This is the baseline system for PVTC2020 which is a satellite event of ISCSLP2021(https://www.iscslp2021.org), for more information about the challenge and dataset you can visit the website https://www.pvtc2020.org.
Our baseline method consists of a wake-up system and a speaker verification system. As shown in the figure below, we designed a two-pass approach to respond whenever the target speaker says the wake-up word. When the wake-up word detection system triggers, the speaker verification system starts to decide whether the audio segment that triggered the wake-up word detector is indeed spoken by the enrolled target speaker.
Pleaes refer to KWS_README.md and SV_README.md for more details
.
In this challenge, we provide a leaderboard, ranked by the metric . As for the
metric, the lower the better. The
is provided as our challenge metric and it is calculated from the miss rate and the false alarm (FA) rate in the following form:
Results are shown in S_kws_task1.jpg and S_kws_task2.jpg. We choose the final score under alpha is equal to nineteen as model's performance criterion. (=0.05,
)
By updating the method of determining the threshold of the speaker verification system (using the mean threshold of EER and minDCF instead of the threshold of EER), we proposed Baseline_v2, which has been greatly improved in the development set.
| Model | Task1 | Task2 |
|---|---|---|
| Baseline_v1 | 0.1981 | 0.3334 |
| Baseline_v2 | 0.1009 | 0.1415 |
All the result are based on the development set.
The run.sh is the current recommended recipe.
Contact us: [email protected]