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,
)
| Model | Task1 | Task2 |
|---|---|---|
| Baseline | 0.2121 | 0.3138 |
The run.sh is the current recommended recipe.