Institution profile

Brno University of Technology

Academic institutioneurope · cz
Official website
Research library199linked papers
Opportunities0open roles
Selected work

Representative Papers

Comparing Specialised Small and General Large Language Models on Text Classification: 100 Labelled Samples to Achieve Break-Even Performance

Feb 20, 2024

This work investigates the minimal labeled data requirements for specialized small-scale models to surpass general-purpose large language models (LLMs) under few-shot settings (10–1,000 samples). Using eight text classification tasks, we systematically evaluate performance inflection points of seven models—including LLaMA and BERT—across fine-tuning, instruction tuning, prompt engineering, and in-context learning. We find that, on average, multi-class tasks require ≤100 samples for small models to match or exceed LLM performance, whereas binary classification may demand up to 5,000 samples. We propose a robust sample-requirement estimation framework that explicitly accounts for performance variance, improving estimation accuracy by 100–200% (up to 1,500% in extreme cases). Our core contribution is the quantitative characterization of how task properties—particularly class count and output variance—affect small-model data efficiency. This yields reproducible, task-aware guidelines for determining optimal annotation budgets in low-resource NLP scenarios.

7 citationsRead paper

Text-only adaptation in LLM-based ASR through text denoising

Jan 28, 2026

This work addresses the performance degradation commonly observed in large language model (LLM)-based speech recognition systems when adapting solely with in-domain text, a process that often disrupts the alignment between speech and text modalities. To mitigate this issue, the authors propose a lightweight text-denoising adaptation approach that reformulates the audio projection task as a text denoising problem. By training the LLM to reconstruct clean transcripts from noisy textual inputs, the method achieves effective domain adaptation without modifying the model architecture or introducing additional parameters. This strategy preserves cross-modal alignment while significantly improving recognition accuracy, yielding up to a 22.1% relative reduction in word error rate on two benchmark datasets—substantially outperforming current state-of-the-art text-only adaptation techniques.

1 citationsRead paper

Reducing Prompt Sensitivity in LLM-based Speech Recognition Through Learnable Projection

Jan 28, 2026

This work addresses the instability and high sensitivity to prompt selection in existing large language model (LLM)-based speech recognition approaches that rely on fixed, manually crafted prompts. To overcome this limitation, the authors propose a model-agnostic, learnable prompt projection module that adaptively maps prompt embeddings into more effective regions of the LLM’s input space, without modifying the underlying LLM architecture. This approach significantly reduces prompt sensitivity and enhances recognition robustness and consistency. Experimental results across four benchmark datasets demonstrate that the proposed method not only consistently outperforms the best handcrafted prompts but also substantially mitigates performance variance, yielding more reliable and stable recognition outcomes.

1 citationsRead paper

SE-DiCoW: Self-Enrolled Diarization-Conditioned Whisper

Jan 27, 2026

This work addresses the challenges faced by existing cross-domain multi-speaker automatic speech recognition (ASR) systems in accurately disentangling speakers in overlapping speech and their limited generalization capability. The authors propose a self-registration mechanism that leverages speaker diarization to identify the most active utterance segment of a target speaker and uses it as a fixed conditioning embedding injected into the cross-attention layers of the Whisper encoder. This approach effectively mitigates the speaker–time–output (STNO) mask ambiguity inherent in multi-talker ASR. Combined with refined data segmentation, model initialization, and augmentation strategies, the method substantially enhances the model’s ability to separate overlapping speakers and improves cross-domain generalization. On the EMMA MT-ASR benchmark, the proposed system achieves a 52.4% relative reduction in macro-averaged tcpWER compared to DiCoW.

1 citationsRead paper
Recent publications

Latest Papers