Presentation is loading. Please wait.

Presentation is loading. Please wait.

HIWIRE MEETING Paris, February 11, 2005 JOSÉ C. SEGURA LUNA GSTC UGR.

Similar presentations


Presentation on theme: "HIWIRE MEETING Paris, February 11, 2005 JOSÉ C. SEGURA LUNA GSTC UGR."— Presentation transcript:

1 HIWIRE MEETING Paris, February 11, 2005 JOSÉ C. SEGURA LUNA GSTC UGR

2 2 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Schedule  AURORA 4 HTK-based setup  Baseline results (AURORA databases)  MFCC with C0 and CMN  AFE  Additional results  CMVN  HEQ  Work in progress  WP1: Improved HEQ  WP2: User independence & robustness

3 3 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna AURORA 4 HTK-based setup  ETSI AURORA 4 evaluation  Baseline system based on ISIP speech recognition system  Main drawbacks:  CPU time for experiments (specially for decoding)  Scripts are excessively complex to use  Described in:  N. Parihar and J. Picone, "DSR Front End LVCSR Evaluation - AU/384/02," Aurora Working Group, ETSI, December 06, 2002.  G. Hirsch, "Experimental Framework for the Performance Evaluation of Speech Recognition Front-ends on a Large Vocabulary Task, Version 2.0," ETSI STQ-Aurora DSR Working Group, November 19, 2002.

4 4 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna AURORA 4 HTK-based setup  HTK-based setup for AURORA 4 evaluations  Features  12MFCC + C0 (CMS) + Δ + Δ Δ  Cross-word tree-based tied-state tri-phones  3 states / 6 Gaussians per state  Back-off bi-gram language model  Same as used in ISIP setup  Pruning is performed as in ISIP setup  Available for partners at: http://www.hiwire.orghttp://www.hiwire.org

5 5 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna AURORA 4 HTK-based setup  Performance comparisons (HTK-based setup vs. ISIP)  Training clean models from scratch takes 3h52‘ on a 2.66GHz Word error rateDecoding time (s) ISIPHTKISIPHTK Test 01 (clean data) 16.2%13.22% 7580 (6.16  RT) 3428 (2.78  RT) Test 02 (car noise) 49.6%24.68% 22195 (18.03  RT) 8002 (6.50  RT) Test 03 (babble noise) 62.2%46.00% 33203 (26.9  RT) 13747 (11.17  RT) 12 MFCCs + C0 (CMS) +  + 

6 6 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna AURORA 4 Baseline results

7 7 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna AURORA 4 Additional results

8 8 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Baseline results  HIWIRE baseline results: 12 MFCCs + C0 (CMS) +  +  AURORA 2

9 9 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Baseline results  AFE AURORA 2

10 10 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Baseline results  AURORA 3 word error rates

11 11 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Work in progress (WP1)  Improved equalization  Modeling Speech & Noise separately  First results with Gaussian models Very promising on AURORA 4 Need to be evaluated on AURORA 2 & 3  Next Use more detailed / nonparametric models Incorporate dynamic features

12 12 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Preliminary results

13 13 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Work in progress (WP1)  VAD & Noise reduction  Baseline evaluations AURORA 2 & 3 already done AURORA 4 to be ready on June  Integration with parametric techniques Speech & Noise equalization

14 14 HIWIRE Meeting – Paris, 11 February, 2005José C. Segura Luna Work in progress (WP2)  HEQ-based user robustness  Ready for AURORA 4  Working in WSJ1 baseline  HEQ-based user adaptation  MLLR baseline  Estimation of MLLR transformations using HEQ  Working in WSJ1 baseline

15 HIWIRE MEETING Paris, February 11, 2005 JOSÉ C. SEGURA LUNA GSTC UGR


Download ppt "HIWIRE MEETING Paris, February 11, 2005 JOSÉ C. SEGURA LUNA GSTC UGR."

Similar presentations


Ads by Google