Real-Time Voice AI Hears but Does Not Listen
Abstract
Speech conveys information through both words and vocal delivery. We evaluate four leading production realtime voice systems-OpenAI's GPT Realtime 2, Google's Gemini 3.1 Flash Live, and Alibaba's Qwen3.5 Omni Plus and Omni Flash-on tasks where the words and the delivery patterns both convey meaningful information. Across three consequential scenarios, all four systems act on the words rather than the voice. They end calls with crying callers who insist nothing is wrong, approve wire transfers authorized in frightened voices, and enroll callers whose agreement is clearly sarcastic. Surprisingly, this is often not a failure of perception. When asked directly, three of the four systems reliably identify the distress, fear, or sarcasm they later ignore when making decisions. We observe a similar pattern when these realtime voice systems estimate accent and age, as their responses frequently follow the biases of the words rather than the acoustic properties of the speaker. We term this disconnect between perception and action the emotional intelligence gap of voice AI. Prompting systems to explicitly attend to vocal delivery improves performance only partially and inconsistently. Our findings show that current realtime voice AI systems often behave as if speech had been reduced to a transcript, suggesting that they should be used with caution in settings where the tone and emotion of delivery convey important information.
The following analysis is generated by AI and may contain errors. It is separate from the authors’ abstract and is not peer review.
Research summary
This paper provides a critical evaluation of real-time voice AI systems, revealing a significant disconnect between their ability to perceive vocal emotions and their willingness to act on them, thereby highlighting urgent safety concerns for deployment in sensitive domains.
Methodology
The paper employs a rigorous empirical evaluation framework to probe the "emotional intelligence gap" in state-of-the-art real-time voice AI systems. The methodology is sophisticated, moving beyond simple classification benchmarks to test consequential decision-making in multi-turn dialogues. By synthesizing speech with ElevenLabs to create controlled conflicts between lexical content (words) and non-lexical content (tone/emotion/accent/age), the authors isolate the model's reliance on text versus audio. The use of single-turn diagnostics to separate perception from action is a critical methodological strength, allowing the authors to demonstrate that models can "hear" distress but choose not to act on it. The inclusion of instruction-tuning variants (attend/override) further deepens the analysis of model alignment and robustness.
Results & evaluation
The experimental design is robust and well-executed. The authors test four leading production systems (OpenAI, Google, Alibaba) across three high-stakes scenarios (emergency welfare, fraud prevention, volunteer recruitment). The results are consistent and striking: all systems prioritized lexical content over vocal cues, leading to potentially dangerous outcomes (e.g., ending calls with distressed callers). The human listener validation confirms that the synthesized stimuli successfully conveyed the intended emotional and demographic cues, ensuring the validity of the conflict conditions. The statistical reporting (5 runs per condition) provides a solid basis for the claims, although the small sample size per condition is a minor limitation. The finding that prompting only partially mitigates the issue is significant and well-supported by the data.
Reproducibility
The paper provides detailed descriptions of the experimental setup, including the specific model identifiers, the use of ElevenLabs for synthesis, and the prompts used for both the agent and the automated caller (GPT-5.5). The appendix contains the full scripts and prompts, which greatly aids reproducibility. However, reliance on proprietary APIs (OpenAI, Google, Alibaba) and a third-party TTS service (ElevenLabs) means that exact replication of results may be subject to API updates or changes in model versions. The use of synthesized speech rather than real human recordings is a trade-off; while it ensures control, it may not fully capture the nuance of human vocalizations.
Limitations
The primary limitation is the use of synthesized speech, which, while high-quality, may lack the subtle imperfections and variations of human speech that models might rely on or be trained to ignore. Additionally, the scenarios, while realistic, are simplified simulations. The reliance on GPT-5.5 to drive the "caller" persona introduces another layer of complexity and potential bias, although it is used consistently across conditions. The study focuses on English-language interactions, limiting generalizability to other languages where prosodic cues may differ. Finally, the sample size for each condition (5 runs) is small, though sufficient to demonstrate the trend.
Broader impact
This paper has significant implications for the safe deployment of voice AI in critical sectors such as healthcare, finance, and emergency services. It highlights a fundamental safety risk: models may fail to detect distress or coercion if the verbal content is neutral or positive. The term "emotional intelligence gap" provides a useful framework for future research and evaluation. The findings suggest that current real-time voice models are not yet safe for unsupervised use in high-stakes scenarios where non-lexical cues are decisive. This work calls for architectural changes or more robust alignment techniques to ensure models attend to the full spectrum of vocal information. This paper provides a critical evaluation of real-time voice AI systems, revealing a significant disconnect between their ability to perceive vocal emotions and their willingness to act on them, thereby highlighting urgent safety concerns for deployment in sensitive domains.
Author of this paper? Send a correction.