NIST Speech Group Website

Test Material

The evaluation test material is distributed on 9 CDROM's.

R65_1_1: Female training data
R65_2_1: Male training data
R65_3_1: Female 1-Speaker detection task test data
R65_4_1: Male 1-Speaker detection task test data
R65_5_1: Male & Female 1-Speaker detection task test segments
R65_6_1: Multi-Speaker test segments for 2-Speaker detection, speaker tracking, and segmentation with known number of speakers
R65_7_1: Multi-Speaker test segments for 2-Speaker detection
R65_8_1: CALLHOME test segments for segmentation with unknown number of speakers
R65_9_1: Spanish AHUMADA training and test data for 1-Speaker detection task

Training Data

Training data is provided separately for males and females.

R65_1_1: ./sid00tr1/female
R65_2_1: ./sid00tr2/males

Training data for the optional Spanish language 1-speaker detection task is provided for males only.

R65_9_1: ./sid00ah1/train

1-Speaker Detection

To fully implement the NIST's basic task of 1-speaker detection using conversational telephone speech, the following index files must be completely processed:

R65_3_1: ./sid00e1f/data/detect1.ndx
R65_4_1: ./sid00e1m/data/detect1.ndx
R65_5_1: ./sid00e1b/data_f/detect1.ndx
R65_5_1: ./sid00e1b/data_m/detect1.ndx

2-Speaker Detection

To fully implement the task of 2-speaker detection using conversational telephone speech, the following index files must be completely processed:

R65_6_1: ./sid00e2a/data/detect2.ndx
R65_7_1: ./sid00e2b/data/detect2.ndx

Speaker Tracking

To fully implement the task of Speaker Tracking using conversational telephone speech, the following index file must be completely processed:

R65_6_1: ./sid00e2a/data/tracking.ndx

Segmentation (known speaker count)

To fully implement the task of segmentation when it is known that there are exactly two speakers present in the test segment, the following index file must be completely processed:

R65_6_1: ./sid00e2a/data/segment.ndx

Segmentation (Unknown speaker count)

To fully implement the task of segmentation on test-segments from various languages, with the number of speakers present in the test segment left unknown, the following index file must be completely processed:

R65_8_1: ./sid00sg1/data/segment.ndx

1-Speaker Detection using spontaneous Spanish data

To fully implement the optional task of 1-speaker detection using Spanish data taken from the Ahumada corpus, the following index file must be completely processed:

R65_9_1: ./sid00ah1/test/detect1.ndx
Replacement CD, same index file: R65_9_2: ./sid00ah1/test/detect1.ndx

Miscellaneous Information

Consult the Evaluation Plan for detailed information regarding this evaluation.

The index files listed contain the proper number of trials for each test segment in the specific task. REMINDER: If sites wish to process the complete matrix of all TARGET speakers against EVERY test segment for ANY task, NIST will score and provide results. These submissions should be made separately, and will be accepted at anytime.

DISC R65_9_2 is a replacement CD for the Ahumada data. Incorrect sample counts were fixed, and handset labeling information was added.

The following test segments have the handset type set to "electret" by default, there are not any probability or likelihood scores.

gszs	FEMALE
gtis	FEMALE
gase	FEMALE
gbcu	FEMALE
gbzq	FEMALE
gdgr	FEMALE
geqj	FEMALE
gjzb	FEMALE
gkjj	FEMALE
gmyu	FEMALE
gbfi	MALE
ghex	MALE
gjga	MALE
glzb	MALE
gsxx	MALE
ghss	MALE
giye	MALE
gjep	MALE
gmky	MALE
gomi	MALE
gpki	MALE
gteu	MALE
guhu	MALE

2001 Speaker Recognition Evaluation Supporting Documentation