Research & Technology Experiments on AURORA 2000 database : Features of DC baseline system: RESPITE workshop " training on N1 ... N4 sets (multi−condition training) Jan.25−27 2001 " NSPS (nonlinear spectral subtraction) Martigny " VTN (vocaltract length normalization) " MFCC features with cepstral mean normalization Experiments on different feature sets; " „cepstral“ interface: „cepstral“ interface comparison with DC baseline system preprocessing, SCHMM feature extraction recognizer Joan Mari „cepstral“ data Hilario files Fritz Class
Research & Technology Comparison DC−baseline / ICSI−Tandem features RESPITE workshop % WER Jan.25−27 2001 Martigny average N1 ... N4 test sets clean SNR 20 SNR 10 SNR 0 Experiments on different feature sets; DC baseline 1.6 2.2 7.9 37.1 without LDA comparison with DC DC baseline baseline system 1.6 1.8 6.6 31.9 with LDA ICSI Tandem feat. 12 plp− 0.9 0.8 2.4 21.4 dd+msg; without LDA ICSI Tandem feat. 12 plp− Joan Mari 1.1 1.1 2.6 21.3 dd+msg; Hilario with LDA Fritz Class
Research & Technology Comparison DC−baseline / ICSI−Tandem features RESPITE workshop Jan.25−27 2001 50,00 Martigny DC−MFCC with LDA 40,00 DC−MFCC without %WER Experiments on 30,00 LDA different feature sets; ICSI−Tandem with 20,00 LDA comparison with DC ICSI−Tandem without baseline system 10,00 LDA 0,00 0 10 20 clean SNR Joan Mari Hilario Fritz Class
Research & Technology Comparison DC−baseline / FPM‘s RESPITE " FPM‘s: word models trained on clean speech workshop " DC: word models multi−condition training Jan.25−27 2001 % Martigny WER average N1 ... N4 test sets Experiments on different feature sets; clean SNR 20 SNR 10 SNR 0 comparison with DC baseline system FPM’s 1.0 3.8 21.6 72.7 log− RA STA FPM’s 1.3 3.1 13.2 55.8 J− RA STA FPM’s with 1.2 2.2 11.8 55.1 DC− models Joan Mari Hilario DC baseline 1.6 1.8 6.6 31.9 Fritz Class
Research & Technology Comparison DC−baseline / FPM‘s RESPITE " FPM‘s: word models trained on clean speech workshop " DC: word models multi−condition training Jan.25−27 2001 Martigny 50,00 Experiments on DC−baseline 40,00 different feature sets; comparison with DC FPM’s log−RASTA %WER 30,00 baseline system FPM’s J−RASTA 20,00 FPM’s with DC− 10,00 models 0,00 Joan Mari Hilario 0 10 20 clean SNR Fritz Class
Research & Technology RESPITE demonstrators RESPITE workshop Jan.25−27 2001 Statements: Martigny " our demonstrator strategy: „show project achivements (possibility of online application of the new techniques), not commercially relevant“!! " a demonstrator makes sence only, if there are better techniques than in the baseline system Discussion about ==> if we have really found such techniques (compared to the baseline system in offline simulations), we can build a demonstrator demonstrators " a full integration of the new techniques means a redesign of the complete system ==> not possible within RESPITE ==> combination of different modules (processes) via interfaces (files) or using DLL‘s under windows " a demonstration could be done e.g. in a car using a laptop Fritz Class
Research & Technology possible demonstration system: RESPITE TANDEM features with DC system workshop Jan.25−27 2001 architecture 1 Martigny TANDEM PL feature vectors Feature P Neural net MS DC−system classifier G calculation Discussion about file with demonstrators „cep“−feat ures process 1 process 2 architecture 2 Feature TANDEM Fritz calculation feature vectors Class DC−system (ICSI−software, 1 process under Windows TANDEM‘s) NT
Research & Technology RESPITE demonstrators RESPITE workshop Jan.25−27 2001 Questions: Martigny " sources (ICSI) for TANDEM features ? " missing data demonstrator ? " What are „potential users with respect to the demonstrators“? Discussion about " Is anywhere a online system available ? Portable? Under which demonstrators system? Fritz Class
Recommend
More recommend