advances in wp1

16
Advances in WP1 Chania Meeting – May 2007 www.loquendo.com

Upload: dalton

Post on 02-Feb-2016

29 views

Category:

Documents


0 download

DESCRIPTION

Chania Meeting – May 2007. Advances in WP1. www.loquendo.com. Summary. Test on Hiwire DB with denoising methods developed in the project: Wiener SNR dep. Spectral Subtraction Ephraim-Malah SNR dep. Spectral Attenuation Loquendo FE – UGR PEQ Integration Details Results on Hiwire db. - PowerPoint PPT Presentation

TRANSCRIPT

Page 1: Advances in WP1

Advances in WP1

Chania Meeting – May 2007

www.loquendo.com

Page 2: Advances in WP1

2

Summary

• Test on Hiwire DB with denoising methods developed in the project:– Wiener SNR dep. Spectral Subtraction– Ephraim-Malah SNR dep. Spectral Attenuation

• Loquendo FE – UGR PEQ Integration– Details– Results on Hiwire db

Page 3: Advances in WP1

HIWIRE DB Test

Chania Meeting – May 2007

www.loquendo.com

Page 4: Advances in WP1

4

Test Conditions

• Test on the last 50 utterances of each speaker (50-99)• The first 50 utterances of each speaker (0-50) left for

development or adaptation• Four noise conditions:

– Clean

– Low Noise (SNR = 10 dB)

– Medium Noise (SNR = 5 dB)

– High Noise (SNR = -5 dB)

• 4049 utterances for each condition, from 81 speakers of 4 nationalities

Page 5: Advances in WP1

5

HMM-ANN Models

Two HMM-ANN models have been trained:

• Telephone 8 kHz: trained with a large telephone corpus (LDC Macrophone + SpeechDat Mobile)

• Microphone 16 kHz: trained with a collection of microphone corpora (timit, wsj0-1, vehic1us-ch0)

Page 6: Advances in WP1

6

Test Results

Models Denoising method

Noise Condition AVG E.R. %

Clean LN MN HN

Telephone

8 kHz(Macrophone)

No Den 88.4 51.1 27.3 2.8 42.4 -

WIE 88.3 70.0 54.1 16.3 57.2 25.7

EM 88.3 74.7 62.0 20.1 61.3 32.8

Microphone 16kHz

(timit-wsj0-1-vehic1us)

No Den 90.5 49.1 27.5 5.0 43.0 -

WIE 90.4 68.5 51.1 14.5 56.2 23.2

EM 90.2 71.9 55.0 16.6 58.4 27.0

Page 7: Advances in WP1

7

Test Results

0

10

20

30

40

50

60

70

80

90

100

No Den WIE EM No Den WIE EM

Telephone 8 kHz Microphone 16kHz

wo

rd a

ccu

racy

% Clean

LN

MN

HN

AVG

Page 8: Advances in WP1

8

Comments on Results

• The 16 kHz models are more accurate on clean speech (90.5% vs. 88.4%)

• Ephraim-Malah noise reduction always outperforms Wiener spectral subtraction (32.8% vs. 25.7% and 25.7% vs. 21.8% E.R.).

Page 9: Advances in WP1

Loquendo FE UGR PEQintegration

Chania Meeting – May 2007

www.loquendo.com

Page 10: Advances in WP1

10

PEQ Integration (Loquendo & UGR)

Loquendo FE

UGR PEQ

Loquendo ASR

Denoise

(Power Spectrum level)

Feature Normalization

(Frame -13 coeff- level)

Phoneme-based

Models

Page 11: Advances in WP1

11

PEQ effects

Page 12: Advances in WP1

12

PEQ Results

Models Den. Norm. Noise Condition AVG

Clean LN MN HN

wsj0 16 kHz NO NO 89.3 44.2 20.9 2.0 39.1

wsj0 16 kHz E.M. NO 89.2 69.6 53.7 15.4 57.0

wsj0 16 kHz NO PEQ 85.7 67.2 50.4 14.7 54.5

wsj0 16 kHz E.M. PEQ 85.2 73.7 59.5 19.8 59.5

The HMM-ANN models employed are:

• WSJ0 models

• WSJ0 models + E.M. denoising

• WSJ0 models + E.M. denoising + PEQ

Page 13: Advances in WP1

13

EM Denoise and PEQ

0

10

20

30

40

50

60

70

80

90

100

Clean LN MN HN AVG

Noise Conditions

wo

rd a

ccu

racy

%

NO-DEN NO-PEQ DEN NO-PEQ NO-DEN PEQ DEN PEQ

Page 14: Advances in WP1

14

Comments on EM denoising - PEQ

• On noisy speech (LN, MN, HN):– both EM denoising and PEQ obtain a good improvement

– best results are obtained when adding the effects of EM de-noising and PEQ normalization.

• On clean speech:– EM denoising does not decrease performances

– PEQ normalization slightly decreases performances

• PEQ is very useful in mismatched conditions

• can (slightly) decrease performances in matched conditions (e.g. clean speech)

Page 15: Advances in WP1

15

Test on TTS American Voice (Dave)

Models Dave Hiwire DB

clean

Telephone 8 kHz

(Macrophone)98.9 88.3

Micro 16 kHz (wsj0) 99.7 88.1

• We have used the American voice DAVE of Loquendo TTS to read the 4049 sentences of the Hiwire DB

• The great difference in results is due to non-native pronounce

• Es. “Range Forty” pronounced

• by Dave

• by a French speaker

• by a Greek speaker

Page 16: Advances in WP1

16

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)