Rethinking Tokenization for Rich Morphology: The Dominance of Unigram over BPE and Morphological Alignment

Prior work on language modeling showed conflicting findings about whether morphologically aligned approaches to tokenization improve performance, particularly for languages with complex morphology. To investigate this, we select a typologically diverse set of languages: Telugu (agglutinative), Hindi (primarily fusional with some agglutination), and English (fusional). We conduct a comprehensive evaluation of language models -- starting from tokenizer training and extending through the finetuning and downstream task evaluation. To account for the consistent performance differences observed across tokenizer variants, we focus on two key factors: morphological alignment and tokenization quality. To assess morphological alignment of tokenizers in Telugu, we create a dataset containing gold morpheme segmentations of 600 derivational and 7000 inflectional word forms. Our experiments reveal that better morphological alignment correlates positively -- though moderately -- with performance in syntax-based tasks such as Parts-of-Speech tagging, Named Entity Recognition and Dependency Parsing. However, we also find that the tokenizer algorithm (Byte-pair Encoding vs. Unigram) plays a more significant role in influencing downstream performance than morphological alignment alone. Naive Unigram tokenizers outperform others across most settings, though hybrid tokenizers that incorporate morphological segmentation significantly improve performance within the BPE framework. In contrast, intrinsic metrics like Corpus Token Count (CTC) and R\'enyi entropy showed no correlation with downstream performance.

SCDF: A Speaker Characteristics DeepFake Speech Dataset for Bias Analysis

Despite growing attention to deepfake speech detection, the aspects of bias and fairness remain underexplored in the speech domain. To address this gap, we introduce the Speaker Characteristics Deepfake (SCDF) dataset: a novel, richly annotated resource enabling systematic evaluation of demographic biases in deepfake speech detection. SCDF contains over 237,000 utterances in a balanced representation of both male and female speakers spanning five languages and a wide age range. We evaluate several state-of-the-art detectors and show that speaker characteristics significantly influence detection performance, revealing disparities across sex, language, age, and synthesizer type. These findings highlight the need for bias-aware development and provide a foundation for building non-discriminatory deepfake detection systems aligned with ethical and regulatory standards.

Iterative refinement, not training objective, makes HuBERT behave differently from wav2vec 2.0

Self-supervised models for speech representation learning now see widespread use for their versatility and performance on downstream tasks, but the effect of model architecture on the linguistic information learned in their representations remains under-studied. This study investigates two such models, HuBERT and wav2vec 2.0, and minimally compares two of their architectural differences: training objective and iterative pseudo-label refinement through multiple training iterations. We find that differences in canonical correlation of hidden representations to word identity, phoneme identity, and speaker identity are explained by training iteration, not training objective. We suggest that future work investigate the reason for the effectiveness of iterative refinement in encoding linguistic information in self-supervised speech representations.

Toward Machine Interpreting: Lessons from Human Interpreting Studies

Current speech translation systems, while having achieved impressive accuracies, are rather static in their behavior and do not adapt to real-world situations in ways human interpreters do. In order to improve their practical usefulness and enable interpreting-like experiences, a precise understanding of the nature of human interpreting is crucial. To this end, we discuss human interpreting literature from the perspective of the machine translation field, while considering both operational and qualitative aspects. We identify implications for the development of speech translation systems and argue that there is great potential to adopt many human interpreting principles using recent modeling techniques. We hope that our findings provide inspiration for closing the perceived usability gap, and can motivate progress toward true machine interpreting.

Optimal Transport Regularization for Speech Text Alignment in Spoken Language Models

Spoken Language Models (SLMs), which extend Large Language Models (LLMs) to perceive speech inputs, have gained increasing attention for their potential to advance speech understanding tasks. However, despite recent progress, studies show that SLMs often struggle to generalize across datasets, even for trained languages and tasks, raising concerns about whether they process speech in a text-like manner as intended. A key challenge underlying this limitation is the modality gap between speech and text representations. The high variability in speech embeddings may allow SLMs to achieve strong in-domain performance by exploiting unintended speech variations, ultimately hindering generalization. To mitigate this modality gap, we introduce Optimal Transport Regularization (OTReg), a method that formulates speech-text alignment as an optimal transport problem and derives a regularization loss to improve SLM training. In each training iteration, OTReg first establishes a structured correspondence between speech and transcript embeddings by determining the optimal transport plan, then incorporates the regularization loss based on this transport plan to optimize SLMs in generating speech embeddings that align more effectively with transcript embeddings. OTReg is lightweight, requiring no additional labels or learnable parameters, and integrates seamlessly into existing SLM training procedures. Extensive multilingual ASR experiments demonstrate that OTReg enhances speech-text alignment, mitigates the modality gap, and consequently improves SLM generalization across diverse datasets.

A Survey on Non-Intrusive ASR Refinement: From Output-Level Correction to Full-Model Distillation

Automatic Speech Recognition (ASR) has become an integral component of modern technology, powering applications such as voice-activated assistants, transcription services, and accessibility tools. Yet ASR systems continue to struggle with the inherent variability of human speech, such as accents, dialects, and speaking styles, as well as environmental interference, including background noise. Moreover, domain-specific conversations often employ specialized terminology, which can exacerbate transcription errors. These shortcomings not only degrade raw ASR accuracy but also propagate mistakes through subsequent natural language processing pipelines. Because redesigning an ASR model is costly and time-consuming, non-intrusive refinement techniques that leave the model's architecture unchanged have become increasingly popular. In this survey, we systematically review current non-intrusive refinement approaches and group them into five classes: fusion, re-scoring, correction, distillation, and training adjustment. For each class, we outline the main methods, advantages, drawbacks, and ideal application scenarios. Beyond method classification, this work surveys adaptation techniques aimed at refining ASR in domain-specific contexts, reviews commonly used evaluation datasets along with their construction processes, and proposes a standardized set of metrics to facilitate fair comparisons. Finally, we identify open research gaps and suggest promising directions for future work. By providing this structured overview, we aim to equip researchers and practitioners with a clear foundation for developing more robust, accurate ASR refinement pipelines.

Keyword Mamba: Spoken Keyword Spotting with State Space Models

Keyword spotting (KWS) is an essential task in speech processing. It is widely used in voice assistants and smart devices. Deep learning models like CNNs, RNNs, and Transformers have performed well in KWS. However, they often struggle to handle long-term patterns and stay efficient at the same time. In this work, we present Keyword Mamba, a new architecture for KWS. It uses a neural state space model (SSM) called Mamba. We apply Mamba along the time axis and also explore how it can replace the self-attention part in Transformer models. We test our model on the Google Speech Commands datasets. The results show that Keyword Mamba reaches strong accuracy with fewer parameters and lower computational cost. To our knowledge, this is the first time a state space model has been used for KWS. These results suggest that Mamba has strong potential in speech-related tasks.

Incorporating Contextual Paralinguistic Understanding in Large Speech-Language Models

Current large speech language models (Speech-LLMs) often exhibit limitations in empathetic reasoning, primarily due to the absence of training datasets that integrate both contextual content and paralinguistic cues. In this work, we propose two approaches to incorporate contextual paralinguistic information into model training: (1) an explicit method that provides paralinguistic metadata (e.g., emotion annotations) directly to the LLM, and (2) an implicit method that automatically generates novel training question-answer (QA) pairs using both categorical and dimensional emotion annotations alongside speech transcriptions. Our implicit method boosts performance (LLM-judged) by 38.41% on a human-annotated QA benchmark, reaching 46.02% when combined with the explicit approach, showing effectiveness in contextual paralinguistic understanding. We also validate the LLM judge by demonstrating its correlation with classification metrics, providing support for its reliability.

How Does a Deep Neural Network Look at Lexical Stress?

Despite their success in speech processing, neural networks often operate as black boxes, prompting the question: what informs their decisions, and how can we interpret them? This work examines this issue in the context of lexical stress. A dataset of English disyllabic words was automatically constructed from read and spontaneous speech. Several Convolutional Neural Network (CNN) architectures were trained to predict stress position from a spectrographic representation of disyllabic words lacking minimal stress pairs (e.g., initial stress WAllet, final stress exTEND), achieving up to 92% accuracy on held-out test data. Layerwise Relevance Propagation (LRP), a technique for CNN interpretability analysis, revealed that predictions for held-out minimal pairs (PROtest vs. proTEST ) were most strongly influenced by information in stressed versus unstressed syllables, particularly the spectral properties of stressed vowels. However, the classifiers also attended to information throughout the word. A feature-specific relevance analysis is proposed, and its results suggest that our best-performing classifier is strongly influenced by the stressed vowel's first and second formants, with some evidence that its pitch and third formant also contribute. These results reveal deep learning's ability to acquire distributed cues to stress from naturally occurring data, extending traditional phonetic work based around highly controlled stimuli.

Freeze and Reveal: Exposing Modality Bias in Vision-Language Models

Vision Language Models achieve impressive multi-modal performance but often inherit gender biases from their training data. This bias might be coming from both the vision and text modalities. In this work, we dissect the contributions of vision and text backbones to these biases by applying targeted debiasing using Counterfactual Data Augmentation and Task Vector methods. Inspired by data-efficient approaches in hate-speech classification, we introduce a novel metric, Degree of Stereotypicality and a corresponding debiasing method, Data Augmentation Using Degree of Stereotypicality - DAUDoS, to reduce bias with minimal computational cost. We curate a gender annotated dataset and evaluate all methods on VisoGender benchmark to quantify improvements and identify dominant source of bias. Our results show that CDA reduces the gender gap by 6% and DAUDoS by 3% but using only one-third of the data. Both methods also improve the model's ability to correctly identify gender in images by 3%, with DAUDoS achieving this improvement using only almost one-third of training data. From our experiment's, we observed that CLIP's vision encoder is more biased whereas PaliGemma2's text encoder is more biased. By identifying whether bias stems more from vision or text encoders, our work enables more targeted and effective bias mitigation strategies in future multi-modal systems.