Mobio Dataset, female subjects¶
Dataset Description¶
Mobio is a collection of English voice recordings. The set of female subjects contains:
Identities |
Sample count |
||
train |
13 |
2406 |
|
dev |
references |
18 |
90 |
probes |
1890 |
||
eval |
references |
20 |
100 |
probes |
2100 |
GMM¶
To run the baseline, use the following command:
bob bio pipeline simple -d mobio-audio-female gmm-mobio -g dev -g eval -l sge -o results/gmm_mobio_female
Then, to generate the scores, use:
bob bio metrics -e ./results/gmm_mobio_female/scores-{dev,eval}.csv
Development |
Evaluation |
|
---|---|---|
Failure to Acquire |
0.0% |
0.0% |
False Match Rate |
20.6% (6632/32130) |
7.8% (3093/39900) |
False Non Match Rate |
20.6% (390/1890) |
26.8% (562/2100) |
False Accept Rate |
20.6% |
7.8% |
False Reject Rate |
20.6% |
26.8% |
Half Total Error Rate |
20.6% |
17.3% |
On 128[1] CPU nodes on the SGE Grid: Ran in 15 minutes (11 minutes of training).
ISV¶
To run the baseline, use the following command:
bob bio pipeline simple -d mobio-audio-female -p isv-default -g dev -g eval -l sge -o results/isv_mobio_female
Then, to generate the scores, use:
bob bio metrics -e ./results/isv_mobio_female/scores-{dev,eval}.csv
Development |
Evaluation |
|
---|---|---|
Failure to Acquire |
0.0% |
0.0% |
False Match Rate |
14.7% (4710/32130) |
20.3% (8103/39900) |
False Non Match Rate |
14.7% (277/1890) |
17.4% (366/2100) |
False Accept Rate |
14.7% |
20.3% |
False Reject Rate |
14.7% |
17.4% |
Half Total Error Rate |
14.7% |
18.9% |
On 128[1] CPU nodes on the SGE Grid: Ran in 8 minutes (3 minutes of training).
Speechbrain ECAPA-TDNN¶
To run the baseline, use the following command:
bob bio pipeline simple -d mobio-audio-female -p speechbrain-ecapa-voxceleb -g dev -g eval -l sge -o results/speechbrain_mobio_female
Then, to generate the scores, use:
bob bio metrics -e ./results/speechbrain_mobio_female/scores-{dev,eval}.csv
Development |
Evaluation |
|
---|---|---|
Failure to Acquire |
0.0% |
0.0% |
False Match Rate |
1.9% (616/32130) |
10.8% (4307/39900) |
False Non Match Rate |
1.9% (36/1890) |
2.5% (52/2100) |
False Accept Rate |
1.9% |
10.8% |
False Reject Rate |
1.9% |
2.5% |
Half Total Error Rate |
1.9% |
6.6% |
On 128[1] CPU nodes on the SGE Grid: 12 minutes (no training).
Footnotes