Searcharxiv⌕ Search

arXiv subjects

Ina Kodrasi

Publications and source records attributed to Ina Kodrasi.

At least 19 recordsLinked to original sources

Geometric Latent Reasoning Induces Shorter Generations in LLMs

Large language models solve complex problems by generating lengthy chains of explicit reasoning tokens. While effective, this makes reasoning expensive, length-sensitive, and constrained to (discrete) natural language. While latent reasoning offers a continuous alternative, determining useful structures for intermediate latent states is an open challenge. In this paper, we formulate latent reasoning as a geometric path-approximation problem within the model's pretrained token-embedding space. We introduce Geometric Latent Reasoning (GLR), which uses a lightweight transition head to predict iterative direction updates in embedding space. Using textual chain-of-thought traces as anchors, GLR learns to approximate discrete reasoning trajectories while permitting continuous deviations from exact token embeddings. Evaluations on mathematical reasoning benchmarks using Qwen3 models reveal an emergent phenomenon: geometric latent reasoning induces substantially shorter generations without an explicit length objective. By replacing early explicit reasoning with continuous latent steps, models often reach correct answers using substantially fewer total generation steps. These findings suggest that continuous trajectories act as compact intermediate reasoning states, exposing a new tradeoff between latent computation budget, output length, and accuracy. Our code is publicly available at: https://github.com/idiap/GLR. The project page is available at: https://glr-llm.github.io/.

cs.CL↗

Discriminator-free adversarial domain adaptation for robust speech classification

The performance of speech classification systems such as speech emotion recognition or pathological speech detection can degrade substantially when there is a domain shift between training and test data. This shift can arise due to e.g., differences in recording setup, acoustic environment, speaker, language, or speaking style. Adversarial domain adaptation addresses this problem by encouraging representations that are less sensitive to domain-specific differences. However, many adversarial approaches rely on training a separate domain discriminator, which can be unstable and difficult to train. In this paper, we propose to use a discriminator-free adversarial learning network (DALN) for unsupervised domain adaptation in speech classification tasks. DALN reuses the task classifier as an implicit discriminator and aligns domains through the nuclear-norm Wasserstein discrepancy on classifier outputs. We extensively evaluate the performance of DALN on intra-lingual, cross-lingual, and noisy adaptations for speech emotion recognition and pathological speech detection. Extensive experimental results show that DALN generally achieves the best performance across the considered tasks and domain shifts, outperforming discriminator-based domain adaptation approaches.

eess.AS↗

SCLAP: Temporal structural alignment improves localization in long-form audio

Audio-text models such as Contrastive Language-Audio Pretraining (CLAP) typically pool frame and token sequences uniformly into global embeddings. This discards temporal structure and impairs both retrieval and localization, especially for short events in long recordings. We introduce SCLAP, a sequence-level contrastive framework that aligns audio frames and text tokens using monotone one-dimensional optimal transport. Its learned marginals and transport plan provide fine-grained temporal supervision from clip-level captions alone. On short-form audio (i.e., Clotho and TACOS datasets with 15-30s recordings), SCLAP remains competitive for retrieval while improving zero-shot localization over CLAP and Text-aware Attention Pooling. On long-form audio (i.e., CASTELLA dataset with up to 300s recordings), SCLAP improves retrieval performance and its localization advantage grows as the event-to-context ratio decreases. At inference, SCLAP requires no optimal transport computation since retrieval can use pooled embeddings and localization can use inexpensive frame-text cosine scores.

eess.AS↗

Rethinking Batch Composition: Promoting Feature Diversity for Speech Enhancement

Speech enhancement systems are commonly trained using randomly sampled mini-batches, which may contain limited diversity in intrinsic acoustic characteristics of the clean speech signals. Motivated by previous findings linking clean speech characteristics to enhancement performance, we propose a feature-stratified mini-batch sampling strategy that divides training examples into quantiles according to a clean-speech characteristic and enforces equal representation within each mini-batch. Experiments across two large-scale English and Spanish speech datasets and a smaller dataset containing neurotypical and pathological speech demonstrate improvements over standard random mini-batch sampling across multiple objective enhancement metrics. Ablation studies demonstrate that the gains are robust across training runs, while their magnitude depends on the acoustic characteristic used for stratification. The proposed strategy requires no modification to the enhancement model or training objective, providing a simple and complementary approach to speech enhancement training.

eess.AS↗

Augmenting Large Audio Language Models with Low-Level Acoustic Features for Dysarthric Speech Detection

Automatic dysarthric speech detection approaches can support traditional clinical diagnosis, which relies on costly and time-consuming evaluation by a speech and language pathologist. Existing automatic approaches predominantly rely on deep learning (DL). More recently, Large Audio Language Models (LALMs) have emerged as a promising alternative given their strong performance across various tasks, but their application to dysarthric speech detection has not yet been established. We propose a framework that fine-tunes LALMs for dysarthric speech detection on speech recordings combined with textual information comprising low-level acoustic features and speaker demographics. Across two LALMs, our framework outperforms DL-based baselines, with Qwen2-Audio-Instruct achieving state-of-the-art performance. An ablation study shows that incorporating acoustic features and speaker demographics during fine-tuning improves LALM performance, while LALMs alone exhibit only chance-level zero-shot performance. These findings establish an effective approach for adapting LALMs to dysarthric speech detection.

eess.AS↗

Multiclass Speech Classification Under Noise Disparity

We investigate noise disparity in multi-class speech classification tasks and develop a strategy to prevent classifiers from exploiting class-dependent noise characteristics. Building on our previous work for binary classification, we introduce a multi-class cross-augmentation scheme that exposes each class to the noise characteristics of the other classes, thereby removing the association between individual noise conditions and class labels. We compare this training-based approach with speech enhancement (SE) as a preprocessing strategy, which aims to suppress noise-related cues directly from the input. Experiments on multi-class emotion recognition show that cross-augmentation effectively mitigates the effect of noise disparity across a range of signal-to-noise-ratios, while SE has a detrimental effect on model performance.

eess.AS↗

Test-Time Compute Scaling for ASR with Depth-Conditioned Looped Transformers

End-to-end ASR systems typically use fixed-depth acoustic encoders at inference, making it difficult to trade additional test-time computation for improved recognition without training a larger model. A natural approach is to reuse a shared Transformer block recurrently, but we find that naive looping does not fully exploit additional recurrent compute. We introduce LARM, a depth-conditioned looped Transformer that turns recurrent encoder depth into a controllable test-time compute axis. LARM combines sparse CTC checkpoints, supervision-clock embeddings, FiLM depth conditioning, and delayed soft-posterior feedback. These components structure the loop into recognition checkpoints separated by latent refinement phases and allow shared weights to specialize across recurrent steps. On LibriSpeech, LARM improves WER as the number of inference loops increases and achieves performance competitive with deeper unshared-parameter baselines. Our results show that test-time compute scaling can extend beyond autoregressive language-model reasoning to continuous non-autoregressive speech recognition.

cs.LG↗

Data Augmentation for Pathological Speech Enhancement

The performance of state-of-the-art speech enhancement (SE) models considerably degrades for pathological speech due to atypical acoustic characteristics and limited data availability. This paper systematically investigates data augmentation (DA) strategies to improve SE performance for pathological speakers affected by Parkinson`s disease, evaluating both predictive and generative SE models. We examine three DA categories, i.e., transformative, generative, and noise augmentation, assessing their impact with objective SE metrics. Experimental results show that noise augmentation consistently delivers the largest and most robust gains, transformative augmentations provide moderate improvements, while generative augmentation yields limited benefits and can harm performance as the amount of synthetic data increases. Furthermore, we show that the effectiveness of DA varies depending on the SE model, with DA being more beneficial for predictive SE models. While our results demonstrate that DA improves SE performance for pathological speakers, a performance gap between neurotypical and pathological speech persists, highlighting the need for future research on targeted DA strategies for pathological speech.

eess.AS↗

CLAP-Based Automatic Word Naming Recognition in Post-Stroke Aphasia

Conventional automatic word-naming recognition systems struggle to recognize words from post-stroke patients with aphasia because of disfluencies and mispronunciations, limiting reliable automated assessment in this population. In this paper, we propose a Contrastive Language-Audio Pretraining (CLAP) based approach for automatic word-naming recognition to address this challenge by leveraging text-audio alignment. Our approach treats word-naming recognition as an audio-text matching problem, projecting speech signals and textual prompts into a shared embedding space to identify intended words even in challenging recordings. Evaluated on two speech datasets of French post-stroke patients with aphasia, our approach achieves up to 90% accuracy, outperforming existing classification-based and automatic speech recognition-based baselines.

eess.AS↗

A Differentiable Alignment Framework for Sequence-to-Sequence Modeling via Optimal Transport

Accurate sequence-to-sequence (seq2seq) alignment is critical for applications like medical speech analysis and language learning tools relying on automatic speech recognition (ASR). State-of-the-art end-to-end (E2E) ASR systems, such as the Connectionist Temporal Classification (CTC) and transducer-based models, suffer from peaky behavior and alignment inaccuracies. In this paper, we propose a novel differentiable alignment framework based on one-dimensional optimal transport, enabling the model to learn a single alignment and perform ASR in an E2E manner. We introduce a pseudo-metric, called Sequence Optimal Transport Distance (SOTD), over the sequence space and discuss its theoretical properties. Based on the SOTD, we propose Optimal Temporal Transport Classification (OTTC) loss for ASR and contrast its behavior with CTC. Experimental results on the TIMIT, AMI, and LibriSpeech datasets show that our method considerably improves alignment performance compared to CTC and the more recently proposed Consistency-Regularized CTC, though with a trade-off in ASR performance. We believe this work opens new avenues for seq2seq alignment research, providing a solid foundation for further exploration and development within the community. Our code is publicly available at: https://github.com/idiap/OTTC

cs.LG↗

Latent Space Factorization in LoRA

Low-rank adaptation (LoRA) is a widely used method for parameter-efficient finetuning. However, existing LoRA variants lack mechanisms to explicitly disambiguate task-relevant information within the learned low-rank subspace, potentially limiting downstream performance. We propose Factorized Variational Autoencoder LoRA (FVAE-LoRA), which leverages a VAE to learn two distinct latent spaces. Our novel Evidence Lower Bound formulation explicitly promotes factorization between the latent spaces, dedicating one latent space to task-salient features and the other to residual information. Extensive experiments on text, audio, and image tasks demonstrate that FVAE-LoRA consistently outperforms standard LoRA. Moreover, spurious correlation evaluations confirm that FVAE-LoRA better isolates task-relevant signals, leading to improved robustness under distribution shifts. Our code is publicly available at: https://github.com/idiap/FVAE-LoRA

cs.LG↗

Influence of Clean Speech Characteristics on Speech Enhancement Performance

Speech enhancement (SE) performance is known to depend on noise characteristics and signal to noise ratio (SNR), yet intrinsic properties of the clean speech signal itself remain an underexplored factor. In this work, we systematically analyze how clean speech characteristics influence enhancement difficulty across multiple state of the art SE models, languages, and noise conditions. We extract a set of pitch, formant, loudness, and spectral flux features from clean speech and compute correlations with objective SE metrics, including frequency weighted segmental SNR and PESQ. Our results show that formant amplitudes are consistently predictive of SE performance, with higher and more stable formants leading to larger enhancement gains. We further demonstrate that performance varies substantially even within a single speaker's utterances, highlighting the importance of intraspeaker acoustic variability. These findings provide new insights into SE challenges, suggesting that intrinsic speech characteristics should be considered when designing datasets, evaluation protocols, and enhancement models.

eess.AS↗

Generalizability of Predictive and Generative Speech Enhancement Models to Pathological Speakers

State of the art speech enhancement (SE) models achieve strong performance on neurotypical speech, but their effectiveness is substantially reduced for pathological speech. In this paper, we investigate strategies to address this gap for both predictive and generative SE models, including i) training models from scratch using pathological data, ii) finetuning models pretrained on neurotypical speech with additional data from pathological speakers, and iii) speaker specific personalization using only data from the individual pathological test speaker. Our results show that, despite the limited size of pathological speech datasets, SE models can be successfully trained or finetuned on such data. Finetuning models with data from several pathological speakers yields the largest performance improvements, while speaker specific personalization is less effective, likely due to the small amount of data available per speaker. These findings highlight the challenges and potential strategies for improving SE performance for pathological speakers.

eess.AS↗

Towards interpretable emotion recognition: Identifying key features with machine learning

Unsupervised methods, such as wav2vec2 and HuBERT, have achieved state-of-the-art performance in audio tasks, leading to a shift away from research on interpretable features. However, the lack of interpretability in these methods limits their applicability in critical domains like medicine, where understanding feature relevance is crucial. To better understand the features of unsupervised models, it remains critical to identify the interpretable features relevant to a given task. In this work, we focus on emotion recognition and use machine learning algorithms to identify and generalize the most important interpretable features for this task. While previous studies have explored feature relevance in emotion recognition, they are often constrained by narrow contexts and present inconsistent findings. Our approach aims to overcome these limitations, providing a broader and more robust framework for identifying the most important interpretable features.

eess.AS↗

Overview of Automatic Speech Analysis and Technologies for Neurodegenerative Disorders: Diagnosis and Assistive Applications

Advancements in spoken language technologies for neurodegenerative speech disorders are crucial for meeting both clinical and technological needs. This overview paper is vital for advancing the field, as it presents a comprehensive review of state-of-the-art methods in pathological speech detection, automatic speech recognition, pathological speech intelligibility enhancement, intelligibility and severity assessment, and data augmentation approaches for pathological speech. It also highlights key challenges, such as ensuring robustness, privacy, and interpretability. The paper concludes by exploring promising future directions, including the adoption of multimodal approaches and the integration of large language models to further advance speech technologies for neurodegenerative speech disorders.

eess.AS↗

Exploring In-Context Learning Capabilities of ChatGPT for Pathological Speech Detection

Automatic pathological speech detection approaches have shown promising results, gaining attention as potential diagnostic tools alongside costly traditional methods. While these approaches can achieve high accuracy, their lack of interpretability limits their applicability in clinical practice. In this paper, we investigate the use of multimodal Large Language Models (LLMs), specifically ChatGPT-4o, for automatic pathological speech detection in a few-shot in-context learning setting. Experimental results show that this approach not only delivers promising performance but also provides explanations for its decisions, enhancing model interpretability. To further understand its effectiveness, we conduct an ablation study to analyze the impact of different factors, such as input type and system prompts, on the final results. Our findings highlight the potential of multimodal LLMs for further exploration and advancement in automatic pathological speech detection.

eess.AS↗

Variational Autoencoder for Personalized Pathological Speech Enhancement

The generalizability of speech enhancement (SE) models across speaker conditions remains largely unexplored, despite its critical importance for broader applicability. This paper investigates the performance of the hybrid variational autoencoder (VAE)-non-negative matrix factorization (NMF) model for SE, focusing primarily on its generalizability to pathological speakers with Parkinson's disease. We show that VAE models trained on large neurotypical datasets perform poorly on pathological speech. While fine-tuning these pre-trained models with pathological speech improves performance, a performance gap remains between neurotypical and pathological speakers. To address this gap, we propose using personalized SE models derived from fine-tuning pre-trained models with only a few seconds of clean data from each speaker. Our results demonstrate that personalized models considerably enhance performance for all speakers, achieving comparable results for both neurotypical and pathological speakers.

eess.AS↗

Graph Neural Networks for Parkinsons Disease Detection

Despite the promising performance of state of the art approaches for Parkinsons Disease (PD) detection, these approaches often analyze individual speech segments in isolation, which can lead to suboptimal results. Dysarthric cues that characterize speech impairments from PD patients are expected to be related across segments from different speakers. Isolated segment analysis fails to exploit these inter segment relationships. Additionally, not all speech segments from PD patients exhibit clear dysarthric symptoms, introducing label noise that can negatively affect the performance and generalizability of current approaches. To address these challenges, we propose a novel PD detection framework utilizing Graph Convolutional Networks (GCNs). By representing speech segments as nodes and capturing the similarity between segments through edges, our GCN model facilitates the aggregation of dysarthric cues across the graph, effectively exploiting segment relationships and mitigating the impact of label noise. Experimental results demonstrate theadvantages of the proposed GCN model for PD detection and provide insights into its underlying mechanisms

cs.LG↗