NV-Bench: Benchmarking Nonverbal Vocalization Synthesis in Expressive Text-to-Speech Models

Abstract

While recent text-to-speech (TTS) systems increasingly integrate nonverbal vocalizations (NVVs), their evaluation lacks standardized metrics and reliable ground truth references. To bridge this gap, we propose NV-Bench, the first benchmark grounded in a functional taxonomy that treats NVVs as communicative acts rather than acoustic artifacts. NV-Bench comprises 1,651 multilingual, in-the-wild utterances with paired human reference audio, balanced across 14 categories. We introduce a dual-dimensional evaluation protocol:

1. Instruction Alignment — utilizes our proposed Paralinguistic Character Error Rate (PCER) to assess controllability.
2. Acoustic Fidelity — quantifies the distributional gap between synthesized and real speech.

Experimental results demonstrate a strong correlation between our objective metrics and human perception, establishing NV-Bench as a standardized evaluation framework.

Dataset Overview

Subsets

Subset	Language	Description	Label Type
zh_single	Chinese	Single nonverbal event per utterance	Single-label
zh_multi	Chinese	Multiple nonverbal events per utterance	Multi-label
en_single	English	Single nonverbal event per utterance	Single-label
en_multi	English	Multiple nonverbal events per utterance	Multi-label

• Single-label Subset: Strictly balanced — exactly one NVV event per utterance (50 samples per category). Isolates fundamental generation capabilities.
• Multi-label Subset: Challenging utterances with 2+ NVV events — tests robustness under dense paralinguistic conditions with relative balance.

Data Fields

Field	Type	Description
text	string	Target text with inline NVV tags (e.g. [Laughter], [Cough])
prompt_text	string	Prompt text for zero-shot speaker cloning
category	string	NVV category label (one of 14 categories)
type	string	Subset identifier (zh_single, zh_multi, en_single, en_multi)
wav	Audio	Ground-truth reference audio (MP3)
prompt_wav	Audio	Speaker prompt audio for zero-shot cloning (MP3)

Functional Taxonomy

NVVs are organized into three functional levels based on communicative intent:

Name	Description	Categories
Vegetative Sounds	Biological reflexes grounding speech in physical realism	[Cough], [Sigh], [Breathing]
Affect Bursts	Valenced vocalizations conveying emotion or instant reactions	[Laughter], [Surprise-ah], [Surprise-oh], [Dissatisfaction-hnn]
Conversational Grunts	Interaction-management cues — filled pauses and prosodic particles	[Uhm], [Confirmation-en], [Question-ei], [Question-ah], [Question-en], [Question-oh], [Question-huh]

Usage

from datasets import load_dataset

# Load a specific subset
dataset = load_dataset("AnonyData/NV-Bench", "zh_single", split="test")

# Access a sample
sample = dataset[0]
print(sample["text"])          # Target text with NVV tags
print(sample["category"])      # NVV category
print(sample["wav"])           # Ground-truth audio
print(sample["prompt_wav"])    # Speaker prompt audio
print(sample["prompt_text"])   # Prompt text

Evaluation Protocol

Instruction Alignment

Measures whether the model can generate the specified NVV events at the correct positions.

Metric	Description
CER	Character Error Rate
PCER	Paralinguistic Character Error Rate
OCER	Overall Character Error Rate

Acoustic Fidelity

Measures how realistic synthesized speech sounds compared to real recordings.

Metric	Description
FAD / FD / KL	Distribution distance metrics
SIM	Speaker similarity
DNSMOS	Perceptual quality score

Pipeline

1. Data Processing — 565K clips (~1,560 hrs) filtered via Emilia-Pipeline & MiMo-Audio for single-speaker verification.
2. Multi-lingual NVASR — SenseVoice-Small fine-tuned on 6 datasets with unified label taxonomy.
3. Human Verification — 1,651 prompt-GT pairs (7.9 hrs).

Citation

Coming soon

License

This dataset is released under the CC BY NC SA 4.0 license.