation support in Windows. • More documentation. • Batch mode recognizer supported on Windows/Linux/FreeBSD/MacOS/Solaris.

  5. Outlook of Sphinx in this year (Sphinx 3.5) • Sphinx 3.5 • Fully support live mode recognition API (now still testing, not in s3.4’s distribution) • ETA (End of July) • Fully support speaker adaptation techniques such as MLLR and VTLN. • ETA (Mid of August) • Sphinx trainer (SphinxTrain) • Better packaging and improvement on the training algorithm • ETA (Beginning of July) • SphinxDoc • A document on using Sphinx to build speech recognition software • ETA (End of October)

  6. Meeting Corpus Transcription and Training • ICSI Transcription Conversion Completed • Processing XML • As well as human transciber’s error • CHMM Training Completed • We chose a very difficult situation to decode • Transcriber meeting. • Many speakers and many cross talk. • Gives us a sense of the worst case performance. • 53% WERR

More Related