Work

Deepcounsel: A Multi-Agent Framework for Simulating Complex Courtroom Audio Environments

Year: 2026
Type: conference-paper
Abstract: The scarcity of high-quality, fully labeled audio data is a massive bottleneck for developing legal AI tech because real-world legal proceedings are tightly guarded by privacy laws, sealed records, and strict nondisclosure bounds.So there is a requirement for developing synthetic courtroom audio data.We propose a multi-agent prompting framework that utilizes Gemini 3.1 and GPT-4o native support for English language and its context-aware pacing to produce distinct legal courtroom scenarios which generates a 600 minute (10 hours) synthetic audio data named Deep-Counsel .We subjected it to human evaluation across metrics including Naturalness, Professionalism, and Authenticity.In a legal framework, GPT-4o excels at analytical, client-facing tasks, leading in Legal Logic (4.82/5.0)and Emotional Prosody (4.27/5.0).It also dominates in technical audio clarity (0.015 Fréchet Audio Distance) and Ambient Integration, alongside superior transcription accuracy (0.021 Word Error Rate and 0.884 BertScore).However, both models suffer from legal hallucinations in simulations.These include procedural drift, legal confabulation, and temporal errors. 1
Author: Aniket Deroy
Language: English
Open Access status: gold
OpenAlex ID: W7127053166
FWCI:
Cites: 39
Cited by:
Related to: 10