Researchers have created a new benchmark called TELEVAL to evaluate Spoken Language Models (SLMs) in Chinese interactive scenarios. Unlike existing benchmarks that focus on semantic correctness, TELEVAL assesses two key aspects: the accuracy of SLMs' responses and their ability to produce natural and appropriate interactions based on auditory cues. Experiments showed that current SLMs struggle with acoustic variability and interactional settings, often producing descriptions of audio signals instead of interactive responses. --- Why it matters: This matters because it highlights the limitations of current Spoken Language Models in handling real-world spoken interactions, which are crucial for applications like voice assistants and customer service chatbots. By creating a more comprehensive benchmark, researchers can better understand these models' weaknesses and develop more effective solutions. Source: https://arxiv.org/abs/2507.18061