en-de
en-es
en-fr
en-sl
en
en-zh
0.25
0.5
0.75
1.25
1.5
1.75
2
Least Squares Filtering of Speech Signals for Robust ASR
Published on Feb 25, 20072939 Views
Chapter list
Least Squares Filtering of Speech Signals for Robust ASR00:01
Introduction00:55
Schematic of least squares filter03:25
N: block length, L: filter length M: # of sinusoids04:15
TITLE05:30
TITLE06:05
TITLE07:27
Alpha=1/N08:55
LeSF gain=(o/p SNR)/(i/p SNR) for fixed filter length ‘L=100’10:25
LeSF gain=(o/p SNR)/(i/p SNR) for fixed frame length ‘N=500’11:02
11:15
An example vocal tract frequency response with two poles11:56
Spectrum enhancement using a 60th order LP filter. Blue: clean spectrum, Magenta: Noisy at SNR 0db AWGN, Red: Spectrum obtained after filtering using a 60th order LP Filter.12:09
Embedded in F-16 cockpit noise SNR 6db12:37
Parameters L, M, N13:19
Intuitive explanation Autocorr. of the broadband noise decays faster. High order LeSF predicts narrow-band components, atten. broadband noise the LP filter can put enough narrow-bandwidth pole14:08
Numbers95 database, additive factory noise. WER reported15:32
Conclusions Derived analytic solution of LeSF filter for nonstationary signals. Analysis of BW and LeSF gain relations. Leads to significant ASR rob16:46
Thank you for your attention17:26