Answer Keys & Supporting Files
Answer Key
The answer key maps the test segments to the true speakers. Note that
the true speaker does NOT have the same PIN as the models used in the
evaluation.
# Column 1 - Test segment
# Column 2 - Language of the test segment
# Column 3 - Source conversation
# Column 4 - channel (A, B, X1 or X2 when summed)
Note, with summed test segments there are two true
speakers so the test segment entry is made twice.
# Column 5 - Segment type
1s - 1 side
1c - 1 conversation (summed)
10 - 10 seconds
30 - 30 seconds
# Column 6 - True speaker (MIXER PIN)
# Column 7 - Gender
# Column 8 - Actual segment length in seconds
# Column 9 - Dialect of the true speaker (nontar if side is not
one of the targets.)
# Column 10 - Microphone type
Speakerphone, Headset, Ear-bud, Regular, Mixed, Unknown
# Column 11 - Phone Type
Cellular, Landline, Cordless, Mixed, Unknown
Models Mapped to Speakers
For this evaluation there were several models created from the same
speaker. This file maps the evaluation MODEL ID to the MIXER true
speaker pin found in the answer key.
# Column 1 - Four digit model id used for evaluation
# Column 2 - Four digit MIXER corpus pin
# Column 3 - Dialect (A=Arabic, E=English, M=Mandarin,
S=Spanish, R=Russian)
# Column 4 - Gender
# Column 5 - Training condition
1 - 1 side training
3 - 3 side training
8 - 8 side training
16 - 16 side training
10s - 10 second training
30s - 30 second training
3c - 3 conversations (summed) training
Training Languages
This file identifies the different languages and combinination of
langauges used for training each model.
# Column 1 - Four digit model id used for evaluation
# Column 2 - Number of Arabic training segments
# Column 3 - Number of English training segments
# Column 4 - Number of Mandarin training segments
# Column 5 - Number of Russian training segments
# Column 6 - Number of Spanish training segments
Training Handsets
This file identifies the different handsets and combination of handsets used for training models with single channel data.
# Column 1 - Four digit model id used for evaluation
# Column 2 - Training Condition
# Column 3 - Number of different telephone numbers
# Column 4 - Microphone type
Speakerphone, Headset, Ear-bud, Regular, Mixed, Unknown
# Column 5 - Phone Type
Cellular, Landline, Cordless, Mixed, Unknown
3-conversation Training
This file defines attributes specific to the 3 conversation training models.
# Column 1 - Four digit model id used for evaluationk
# Column 2 - Gender mix (there are six possible sides)
nFnM (Number of six that are Female, number of six
that are male)
6F0M, 5F1M, 4F2M, 3F3M, 2F4M, 1F5M 0F6M
Sub-Models of the 16-sides Models
This file identifies all pure sub-models of a given 16-side training
model. By "pure" we refer to only those sub-models that will have an
8-side, 3side, 1side, 30-sec and 10sec sub training model.
# Column 1 - M16 marker
# Column 2 - Model 16 ID
# Column 3 - M8 marker
# Column 4 - Model 8 ID (sub of 16)
# Column 5 - M3 marker
# Column 6 - Model 3 ID (sub of 8)
# Column 7 - M1 marker
# Column 8 - Model 1 ID (sub of 3)
# Column 9 - M30s marker
# Column 10 - Model 30-second ID (sub of 1)
# Column 11 - M10s marker
# Column 12 - Model 10-second ID (sub of 30-second)
# Column 13 - Model 3-conversation marker
# Column 14 - Model 3-conversation ID (same as 3-sides)
Same as above, but starting point is models of size 8.
Same as above, but starting point is models of size 3.
Mapping Models of 3-sides to Models of 3-conversations
This file maps each 3-sides training model to the corresponding 3-conversation model.
# Column 1 - Model 3-sides marker
# Column 2 - Model 3 sides id
# Column 3 - Model 3-conversations marker
# Column 4 - Model 3 convs id