Segmentation.tar
Collection of segmenter utilities for 96/97 Hub-4 CSR task.

man:    Matthew Siegler
        msiegler@cs.cmu.edu
        August 1997


This collection contains:

wave2mfcc		convert audio files into mfcc (cepstra) files
UTT_Kseg		find statistical changepoints
UTT_findsil		find silences near changepoints
UTT_gauss_class		acoustically classify segments
UTT_cluster		cluster utterances based on KL2 metric
 (two version exist, 96 and 97, the latter working better in the UE for H4-97)

lib/mfc_io		mfc read and write routines
lib/fe			Front End for wave2mfcc, (requires NIST sphere 2.5 or greater,
			contact jonathan.fiscus@nist.gov)
lib/NIST		NIST sphere package (not included in this distribution)
			See the readme for directory format.


Each utility has a README with it to aid in running the program, including the
parameters that CMU used to run the Hub 4 1997 evaluation. In the evaluation, we ran
the following (in order)

(for UE data)
UTT_Kseg		To locate change in acoustic conditions
UTT_findsil		To find the places to legally draw a break
UTT_gauss_class		To detect telephone BW speech
UTT_cluster		To cluster the utterances for MLLR adaptation

(for PE data)
UTT_findsil 
UTT_gauss_class
UTT_cluster


Most of the algorithms used in the segmentation, classification and clustering are found in:

	Siegler, Jain, Raj and Stern,
	"Acoustic Segmentation, Classification and Clustering of Broadcast News Audio,"
	Processdings of the Speech Recognition Workshop, 1997.


