Speech-to-SOAP: End-to-End Summarization of Medical Dialogues: KIT@BeTraC 2026
Abstract Overview
This paper presents KIT's BeTraC 2026 lightweight-track system for generating SOAP clinical notes directly from medical dialogue audio. The core approach adapts the Qwen2.5-Omni-3B speech-language model with LoRA and trains it on a unified collection of heterogeneous medical dialogue datasets converted into Audio→SOAP, Transcript→SOAP, and related formats. A central component is a scalable augmentation pipeline that synthesizes speech for text-only datasets and creates SOAP-style supervision automatically when reference notes are unavailable. The study evaluates prompt design, speech-only versus joint speech-text training, multi-stage adaptation, duration filtering, audio cleaning, chain-of-thought supervision, and checkpoint averaging under the shared-task evaluation protocol.
Novelty
The paper's main novelty is a scalable data augmentation pipeline that unifies diverse medical dialogue resources by combining synthetic speech generation with automatically generated SOAP supervision for end-to-end speech-to-SOAP training. It is also distinctive in systematically analyzing prompt placement, intermediate adaptation strategies, and checkpoint averaging for direct clinical note generation from speech.
Results
On the development set, joint audio-text supervision improved Concept-F1 over audio-only training, while intermediate adaptation strategies provided substantial gains over baseline direct ASR initialization. The best development configuration was a merged submission obtained by checkpoint averaging, achieving 0.4986 Concept-F1, 0.3537 ROUGE-2, and 0.2417 ROUGE-3. In the official evaluation, this primary merged model consistently outperformed the contrastive submission across the DoPaCo, Mock Dialogue, and Realistic test sets, demonstrating improved robustness under domain shift.
Key Points
- The system trains an end-to-end speech-to-SOAP model by unifying multiple medical dialogue datasets through synthetic audio generation and automatic SOAP note supervision.
- Experimental ablations show that detailed task guidance functions better as instruction prompts than as system prompts, and that transcript supervision primarily improves clinical concept extraction rather than lexical overlap.
- Checkpoint averaging of diversely trained models produced the strongest overall system and outperformed the contrastive submission across all official test sets.