<div dir="ltr"><span style="font-family:monospace">(Apologies for cross-posting)</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">Dear colleagues,</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">We are pleased to announce BeTraC, the Beyond Transcription Challenge, an</span><br style="font-family:monospace"><span style="font-family:monospace">IEEE SLT 2026 shared task addressing a foundational question in audio AI:</span><br style="font-family:monospace"><span style="font-family:monospace">Can a model reason over speech without first converting it to text?</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">== Motivation ==</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">Current speech models still struggle to extract meaning directly from</span><br style="font-family:monospace"><span style="font-family:monospace">audio, particularly when the signal includes overlapping speakers,</span><br style="font-family:monospace"><span style="font-family:monospace">ambient sounds, and room acoustics. Clinical note generation from</span><br style="font-family:monospace"><span style="font-family:monospace">doctor–patient conversations is an ideal stress test: the model must</span><br style="font-family:monospace"><span style="font-family:monospace">attend to who said what, filter environmental noise, and produce</span><br style="font-family:monospace"><span style="font-family:monospace">faithful structured output.</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">On the Synth-DoPaCo dataset, end-to-end models hallucinate at alarming</span><br style="font-family:monospace"><span style="font-family:monospace">rates — 99–100% of clinical claims are unsupported by the source audio,</span><br style="font-family:monospace"><span style="font-family:monospace">compared to 21–23% for traditional transcribe-then-summarize pipelines.</span><br style="font-family:monospace"><span style="font-family:monospace">BeTraC aims to close this gap.</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">== Tracks (open-weight models only; no intermediate transcription) ==</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">  * Lightweight (≤ 6B parameters): direct end-to-end audio-to-SOAP.</span><br style="font-family:monospace"><span style="font-family:monospace">    No tool use, no agentic pipelines.</span><br style="font-family:monospace"><span style="font-family:monospace">  * Heavyweight (≤ 36B parameters): tool use and agentic architectures</span><br style="font-family:monospace"><span style="font-family:monospace">    permitted; only the final model generates text from audio.</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">== Synth-DoPaCo dataset ==</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">  * 8,800 synthetic doctor–patient conversations (~1,329 hours)</span><br style="font-family:monospace"><span style="font-family:monospace">  * 66 ambient sound classes, room reverberation, Opus compression</span><br style="font-family:monospace"><span style="font-family:monospace">  * Available on Hugging Face: </span><a href="https://huggingface.co/datasets/betrac" target="_blank" style="font-family:monospace">https://huggingface.co/datasets/betrac</a><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">== Key dates ==</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">  * Apr 2, 2026   — Training + dev data release (live now)</span><br style="font-family:monospace"><span style="font-family:monospace">  * May 4, 2026   — Open-source inclusion proposals deadline</span><br style="font-family:monospace"><span style="font-family:monospace">  * Jun 24, 2026  — System description submission deadline</span><br style="font-family:monospace"><span style="font-family:monospace">  * ~Jul 1, 2026  — Test SOAP notes due (~1 week after test audio release)</span><br style="font-family:monospace"><span style="font-family:monospace">  * Jul 8, 2026   — Challenge paper submission</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">Baselines are posted; team registration is open. If you work on speech,</span><br style="font-family:monospace"><span style="font-family:monospace">audio understanding, or multimodal AI, we would love to have you</span><br style="font-family:monospace"><span style="font-family:monospace">compete.</span><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">Website:    </span><a href="https://betrac.github.io/" target="_blank" style="font-family:monospace">https://betrac.github.io</a><br style="font-family:monospace"><span style="font-family:monospace">Baselines:  </span><a href="https://github.com/betrac/betrac-2026-baseline" target="_blank" style="font-family:monospace">https://github.com/betrac/betrac-2026-baseline</a><br style="font-family:monospace"><span style="font-family:monospace">Contact:    </span><a href="mailto:betrac@googlegroups.com" target="_blank" style="font-family:monospace">betrac@googlegroups.com</a><br style="font-family:monospace"><br style="font-family:monospace"><span style="font-family:monospace">The BeTraC organizers:</span><br style="font-family:monospace"><span style="font-family:monospace">  Andrew Perrault      (The Ohio State University)</span><br style="font-family:monospace"><span style="font-family:monospace">  Jiyun (Amy) Chun     (The Ohio State University)</span><br style="font-family:monospace"><span style="font-family:monospace">  Samuele Cornell      (Carnegie Mellon University)</span><br style="font-family:monospace"><span style="font-family:monospace">  Siddhant Arora       (Carnegie Mellon University)</span><br style="font-family:monospace"><span style="font-family:monospace">  Syed-Amad Hussain    (The Ohio State University / Nationwide Children's Hospital)</span><br style="font-family:monospace"><span style="font-family:monospace">  Thomas Schaaf        (Solventum / CMU LTI)</span><br style="font-family:monospace"><span style="font-family:monospace">  Markus Müller        (Amazon)</span><br style="font-family:monospace"><span style="font-family:monospace">  Leibny Paola Garcia  (Johns Hopkins University, CLSP)</span><br style="font-family:monospace"><span style="font-family:monospace">  Ahmed Hassoon        (Johns Hopkins Bloomberg School of Public Health)</span></div>