NV-Bench: 1,651 Samples for Nonverbal Vocalization Synthesis Benchmarking

NV-Bench is a benchmark dataset for evaluating nonverbal vocalization synthesis in text-to-speech models, created by AnonyData and last updated on March 1, 2026. It comprises 1,651 samples grounded in a functional taxonomy that treats nonverbal vocalizations as communicative acts. The dataset is hosted on Hugging Face and aims to provide standardized metrics and reliable ground truth references for this expressive TTS subfield.

Use Cases

• Benchmarking the performance of TTS models on nonverbal vocalization synthesis based on the provided taxonomy.
• Training expressive TTS models to generate specific communicative acts like laughter or hesitation.
• Conducting comparative analysis of different TTS architectures using the standardized evaluation framework.
• Studying the relationship between textual prompts and synthesized nonverbal vocal sounds.

Strengths

• Contains 1,651 samples, providing a substantial basis for benchmarking.
• Based on a defined functional taxonomy for nonverbal vocalizations as communicative acts.
• Explicitly designed to address a lack of standardized evaluation in expressive TTS.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is known but total dataset size and file formats are unknown, which may limit suitability assessment.

Provenance

Source

Hugging Face, uploaded by author AnonyData.

Collection Method

Created as a benchmark dataset; specific collection method is not detailed in the provided metadata.

Time Range

The dataset's temporal coverage is not specified.

Freshness

Last updated 2026-03-01 07:02:19; freshness should be verified.

Geography

The dataset's geographic coverage is not specified.