Sim911: Towards Effective and Equitable 9-1-1 Dispatcher Training with an LLM-Enabled Simulation

Zirong Chen, Elizabeth Chason, Noah Mladenovski, Erin Wilson, Kristin Mullen, Stephen Martini, Meiyi Ma

Abstract

Emergency response services are vital for enhancing public safety by safeguarding the environment, property, and human lives. As frontline members of these services, 9-1-1 dispatchers have a direct impact on response times and the overall effectiveness of emergency operations. However, traditional dispatcher training methods, which rely on role-playing by experienced personnel, are labor-intensive, time-consuming, and often neglect the specific needs of underserved communities. To address these challenges, we introduce Sim911, the first training simulation for 9-1-1 dispatchers powered by Large Language Models (LLMs). Sim911 enhances training through three key technical innovations: (1) knowledge construction, which utilizes archived 9-1-1 call data to generate simulations that closely mirror real-world scenarios; (2) context-aware controlled generation, which employs dynamic prompts and vector bases to ensure that LLM behavior aligns with training objectives; and (3) validation with looped correction, which filters out low-quality responses and refines the system performance.

Introduction

Emergency response services in the U.S. are currently facing a severe staffing crisis, with approximately 25,000 unfilled dispatcher positions nationwide. A primary bottleneck in addressing this shortage is the traditional training process. Currently, training relies heavily on manual role-playing, where experienced dispatchers must divert from active duty to simulate callers for trainees. This method is not only labor-intensive and inconsistent but also frequently fails to represent the full spectrum of real-world scenarios, particularly those involving vulnerable populations such as non-native English speakers, the unhoused, or individuals in mental health crises.

Sim911 addresses these gaps by providing a scalable, high-fidelity, and equitable training environment. By automating the role of the caller using Large Language Models (LLMs), Sim911 ensures that trainees are exposed to diverse and realistic scenarios without taxing the resources of senior personnel.

Methodology

The Sim911 framework is built upon a three-pillar technical architecture designed to transform a general-purpose LLM into a specialized emergency simulator.

Overview of the Sim911 framework, illustrating the three core pillars: Knowledge Construction, Context-aware Controlled Generation, and Validation with Looped Correction.

1. Knowledge Construction (KC)

The system leverages a database of over 11,000 archived 9-1-1 calls from the Metro Nashville Department of Emergency Communications (MNDEC). From these, it extracts "Incident Specifications" (e.g., fire accidents, firearm incidents) and "Caller Images" (e.g., age group, emotional state, and vulnerability tags). This ensures the simulation is grounded in real-world patterns.

2. Context-aware Controlled Generation (CaCG)

To maintain realism and prevent hallucinations, Sim911 employs several prompting strategies:

  • Retrieval-Augmented Generation (RAG): Integrates factual data such as local address lists and official protocols.
  • Chain-of-Thought (CoT): Guides the model to plan responses according to specific training objectives.
  • Few-Shot Prompting (FSP): Uses historical interaction examples to mirror realistic speech patterns and tones.

3. Validation with Looped Correction (VLC)

A real-time "co-pilot" validates the LLM's output against formatting requirements and factual accuracy. For instance, it uses the Google Maps API to verify that any addresses mentioned by the simulated caller are geographically valid. If a response fails validation, the system automatically triggers a correction loop to regenerate the output.

Experimental Results

Sim911 was evaluated using real-world data and CAD logs from MNDEC. The results demonstrate that the system significantly outperforms standard LLM baselines in both realism and authenticity.

Performance comparison and ablation study of Sim911 across realism and authenticity metrics.

Key findings include:

  • Superior Realism: Sim911 achieved a Perplexity (PPL) of 11.07, compared to 61.99 for a non-optimized GPT-4o baseline.
  • High Authenticity: The system reached a 99.19% success rate in providing accurate location data, a critical requirement for dispatcher training.
  • Operational Efficiency: In Nashville deployments, Sim911 saved over 26 hours of active simulation time for senior dispatchers, acting as a significant "force multiplier."
  • User Acceptance: 90% of MNDEC personnel surveyed found the system to be as effective as or superior to traditional human-led training.

Discussion and Future Work

While Sim911 represents a major step forward, challenges remain regarding the persistence of subtle hallucinations and the need for real-time knowledge updates as city infrastructure changes. Future research will explore fine-tuning open-source models (e.g., Llama) to facilitate local, private hosting within secure dispatch environments, further enhancing data privacy and system control.

Summary updated

Abstract

Emergency response services are vital for enhancing public safety by safeguarding the environment, property, and human lives. As frontline members of these services, 9-1-1 dispatchers have a direct impact on response times and the overall effectiveness of emergency operations. However, traditional dispatcher training methods, which rely on role-playing by experienced personnel, are labor-intensive, time-consuming, and often neglect the specific needs of underserved communities. To address these challenges, we introduce Sim911, the first training simulation for 9-1-1 dispatchers powered by Large Language Models (LLMs). Sim911 enhances training through three key technical innovations: (1) knowledge construction, which utilizes archived 9-1-1 call data to generate simulations that closely mirror real-world scenarios; (2) context-aware controlled generation, which employs dynamic prompts and vector bases to ensure that LLM behavior aligns with training objectives; and (3) validation with looped correction, which filters out low-quality responses and refines the system performance.
BibTeX
@inproceedings{chen2025sim,
  author = {Zirong Chen and Elizabeth Chason and Noah Mladenovski and Erin Wilson and Kristin Mullen and Stephen Martini and Meiyi Ma},
  title = {{Sim911: Towards Effective and Equitable 9-1-1 Dispatcher Training with an LLM-Enabled Simulation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i27.35006},
}
Figure