Speech & Audio
A redesigned speaker verification network significantly outperformed state-of-the-art TDNN-based systems on VoxCeleb test sets and the 2019 VoxCeleb Speaker Recognition Challenge
Speaker verification is the task of confirming that a voice recording belongs to the person who claims to be speaking. In a 2020 paper, researchers at Ghent University describe a network called ECAPA-TDNN that significantly outperformed state-of-the-art systems based on time delay neural networks, or TDNNs, on the VoxCeleb test sets and the 2019 VoxCeleb Speaker Recognition Challenge, two speech benchmarks for speaker verification. The design combines several changes that give the network a wider view of a speech recording.
What did the researchers change?
The authors start from a time delay neural network, a design that processes a sequence of speech frames. The enhancements draw on recent trends in face verification and computer vision. The authors propose that the initial frame layers can be restructured into one-dimensional Res2Net modules with skip connections, which pass information around parts of the network. They add Squeeze-and-Excitation blocks. Those blocks rescale the network's internal channels using the global properties of the recording, which widens the temporal context each frame layer sees. The authors also combine features from different levels of hierarchy. Finally, they update the statistics pooling step, which compresses a variable-length utterance into a fixed-length speaker representation, with channel-dependent frame attention. This allows the network to focus on different subsets of frames while estimating each channel's statistics.
What did the evaluation show?
The paper reports that ECAPA-TDNN significantly outperformed state-of-the-art TDNN-based systems on the VoxCeleb test sets and the 2019 VoxCeleb Speaker Recognition Challenge.
What else should a reader know?
What does channel-dependent frame attention mean?
It means the network weighs different frames differently for each internal channel when it compresses the speech into speaker statistics.
Why add Squeeze-and-Excitation blocks?
These blocks explicitly model relationships among channels and rescale channels using the global properties of the recording. That expands the temporal context of each frame layer.
Did the paper report exact benchmark scores?
The abstract does not list exact error rates or other scores. It reports only that the proposed model significantly outperformed state-of-the-art TDNN-based systems on those tests.
Limits and what's next
The abstract does not state limitations or planned next steps.
Source: Brecht Desplanques et al., “ECAPA-TDNN: Emphasized Channel Attention, Propagation and Aggregation in TDNN Based Speaker Verification,” 2020.