OmniACBench: A Benchmark for Evaluating Context-Grounded Acoustic Control in Omni-Modal Models

TL;DR

This paper introduces OmniACBench, a comprehensive benchmark for evaluating how well omni-modal models can generate context-appropriate speech based on multimodal inputs, highlighting current limitations and failure modes.

Contribution

The paper presents OmniACBench, a novel benchmark for assessing acoustic control in omni-modal models, and provides detailed analysis of model limitations and failure modes.

Findings

Models struggle with integrating multimodal context for speech

Current models perform poorly on acoustic control tasks

Identified key failure modes in speech generation

Abstract

Most testbeds for omni-modal models assess multimodal understanding via textual outputs, leaving it unclear whether these models can properly speak their answers. To study this, we introduce OmniACBench, a benchmark for evaluating context-grounded acoustic control in omni-modal models. Given a spoken instruction, a text script, and an image, a model must read the script aloud with an appropriate tone and manner. OmniACBench comprises 3,559 verified instances covering six acoustic features: speech rate, phonation, pronunciation, emotion, global accent, and timbre. Extensive experiments on eight models reveal their limitations in the proposed setting, despite their strong performance on prior textual-output evaluations. Our analyses show that the main bottleneck lies not in processing individual modalities, but in integrating multimodal context for faithful speech generation. Moreover, we identify three common failure modes-weak direct control, failed implicit inference, and failed multimodal grounding-providing insights for developing models that can verbalize responses effectively.