Peichun Hua

Accepted · PRICAI 2026 · Regular paper, oral presentation

Enroll Once, Identify Everywhere: Few-Shot Cross-Scenario Visual Speaker Recognition via Event-driven SlowFast Networks

Junguang Yao, Peichun Hua, Yue Zheng

Read paper (PDF) Code

Overview

This paper studies visual speaker recognition from lip motion captured by event cameras. Its few-shot enrollment protocol separates training and test identities and recognizes new speakers from three enrollment samples, including changes in illumination and viewpoint.

Method

A SlowFast network processes two-channel polarity-aware event count volumes. Complementary temporal pathways and fast-to-slow lateral fusion capture lip articulation. After training on known identities, a frozen feature extractor supports nearest-neighbor identification without retraining for new users.

Event-camera speaker-recognition pipeline: event streams form count volumes, a SlowFast network extracts embeddings, and probe embeddings are matched to an enrollment gallery.
From lip-motion events to speaker embeddings: a frozen SlowFast encoder supports enrollment and nearest-neighbor identification. Figure 1 in the paper

Evaluation

On DVSpeaker, Rank-1 accuracy is 99.18% in matched frontal-bright conditions, 87.25% under low-light shift, 74.95% at 45°, and 50.20% at 90°. Ablations examine the contribution of temporal pathways and lateral fusion.

Speaker identification across enrollment and probe conditions
Enrollment ↓ / Probe →Frontal, bright45°90°Frontal, low light
Frontal, bright99.18%74.95%50.20%87.25%
45°77.45%97.32%58.75%70.05%
90°50.80%53.10%98.04%35.75%
Frontal, low light90.90%65.90%39.00%96.34%
Rank-1 accuracy on DVSpeaker with 30 training identities and 20 disjoint test identities, using three enrollment samples per test identity. Rows specify gallery conditions; columns specify probe conditions. Diagonal entries use matched scenarios. Table 2 in the paper

Citation

@inproceedings{yao2026enroll,
  title={Enroll Once, Identify Everywhere: Few-Shot Cross-Scenario Visual Speaker Recognition via Event-driven SlowFast Networks},
  author={Junguang Yao and Peichun Hua and Yue Zheng},
  booktitle={The 23rd Pacific Rim International Conference on Artificial Intelligence (PRICAI)},
  year={2026},
  organization={Springer}
}