Presentation is loading. Please wait.

Presentation is loading. Please wait.

Advances in WP1 Trento Meeting 11-12 January 2007 www.loquendo.com.

Similar presentations


Presentation on theme: "Advances in WP1 Trento Meeting 11-12 January 2007 www.loquendo.com."— Presentation transcript:

1 Advances in WP1 Trento Meeting 11-12 January 2007 www.loquendo.com

2 2 WP1: Environment & Sensor Robustness T1.2 Noise Independence Noise Reduction: –Spectral Subtraction (YEAR 1) and Spectral Attenuation (YEAR2) –Evaluation of feature normalization techniques in Loquendo ASR (HEQ study + Revision 2) (Q3/4 YEAR2) (PEQ study) (YEAR3)

3 3 WP1: Speech Databases for Noise Reduction Aurora 2 - Connected digits - TIdigits data down sampled to 8 kHz, filtered with a G712 characteristic and noise artificially added at several SNRs (20dB, 15dB, 10 dB, 5dB, 0dB, -5dB). There are three test sets: –A: same noises as in train: subway, babble, car noise, exhibition hall; –B: 4 different noises: restaurant, street, airport, train station; –C: same noises as A but filtered with a different microphone Aurora 3 - Connected digits recorded in car environment - Signal collected by hand free (ch1) and close talk (ch0) microphones. In HIWIRE we use Italian and Spanish recordings. There are two test sets: –WM: ch0 and ch1 recordings used in training and testing lists; –HM: ch0 for training and ch1 for testing Aurora 4 - Continuous speech 5k vocabulary - It is WSJ0 5K with added noise of 6 kinds: Car, Babble, Restaurant, Street, Airport, Train station. It uses the standard Bi-Gram language modeling. CLEAN / MULTI-CONDITIONS training modes.

4 4 Y1,2+PEQ Front End configuration Front EndConfiguration RPLPRasta-PLP frame with energy plus 12 CEP, plus first(D) and second(DD) derivative + PEQParameter Equalization (UGR) applyied to the RPLP frame, plus D, DD + WIE SNRModified Rasta-PLP with Wiener-based denoising technique (PowSpec subtraction), plus D, DD + EM SNRModified Rasta-PLP with Ephraim-Malah-based denoising technique (PowSpec attenuation), plus D,DD + EM SNR + PEQModified Rasta-PLP with Ephraim-Malah plus Parameter Equalization, plus D, DD

5 PEQ + Loquendo ASR on Aurora2 speech databases

6 6 Y1,2+PEQ Performance evaluations Performances in terms of Word Accuracy and (Error Reduction – with respect to RPLP experiment) CLEAN ModelsTest ATest BTest CA-B-C RPLP75.677.575.376.3 + PEQ87.1(47.1)87.3(43.5)86.9(47.0)87.1(45.6) + WIE SNR84.0(34.4)84.4(30.7)83.3(32.4)84.0(32.5) + EM SNR85.3(39.7)84.2(29.5)84.8(34.5)84.8(35.9) + EM SNR + PEQ86.4(44.3)85.8(36.9)86.6(45.7)86.2(41.8)

7 7 Y1,2+PEQ Performance evaluations MULTI ModelsTest ATest BTest CA-B-C RPLP93.591.190.291.9 + PEQ92.8(-10.8)91.4(3.4)92.2(20.4)92.1(2.5) + WIE SNR93.9(6.1)92.1(11.2)90.5(3.1)92.5(7.4) + EM SNR94.0(7.7)92.0(10.1)91.1(9.2)92.6(8.6) + EM SNR + PEQ92.7(-12.3)91.1(0.0)91.7(15.3)91.9(0.0) Performances in terms of Word Accuracy and (Error Reduction – with respect to RPLP experiment)

8 PEQ + Loquendo ASR on Aurora3 speech databases

9 9 Y1,2+PEQ Performance evaluations Performances in terms of Word Accuracy and (Error Reduction – with respect to RPLP experiment) Aurora3 8kHzIta WMIta HMSpa WMSpa HM RPLP98.246.697.374.6 + PEQ97.6 (-33.3)79.7 (62.0)97.3 ( 0.0)87.1 (49.2) + WIE SNR98.3 ( 5.5)77.5 (59.4)97.6 (11.1)89.8 (59.8) + EM SNR98.4 (11.1)82.2 (66.7)97.7 (14.8)88.8 (55.8) + EM SNR + PEQ98.0 (-11.0)87.0 (75.6)97.8 (18.5)90.8 (63.8)

10 PEQ + Loquendo ASR on Aurora4 speech databases

11 11 Y1,2+PEQ Performance evaluations (Sennheiser microphone) CLEAN 8kHz CleanCarBabbleRest.StreetAirportTrain Station Noise Avg. RPLP85.254.323.129.434.029.332.333.7 + PEQ84.8 (-2.7) 69.5 (33.2) 54.3 (40.6) 50.2 (29.5) 52.7 (28.3) 53.5 (34.2) 53.8 (31.7) 55.7 (33.2) + WIE SNR 85.2 (0.0) 67.0 (27.8) 36.6 (17.5) 30.7 (1.8) 43.1 (13.8) 31.9 (3.7) 48.8 (24.4) 43.0 (14.0) + EM SNR85.5 (2.0) 70.4 (35.2) 37.1 (18.2) 31.6 (3.1) 45.8 (13.8) 31.6 (3.2) 53.7 (31.6) 45.0 (17.0) + EM SNR + PEQ85.4 (1.3) 70.9 (36.3) 53.9 (40.0) 49.5 (28.5) 54.9 (31.7) 51.3 (31.1) 59.1 (39.6) 56.6 (34.5) Performances in terms of Word Accuracy and (Error Reduction – with respect to RPLP experiment)

12 12 Y1,2+PEQ Performance evaluations (second microphone) CLEAN 8kHz CleanCarBabbleRest.StreetAirportTrain Station Noise Avg. RPLP59.435.716.221.422.919.322.823.1 + PEQ72.4 (33.0) 58.2 (34.9) 46.5 (36.1) 41.9 (26.1) 43.2 (26.3) 45.1 (31.9) 45.4 (29.3) 46.7 (30.7) + WIE SNR 60.1 (1.7) 50.2 (22.5) 25.7 (11.3) 23.6 (2.8) 29.2 (8.2) 22.6 (4.1) 35.1 (15.9) 31.1 (10.4) + EM SNR60.7 (3.2) 52.3 (25.8) 27.8 (13.8) 23.4 (2.5) 31.0 (10.5) 23.5 (5.2) 39.1 (21.1) 32.9 (12.7) + EM SNR + PEQ73.6 (34.9) 58.0 (34.6) 46.0 (35.5) 39.4 (22.9) 42.7 (25.7) 43.2 (29.6) 49.4 (34.4) 46.4 (30.3) Performances in terms of Word Accuracy and (Error Reduction – with respect to RPLP experiment)

13 13 WP1: Workplan Selection of suitable benchmark databases; (m6) Completion of LASR baseline experimentation of Spectral Subtraction (Wiener SNR dependent) (m12) Discriminative VAD (training+AURORA3 testing) (m16) Exprimentation of Spectral Attenuation rule (Ephraim-Malah SNR dependent) (m21) Preliminary results on spectral subtraction and HEQ techniques (m24) Integration of denoising and normalization techniques (PEQ) (m33)


Download ppt "Advances in WP1 Trento Meeting 11-12 January 2007 www.loquendo.com."

Similar presentations


Ads by Google