NVV-Locator: From Transcript Tags to Acoustic Boundaries
for Fine-Grained Nonverbal Vocalization Grounding

Nonverbal vocalizations (NVVs) convey affective and interactional cues, yet transcript-level tags leave their acoustic boundaries unspecified. We present NVV-Locator, built on Qwen3-ASR-0.6B, which uses non-autoregressive slot filling to jointly predict lexical timestamps, NVV categories, and event boundaries from audio and a lexical transcript in one forward pass. We harmonize public resources into a 26-category taxonomy and construct timestamp-supervised training data through dual-LLM verification, transcript-guided alignment, energy-based refinement, and decoupled augmentation. We also introduce NVV-TimeBench, an expert-refined benchmark comprising 667 utterances and 1,094 events. NVV-Locator achieves 71.0% micro F1 and 70.2% macro F1 at temporal IoU ≥ 0.5, outperforming the evaluated prompted large audio models. On matched events, it achieves 80.4% macro mIoU and 59.6 ms macro mMAE. External evaluation without target-corpus training further suggests cross-corpus NVV recognition transfer.