SpeechParaling-Bench: A Comprehensive Benchmark for Paralinguistic-Aware Speech Generation

Paper Overview

Field: NLP Authors: Ruohan Liu, Shukang Yin, Tao Wang Released: 2026-04-22 arXiv: 2604.20842

Introduction

Paralinguistic cues—the non-verbal aspects of speech such as tone, emotion, and prosody—are essential for natural human-computer interaction. However, evaluating these capabilities in Large Audio-Language Models (LALMs) has been limited by coarse feature coverage and the inherent subjectivity of assessment methods.

SpeechParaling-Bench

To address these challenges, the authors introduce SpeechParaling-Bench, a comprehensive benchmark for paralinguistic-aware speech generation with the following features:

• Expanded feature coverage: Grows from fewer than 50 to over 100 fine-grained paralinguistic features.
• Bilingual queries: More than 1,000 English-Chinese parallel speech queries.
• Three progressively challenging tasks:
1. Fine-grained control 2. Intra-utterance variation 3. Context-aware adaptation

Pairwise Comparison Evaluation Pipeline

To enable reliable evaluation, the benchmark uses a pairwise comparison pipeline in which candidate responses are evaluated against a fixed baseline. By shifting the evaluation framework from absolute scoring to relative preference, this approach:

• Effectively mitigates subjectivity
• Achieves more stable and scalable evaluation
• Avoids the need for costly human annotation

Key Findings

Extensive experiments reveal significant limitations in current LALMs:

• Even leading proprietary models struggle to comprehensively control static paralinguistic features and dynamic modulation.
• In situated dialogue, errors caused by misinterpreting paralinguistic cues account for up to 43.3% of failures.

These findings highlight the urgent need for more robust paralinguistic modeling to advance speech assistants toward human-aligned behavior.

---

*Auto-collected on 2026-04-24*