Streaming Speaker Verification with Confidence-Based Stopping for Short Utterances

Conventional speaker verification systems operate on pre-segmented utterances of fixed duration, limiting their applicability in real-time and interactive scenarios. We introduce a streaming speaker verification framework for short utterances based on a lightweight causal recurrent module, termed the Streaming Adapter. This module aggregates incremental speaker embeddings from a frozen pretrained extractor and outputs a calibrated, time-evolving probability of speaker match every 100 ms. The resulting score trajectory enables early termination for acoustically clear trials while allowing extended analysis for ambiguous cases, thereby dynamically balancing latency and reliability. Experiments on the VoxCeleb dataset demonstrate that the proposed approach improves verification performance in the short-utterance regime and enables reliable early decision making compared to conventional cosine scoring. Ablation studies confirm that the gains generalize across different embedding extractors and are not explained by additional extractor training alone.