Vassilis Katsouros

Leon Voukoutis, Dimitris Roussis, Georgios Paraskevopoulos et al.

We describe the development and capabilities of Meltemi 7B, the first open Large Language Model for the Greek language. Meltemi 7B has 7 billion parameters and is trained on a 40 billion token Greek corpus. For the development of Meltemi 7B, we adapt Mistral, by continuous pretraining on the Greek Corpus. Meltemi 7B contains up-to-date information up to September 2023. Furthermore, we have translated and curated a Greek instruction corpus, which has been used for the instruction-tuning of a chat model, named Meltemi 7B Instruct. Special care has been given to the alignment and the removal of toxic content for the Meltemi 7B Instruct. The developed models are evaluated on a broad set of collected evaluation corpora, and examples of prompts and responses are presented. Both Meltemi 7B and Meltemi 7B Instruct are available at https://huggingface.co/ilsp under the Apache 2.0 license.

Georgios Paraskevopoulos, Theodoros Kouzelis, Georgios Rouvalis et al.

Modern speech recognition systems exhibits rapid performance degradation under domain shift. This issue is especially prevalent in data-scarce settings, such as low-resource languages, where diversity of training data is limited. In this work we propose M2DS2, a simple and sample-efficient finetuning strategy for large pretrained speech models, based on mixed source and target domain self-supervision. We find that including source domain self-supervision stabilizes training and avoids mode collapse of the latent representations. For evaluation, we collect HParl, a $120$ hour speech corpus for Greek, consisting of plenary sessions in the Greek Parliament. We merge HParl with two popular Greek corpora to create GREC-MD, a test-bed for multi-domain evaluation of Greek ASR systems. In our experiments we find that, while other Unsupervised Domain Adaptation baselines fail in this resource-constrained environment, M2DS2 yields significant improvements for cross-domain adaptation, even when a only a few hours of in-domain audio are available. When we relax the problem in a weakly supervised setting, we find that independent adaptation for audio using M2DS2 and language using simple LM augmentation techniques is particularly effective, yielding word error rates comparable to the fully supervised baselines.

Manos Plitsis, Theodoros Kouzelis, Georgios Paraskevopoulos et al.

In this work, we investigate the personalization of text-to-music diffusion models in a few-shot setting. Motivated by recent advances in the computer vision domain, we are the first to explore the combination of pre-trained text-to-audio diffusers with two established personalization methods. We experiment with the effect of audio-specific data augmentation on the overall system performance and assess different training strategies. For evaluation, we construct a novel dataset with prompts and music clips. We consider both embedding-based and music-specific metrics for quantitative evaluation, as well as a user study for qualitative evaluation. Our analysis shows that similarity metrics are in accordance with user preferences and that current personalization approaches tend to learn rhythmic music constructs more easily than melody. The code, dataset, and example material of this study are open to the research community.

Gerasimos Chatzoudis, Manos Plitsis, Spyridoula Stamouli et al.

Aphasia is a common speech and language disorder, typically caused by a brain injury or a stroke, that affects millions of people worldwide. Detecting and assessing Aphasia in patients is a difficult, time-consuming process, and numerous attempts to automate it have been made, the most successful using machine learning models trained on aphasic speech data. Like in many medical applications, aphasic speech data is scarce and the problem is exacerbated in so-called "low resource" languages, which are, for this task, most languages excluding English. We attempt to leverage available data in English and achieve zero-shot aphasia detection in low-resource languages such as Greek and French, by using language-agnostic linguistic features. Current cross-lingual aphasia detection approaches rely on manually extracted transcripts. We propose an end-to-end pipeline using pre-trained Automatic Speech Recognition (ASR) models that share cross-lingual speech representations and are fine-tuned for our desired low-resource languages. To further boost our ASR model's performance, we also combine it with a language model. We show that our ASR-based end-to-end pipeline offers comparable results to previous setups using human-annotated transcripts.

Efthymios Georgiou, Kosmas Kritsis, Georgios Paraskevopoulos et al.

Recent deep learning Text-to-Speech (TTS) systems have achieved impressive performance by generating speech close to human parity. However, they suffer from training stability issues as well as incorrect alignment of the intermediate acoustic representation with the input text sequence. In this work, we introduce Regotron, a regularized version of Tacotron2 which aims to alleviate the training issues and at the same time produce monotonic alignments. Our method augments the vanilla Tacotron2 objective function with an additional term, which penalizes non-monotonic alignments in the location-sensitive attention mechanism. By properly adjusting this regularization term we show that the loss curves become smoother, and at the same time Regotron consistently produces monotonic alignments in unseen examples even at an early stage (13\% of the total number of epochs) of its training process, whereas the fully converged Tacotron2 fails to do so. Moreover, our proposed regularization method has no additional computational overhead, while reducing common TTS mistakes and achieving slighlty improved speech naturalness according to subjective mean opinion scores (MOS) collected from 50 evaluators.

Theodoros Kouzelis, Vassilis Katsouros

In recent years, datasets of paired audio and captions have enabled remarkable success in automatically generating descriptions for audio clips, namely Automated Audio Captioning (AAC). However, it is labor-intensive and time-consuming to collect a sufficient number of paired audio and captions. Motivated by the recent advances in Contrastive Language-Audio Pretraining (CLAP), we propose a weakly-supervised approach to train an AAC model assuming only text data and a pre-trained CLAP model, alleviating the need for paired target data. Our approach leverages the similarity between audio and text embeddings in CLAP. During training, we learn to reconstruct the text from the CLAP text embedding, and during inference, we decode using the audio embeddings. To mitigate the modality gap between the audio and text embeddings we employ strategies to bridge the gap during training and inference stages. We evaluate our proposed method on Clotho and AudioCaps datasets demonstrating its ability to achieve a relative performance of up to ~$83\%$ compared to fully supervised approaches trained with paired target data.

Manos Plitsis, Giorgos Bouritsas, Vassilis Katsouros et al.

Text-to-image (TTI) diffusion models have achieved remarkable visual quality, yet they have been repeatedly shown to exhibit social biases across sensitive attributes such as gender, race and age. To mitigate these biases, existing approaches frequently depend on curated prompt datasets - either manually constructed or generated with large language models (LLMs) - as part of their training and/or evaluation procedures. Beside the curation cost, this also risks overlooking unanticipated, less obvious prompts that trigger biased generation, even in models that have undergone debiasing. In this work, we introduce Bias-Guided Prompt Search (BGPS), a framework that automatically generates prompts that aim to maximize the presence of biases in the resulting images. BGPS comprises two components: (1) an LLM instructed to produce attribute-neutral prompts and (2) attribute classifiers acting on the TTI's internal representations that steer the decoding process of the LLM toward regions of the prompt space that amplify the image attributes of interest. We conduct extensive experiments on Stable Diffusion 1.5 and a state-of-the-art debiased model and discover an array of subtle and previously undocumented biases that severely deteriorate fairness metrics. Crucially, the discovered prompts are interpretable, i.e they may be entered by a typical user, quantitatively improving the perplexity metric compared to a prominent hard prompt optimization counterpart. Our findings uncover TTI vulnerabilities, while BGPS expands the bias search space and can act as a new evaluation tool for bias mitigation.

Maximos Kaliakatsos-Papakostas, Dimos Makris, Konstantinos Soiledis et al.

Melodic harmonization, the task of generating harmonic accompaniments for a given melody, remains a central challenge in computational music generation. Recent single encoder transformer approaches have framed harmonization as a masked sequence modeling problem, but existing training curricula inspired by discrete diffusion often result in weak (cross) attention between melody and harmony. This leads to limited exploitation of melodic cues, particularly in out-of-domain contexts. In this work, we introduce a training curriculum, FF (full-to-full), which keeps all harmony tokens masked for several training steps before progressively unmasking entire sequences during training to strengthen melody-harmony interactions. We systematically evaluate this approach against prior curricula across multiple experimental axes, including temporal quantization (quarter vs. sixteenth note), bar-level vs. time-signature conditioning, melody representation (full range vs. pitch class), and inference-time unmasking strategies. Models are trained on the HookTheory dataset and evaluated both in-domain and on a curated collection of jazz standards, using a comprehensive set of metrics that assess chord progression structure, harmony-melody alignment, and rhythmic coherence. Results demonstrate that the proposed FF curriculum consistently outperforms baselines in nearly all metrics, with particularly strong gains in out-of-domain evaluations where harmonic adaptability to novel melodic queues is crucial. We further find that quarter-note quantization, intertwining of bar tokens, and pitch-class melody representations are advantageous in the FF setting. Our findings highlight the importance of training curricula in enabling effective melody conditioning and suggest that full-to-full unmasking offers a robust strategy for single encoder harmonization.

Maximos Kaliakatsos-Papakostas, Konstantinos Soiledis, Theodoros Tsamis et al.

Transformer architectures offer significant advantages regarding the generation of symbolic music; their capabilities for incorporating user preferences toward what they generate is being studied under many aspects. This paper studies the inclusion of predefined chord constraints in melodic harmonization, i.e., where a desired chord at a specific location is provided along with the melody as inputs and the autoregressive transformer model needs to incorporate the chord in the harmonization that it generates. The peculiarities of involving such constraints is discussed and an algorithm is proposed for tackling this task. This algorithm is called B* and it combines aspects of beam search and A* along with backtracking to force pretrained transformers to satisfy the chord constraints, at the correct onset position within the correct bar. The algorithm is brute-force and has exponential complexity in the worst case; however, this paper is a first attempt to highlight the difficulties of the problem and proposes an algorithm that offers many possibilities for improvements since it accommodates the involvement of heuristics.

Dimitrios Damianos, Leon Voukoutis, Georgios Skyrianos et al.

Generative Vision-Language Models (VLMs) perform well on multimodal reasoning, but how visual inputs are transformed to text remains poorly understood. Existing interpretability work on VLMs uses Sparse Autoencoders (SAEs), which decompose static residual representations and miss the functional updates that drive cross-modal interaction. We adopt a function-centric framework based on Transcoders, sparse approximations of MLP sublayers that act as a causal proxy for layer-wise computation. Applied to Gemma 3-4B-IT, the framework decomposes the model into interpretable computational pathways linking image patches to directions in token generation. Transcoder attributions produce stronger and more stable effects on visually grounded tokens under patch ablation than SAE attributions, and align better with semantically relevant image regions. A False Visual Grounding counterfactual analysis confirms that the recovered pathways are specific to vision-language interaction.Finally, we perform a structural analysis of hallucinated generations, by extracting graph-based indicators from circuit traces produced by the transcoders. A logistic classifier over these mechanistic graph features predicts hallucinations at AUC $0.68$. These results show that function-centric circuit decomposition yields interpretable and predictive accounts of multimodal computation in VLMs.

Christos Chatzisavvas, Panagiotis Rigas, George Ioannakis et al.

Slot-based object-centric learning represents an image as a set of latent slots with a decoder that combines them into an image or features. The decoder specifies how slots are combined into an output, but the slot set is typically fixed: the number of slots is chosen upfront and slots are only refined. This can lead to multiple slots competing for overlapping regions of the same entity rather than focusing on distinct regions. We introduce slot merging: a drop-in, lightweight operation on the slot set that merges overlapping slots during training. We quantify overlap with a Soft-IoU score between slot-attention maps and combine selected pairs via a barycentric update that preserves gradient flow. Merging follows a fixed policy, with the decision threshold inferred from overlap statistics, requiring no additional learnable modules. Integrated into the established feature-reconstruction pipeline of DINOSAUR, the proposed method improves object factorization and mask quality, surpassing other adaptive methods in object discovery and segmentation benchmarks.

Dimitris Roussis, Leon Voukoutis, Georgios Paraskevopoulos et al.

We introduce Llama-Krikri-8B, a cutting-edge Large Language Model tailored for the Greek language, built on Meta's Llama 3.1-8B. Llama-Krikri-8B has been extensively trained on high-quality Greek data to ensure superior adaptation to linguistic nuances. With 8 billion parameters, it offers advanced capabilities while maintaining efficient computational performance. Llama-Krikri-8B supports both Modern Greek and English, and is also equipped to handle polytonic text and Ancient Greek. The chat version of Llama-Krikri-8B features a multi-stage post-training pipeline, utilizing both human and synthetic instruction and preference data, by applying techniques such as MAGPIE. In addition, for evaluation, we propose three novel public benchmarks for Greek. Our evaluation on existing as well as the proposed benchmarks shows notable improvements over comparable Greek and multilingual LLMs in both natural language understanding and generation as well as code generation.

Maximos Kaliakatsos-Papakostas, Gregoris Bastas, Dimos Makris et al.

Guitar tablature transcription consists in deducing the string and the fret number on which each note should be played to reproduce the actual musical part. This assignment should lead to playable string-fret combinations throughout the entire track and, in general, preserve parsimonious motion between successive combinations. Throughout the history of guitar playing, specific chord fingerings have been developed across different musical styles that facilitate common idiomatic voicing combinations and motion between them. This paper presents a method for assigning guitar tablature notation to a given MIDI-based musical part (possibly consisting of multiple polyphonic tracks), i.e. no information about guitar-idiomatic expressional characteristics is involved (e.g. bending etc.) The current strategy is based on machine learning and requires a basic assumption about how much fingers can stretch on a fretboard; only standard 6-string guitar tuning is examined. The proposed method also examines the transcription of music pieces that was not meant to be played or could not possibly be played by a guitar (e.g. potentially a symphonic orchestra part), employing a rudimentary method for augmenting musical information and training/testing the system with artificial data. The results present interesting aspects about what the system can achieve when trained on the initial and augmented dataset, showing that the training with augmented data improves the performance even in simple, e.g. monophonic, cases. Results also indicate weaknesses and lead to useful conclusions about possible improvements.

Georgios Chatzichristodoulou, Despoina Kosmopoulou, Antonios Kritikos et al.

SER is a challenging task due to the subjective nature of human emotions and their uneven representation under naturalistic conditions. We propose MEDUSA, a multimodal framework with a four-stage training pipeline, which effectively handles class imbalance and emotion ambiguity. The first two stages train an ensemble of classifiers that utilize DeepSER, a novel extension of a deep cross-modal transformer fusion mechanism from pretrained self-supervised acoustic and linguistic representations. Manifold MixUp is employed for further regularization. The last two stages optimize a trainable meta-classifier that combines the ensemble predictions. Our training approach incorporates human annotation scores as soft targets, coupled with balanced data sampling and multitask learning. MEDUSA ranked 1st in Task 1: Categorical Emotion Recognition in the Interspeech 2025: Speech Emotion Recognition in Naturalistic Conditions Challenge.

Efthymios Georgiou, Vassilis Katsouros, Yannis Avrithis et al.

While multimodal fusion has been extensively studied in Multimodal Sentiment Analysis (MSA), the role of fusion depth and multimodal capacity allocation remains underexplored. In this work, we position fusion depth, scalability, and dedicated multimodal capacity as primary factors for effective fusion. We introduce DeepMLF, a novel multimodal language model (LM) with learnable tokens tailored toward deep fusion. DeepMLF leverages an audiovisual encoder and a pretrained decoder LM augmented with multimodal information across its layers. We append learnable tokens to the LM that: 1) capture modality interactions in a controlled fashion and 2) preserve independent information flow for each modality. These fusion tokens gather linguistic information via causal self-attention in LM Blocks and integrate with audiovisual information through cross-attention MM Blocks. Serving as dedicated multimodal capacity, this design enables progressive fusion across multiple layers, providing depth in the fusion process. Our training recipe combines modality-specific losses and language modelling loss, with the decoder LM tasked to predict ground truth polarity. Across three MSA benchmarks with varying dataset characteristics, DeepMLF achieves state-of-the-art performance. Our results confirm that deeper fusion leads to better performance, with optimal fusion depths (5-7) exceeding those of existing approaches. Additionally, our analysis on the number of fusion tokens reveals that small token sets ($\sim$20) achieve optimal performance. We examine the importance of representation learning order (fusion curriculum) through audiovisual encoder initialization experiments. Our ablation studies demonstrate the superiority of the proposed fusion design and gating while providing a holistic examination of DeepMLF's scalability to LLMs, and the impact of each training objective and embedding regularization.

Dimitrios Damianos, Leon Voukoutis, Georgios Paraskevopoulos et al.

We present a multimodal fusion framework that bridges pre-trained decoder-based large language models (LLM) and acoustic encoder-decoder architectures such as Whisper, with the aim of building speech-enabled LLMs. Instead of directly using audio embeddings, we explore an intermediate audio-conditioned text space as a more effective mechanism for alignment. Our method operates fully in continuous text representation spaces, fusing Whisper's hidden decoder states with those of an LLM through cross-modal attention, and supports both offline and streaming modes. We introduce \textit{VoxKrikri}, the first Greek speech LLM, and show through analysis that our approach effectively aligns representations across modalities. These results highlight continuous space fusion as a promising path for multilingual and low-resource speech LLMs, while achieving state-of-the-art results for Automatic Speech Recognition in Greek, providing an average $\sim20\%$ relative improvement across benchmarks.

Georgios Paraskevopoulos, Chara Tsoukala, Athanasios Katsamanis et al.

The development of speech technologies for languages with limited digital representation poses significant challenges, primarily due to the scarcity of available data. This issue is exacerbated in the era of large, data-intensive models. Recent research has underscored the potential of leveraging weak supervision to augment the pool of available data. In this study, we compile an 800-hour corpus of Modern Greek from podcasts and employ Whisper large-v3 to generate silver transcriptions. This corpus is utilized to fine-tune our models, aiming to assess the efficacy of this approach in enhancing ASR performance. Our analysis spans 16 distinct podcast domains, alongside evaluations on established datasets for Modern Greek. The findings indicate consistent WER improvements, correlating with increases in both data volume and model size. Our study confirms that assembling large, weakly supervised corpora serves as a cost-effective strategy for advancing speech technologies in under-resourced languages.

Theodoros Kouzelis, Georgios Paraskevopoulos, Athanasios Katsamanis et al.

The study of speech disorders can benefit greatly from time-aligned data. However, audio-text mismatches in disfluent speech cause rapid performance degradation for modern speech aligners, hindering the use of automatic approaches. In this work, we propose a simple and effective modification of alignment graph construction of CTC-based models using Weighted Finite State Transducers. The proposed weakly-supervised approach alleviates the need for verbatim transcription of speech disfluencies for forced alignment. During the graph construction, we allow the modeling of common speech disfluencies, i.e. repetitions and omissions. Further, we show that by assessing the degree of audio-text mismatch through the use of Oracle Error Rate, our method can be effectively used in the wild. Our evaluation on a corrupted version of the TIMIT test set and the UCLASS dataset shows significant improvements, particularly for recall, achieving a 23-25% relative improvement over our baselines.