NV-Bench: Benchmark for Nonverbal Vocalizations in Text-to-Speech

NV-Bench is the first benchmark for evaluating nonverbal vocalizations in text-to-speech systems, grounded in a functional taxonomy. It was created by CharlesNi and last updated on June 12, 2026. The benchmark aims to provide standardized metrics and reliable ground truth references for this expressive component of speech synthesis.

Use Cases

• Benchmarking the performance of text-to-speech models on generating nonverbal vocalizations based on the proposed functional taxonomy.
• Developing standardized evaluation metrics for expressive speech synthesis based on the benchmark's ground truth references.
• Training models to produce communicative nonverbal acts, such as laughs or sighs, based on the dataset's structured approach.

Strengths

• It is described as the first benchmark for this specific task, filling a gap in standardized evaluation.
• The benchmark is grounded in a functional taxonomy that treats nonverbal vocalizations as communicative acts.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

huggingface

Freshness

Last updated 2026-06-12 11:53:31; freshness should be verified.