Unsupervised Spectral Subtraction for Noise-Robust ASR on Unknown Transmission Channels

This directory contains a complete MATLAB implementation of channel normalization (CHN) followed by Unsupervised Spectral Subtraction (USS). USS was originally introduced in IEEE ASRU 2005. See IDIAP RR 06-09 for more details.

Below is a frequency-domain implementation.
You can also find a time-domain implementation there.

MAIN.m is the point where to start (documentation and MATLAB examples).
It contains several examples calling "uss_filter.m".

uss_filter.m is the main function called by the examples:
Input: one audio waveform file (WAV, SPHERE or raw), in 16-bit format.
Output(s): "cleaned" spectrogram, filterbank values, MFCCs.

Questions? ->

Click here if you want to download the whole directory.

/com/mmm/data/Lathoud/2006-CHN-USS/: DATA/; MAIN.m; check_param.m; chn.m; chrono_check.m; chrono_start.m; chrono_stop.m; compute_post_raylsherl.m; fill_default.m; fit_raylsherl.m; gmn_uss.m; my_htk_read.m; my_logsum_fast.m; my_logsum_fast.pdf; my_spectrum.m; rr-06-09.pdf; rse_uss.m; running_mean.m; uss_filter.m

Back to the audio resources page

Code and webpage by Guillaume Lathoud ( lathoud AT idiap DOT ch )
Last updated: 2006-12-22