Speech & Audio
Metric learning matched or beat classification-based losses for speaker recognition on the VoxCeleb dataset, a 2020 study reports
In a 2020 paper at Interspeech (USB), researchers at Naver compared popular loss functions for speaker recognition. Metric learning, the approach they defend, trains a model so that numerical summaries of one speaker’s clips end up close together and summaries of different speakers end up farther apart. The researchers found that a simple metric-learning loss called the vanilla triplet loss was competitive with classification-based losses. Models trained with their proposed metric-learning objective then outperformed what the authors describe as state-of-the-art methods on the VoxCeleb dataset.
What did the study compare?
The study evaluated popular loss functions for open-set speaker recognition, where a system is tested on speakers it did not see during training. The tests used the VoxCeleb dataset. The researchers compared metric-learning losses with classification-based losses. The comparison addressed a common belief that networks trained with classification objectives do better than metric-learning methods.
What did the study find?
The vanilla triplet loss, a simple metric-learning approach, was competitive with classification-based losses. A proposed metric-learning objective did better. The authors report that models trained with this objective outperformed state-of-the-art methods on VoxCeleb.
What is metric learning?
Metric learning trains a model so that speech clips from the same speaker produce summaries that sit close together, while clips from different speakers produce summaries that sit farther apart. In this paper, the goal was a compact summary of each short speech clip, with small distance between clips from the same speaker and large distance between clips from different speakers.
Why does the VoxCeleb result matter?
VoxCeleb is a widely used benchmark for speaker recognition, so results on it give other researchers a common point of comparison. Still, the abstract reports results on this dataset only.
Limits and what's next
The abstract describes one study on the VoxCeleb dataset. It does not report the number of speakers, utterances, or test clips used. The abstract also reports no independent replication or test on a second dataset, and it does not state the size of the performance advantage. The abstract does not say whether the same pattern holds outside VoxCeleb.
Source: Joon Son Chung et al., “In Defence of Metric Learning for Speaker Recognition,” Interspeech (USB), 2020.