This paper describes TalTech's submissions to the Beyond Transcription Challenge (BeTraC), which requires generating SOAP notes directly from long doctor-patient conversation recordings, without intermediate transcription. After screening open-weight speech LLMs for long-audio robustness, we adapted Voxtral Mini (lightweight track) and Voxtral Small (heavyweight track) with LoRA supervised fine-tuning followed by DAPO reinforcement learning that uses the challenge metric, Open Medical Concept F1, as its reward. Our systems ranked first in both tracks, and an independent LLM-as-a-judge evaluation showed the lowest hallucination rate among all submissions, indicating that reinforcement learning against a concept-matching metric need not compromise factual reliability. We also find that fine-tuning on text transcripts transfers well to speech input and appears to improve robustness on out-of-domain real recordings.
Audio-capable foundation models enable end-to-end spoken interaction, but they also introduce safety risks beyond transcript content. It remains unclear how much jailbreak capability can arise from matched-text variation in speech delivery rather than from lexical rewriting or broader style transfer. We study this question by holding transcript content fixed and varying six speech-delivery presets whose acoustic attributes may co-vary. We present PJ-Break, a black-box evaluation protocol with presets targeting arousal, authority, and speaking rate, together with AdvAudio-Prosody, a 600-sample benchmark with acoustically verified attributes. On the exact post-QC Qwen2-Audio panel, the Q=1 Panic (38/95), Anger (35/95), and Fast (32/95) presets are all well above Neutral (4/95). The fixed six-query pool covers 44/95 Qwen2-Audio seeds and 15/95 GPT-4o seeds and exceeds a matched-budget StyleBreak reimplementation (27/95) on Qwen2-Audio. A same-voice pool excluding the confounded Commanding condition still reaches 40/95, and a retained-panel ablation shows emotional-delivery audio alone (44/95) is far more effective than emotional text alone (11/95). Exploratory surrogate diagnostics and pilot mitigation observations are secondary, non-core analyses. Overall, matched-text speech delivery should be treated as a first-class factor in Audio LLM safety evaluation
High-quality, large-scale conversational datasets are scarce, making it difficult to train on-device language models (OD-LLMs, ∼ 1B parameters) as effective assistants. We introduce CoSy ( Co nversational Sy nthesis), a novel framework for generating diverse, steerable, multi-turn conversations at scale. CoSy combines three key mechanisms: (1) conversational graphs that ensure natural dialogue flow, (2) turn-based prompt augmen-tations for diversity, and (3) explicit linguistic phenomena for coherence. We evaluate CoSy on conversational grounded reasoning tasks (i.e., answering questions based on contextual information), a core on-device use case. Our on-device sized models trained on CoSy-synthesized data achieve competitive performance with human-annotated baselines and outperform instruction-tuned models of up to 70B parameters in zero-shot settings.
Patrick Huber, Arash Einolghozati, Rylan Conway et al.· IEEE Games Entertainment Med...· 0 citations
This paper describes our self-designed system for Task 1 of the MLC-SLM 2026 Challenge for multilingual two-speaker conversational speech. The system combines a modular speaker diarization front end with a challenge-adapted Qwen3-ASR-1.7B recognizer. The diarization front end performs voice activity detection, subsegment generation, CAMPPlus speaker embedding extraction, two-speaker spectral clustering, and RTTM-based audio segmentation. The resulting speaker-attributed segments are grouped by language or region and decoded by the adapted ASR model. For ASR adaptation, we first perform supervised full fine-tuning on the official training data, then apply LoRA fine-tuning with synthetic speech generated by a three-pipeline TTS-based synthetic speech augmentation framework, and finally refine the model using GRPO reinforcement learning with rewards based on WER/CER and penalties for hallucination, repetition, and length deviation. On the official development set, the full system achieves an average tcpMER of 23.70, reducing the error rate by 6.83 absolute points relative to the released Qwen-ASR-1.7B performance. On the final evaluation set, the system achieves an average tcpMER of 17.97. Ablation results show that supervised fine-tuning provides the largest gain, while synthetic-speech LoRA adaptation and reinforcement learning further improve robustness.
DCASE~2026 Task~5 introduces Audio-Dependent Question Answering (ADQA), which tests whether large audio-language models answer from the audio rather than from textual priors. An Audio-Dependency Filtering (ADF) pipeline combines silent-audio probing, per-option perplexity, a large language model (LLM) commonsense check, and human review to remove items solvable from text alone. The 3000 items that pass form the ADQA-Bench evaluation set, spanning music, speech, and environmental audio. The inaugural edition draws 14 teams and 36 submissions across two tracks defined by total parameter count (up to 100B and under 10B). A Chung-Ang University ensemble of MOSS-Audio-8B-Thinking and Qwen3-Omni-30B reaches the top overall accuracy at \pct{58.33}, and a MOSS-only configuration from the same team leads the sub-10B track at \pct{57.30}. Across the 30 submissions with a comparable development score, evaluation accuracy falls by 11.91 percentage points (pp) on average (median 10.91\,pp) on the hidden evaluation split, which is designed to be harder than the development split. The most common building blocks are: the MOSS-Audio-8B-Thinking backbone (13 of 36 submissions), Low-Rank Adaptation (LoRA) fine-tuning on AudioMCQ-StrongAC, and preference or reinforcement-learning objectives -- Group Relative Policy Optimization (GRPO) in five teams, Group reward-Decoupled Normalization Policy Optimization (GDPO) in two. At test time, prompt engineering is near-universal, and majority or choice-permutation voting is common. Every system misses the same set of 233 evaluation items.
Haolin He, Renhe Sun, Zheqi Dai et al.· 0 citations
This paper presents an automated pipeline for transcription and summarization of video conferences held on the Jitsi Meet platform. The proposed system integrates the Jigasi SIP gateway as a transparent in-call participant that captures perspeaker RTP audio streams in real time. Each stream is transcribed using the Vosk large-vocabulary speech recognition model (vosk-model-en-us-0.22), which delivers word-level timestamps and speaker attribution without relying on cloud services. A monitoring script continuously watches the Windows Downloads folder for new WebM recordings produced by Jitsi; upon detection the file is moved to a dedicated recordings directory and queued for processing. After transcription, the resulting structured text is forwarded to a Claude-based summarization engine that extracts the agenda, key decisions, action items, and participant contributions, saving the final report as a plain-text file. Experiments on 20 real-world meeting recordings demonstrate a word error rate of 14.3% for English speech and an average end-to-end latency of 38 seconds per hour of audio on a CPU-only server. The system requires no proprietary cloud API keys and is deployable on-premise via Docker Compose, making it suitable for organizations with strict data-privacy requirements.
G. Amirkhanova, L. Bektemir, Shyrailym Adilkyzy et al.· International Conference on...· 0 citations
Self-repairs, in-utterance revisions in which a speaker abandons and reformulates their speech, are a promising interpretable marker for speech-based cognitive screening. Detecting them automatically is difficult because a self-repair is defined by its relationship to surrounding speech rather than by fixed lexical cues. On the DementiaBank ADReSS corpus, we compared the capability of generative LLMs under a five-condition prompt ablation against a fine-tuned DistilBERT token classifier at detecting self-repairs. GPT-5 performed best (test F1 = 0.73) and was largely insensitive to prompt design, whereas the LLaMA (open-weight alternative) was both weaker and far more prompt-sensitive (test F1 = 0.47). DistilBERT, nearly 100 times smaller, matched the open-weight LLM at a fraction of the computational cost. These results suggest that a locally deployable encoder, given sufficient in-domain annotation, is a more plausible route to clinical self-repair detection than scaling model size or prompt complexity.
R. Wu, S. Pugh, K. O'Connor et al.· medRxiv· 0 citations