Social Hatred: Efficient Multimodal Detection of Hatemongers

Automatic detection of online hate speech serves as a crucial step in the detoxification of the online discourse. Moreover, accurate classification can promote a better understanding of the proliferation of hate as a social phenomenon. While most prior work focus on the detection of hateful utterances, we argue that focusing on the user level is as important, albeit challenging. In this paper we consider a multimodal aggregative approach for the detection of hate-mongers, taking into account the potentially hateful texts, user activity, and the user network. Evaluating our method on three unique datasets X (Twitter), Gab, and Parler we show that processing a user's texts in her social context significantly improves the detection of hate mongers, compared to previously used text and graph-based methods. We offer comprehensive set of results obtained in different experimental settings as well as qualitative analysis of illustrative cases. Our method can be used to improve the classification of coded messages, dog-whistling, and racial gas-lighting, as well as to inform intervention measures. Moreover, we demonstrate that our multimodal approach performs well across very different content platforms and over large datasets and networks.

A Robust Method for Pitch Tracking in the Frequency Following Response using Harmonic Amplitude Summation Filterbank

The Frequency Following Response (FFR) reflects the brain's neural encoding of auditory stimuli including speech. Because the fundamental frequency (F0), a physical correlate of pitch, is one of the essential features of speech, there has been particular interest in characterizing the FFR at F0, especially when F0 varies over time. The standard method for extracting F0 in FFRs has been the Autocorrelation Function (ACF). This paper investigates harmonic-structure-based F0 estimation algorithms, originally developed for speech and music, and resolves their poor performance when applied to FFRs in two steps. Firstly, given that unlike in speech or music, stimulus F0 of FFRs is already known, we introduce a stimulus-aware filterbank that selectively aggregates amplitudes at F0 and its harmonics while suppressing noise at non-harmonic frequencies. This method, called Harmonic Amplitude Summation (HAS), evaluates F0 candidates only within a range centered around the stimulus F0. Secondly, unlike other pitch tracking methods that select the highest peak, our method chooses the most prominent one, as it better reflects the underlying periodicity of FFRs. To the best of our knowledge, this is the first study to propose an F0 estimation algorithm for FFRs that relies on harmonic structure. Analyzing recorded FFRs from 16 normal hearing subjects to 4 natural speech stimuli with a wide F0 variation from 89 Hz to 452 Hz showed that this method outperformed ACF by reducing the average Root-Mean-Square-Error (RMSE) within each response and stimulus F0 contour pair by 8.8% to 47.4%, depending on the stimulus.

ClearerVoice-Studio: Bridging Advanced Speech Processing Research and Practical Deployment

This paper introduces ClearerVoice-Studio, an open-source, AI-powered speech processing toolkit designed to bridge cutting-edge research and practical application. Unlike broad platforms like SpeechBrain and ESPnet, ClearerVoice-Studio focuses on interconnected speech tasks of speech enhancement, separation, super-resolution, and multimodal target speaker extraction. A key advantage is its state-of-the-art pretrained models, including FRCRN with 3 million uses and MossFormer with 2.5 million uses, optimized for real-world scenarios. It also offers model optimization tools, multi-format audio support, the SpeechScore evaluation toolkit, and user-friendly interfaces, catering to researchers, developers, and end-users. Its rapid adoption attracting 3000 GitHub stars and 239 forks highlights its academic and industrial impact. This paper details ClearerVoice-Studio's capabilities, architectures, training strategies, benchmarks, community impact, and future plan. Source code is available at https://github.com/modelscope/ClearerVoice-Studio.

Enhanced Hybrid Transducer and Attention Encoder Decoder with Text Data

A joint speech and text optimization method is proposed for hybrid transducer and attention-based encoder decoder (TAED) modeling to leverage large amounts of text corpus and enhance ASR accuracy. The joint TAED (J-TAED) is trained with both speech and text input modalities together, while it only takes speech data as input during inference. The trained model can unify the internal representations from different modalities, and be further extended to text-based domain adaptation. It can effectively alleviate data scarcity for mismatch domain tasks since no speech data is required. Our experiments show J-TAED successfully integrates speech and linguistic information into one model, and reduce the WER by 5.8 ~12.8% on the Librispeech dataset. The model is also evaluated on two out-of-domain datasets: one is finance and another is named entity focused. The text-based domain adaptation brings 15.3% and 17.8% WER reduction on those two datasets respectively.

MFTCXplain: A Multilingual Benchmark Dataset for Evaluating the Moral Reasoning of LLMs through Hate Speech Multi-hop Explanation

Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via hate speech multi-hop explanation using Moral Foundation Theory (MFT). The dataset comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Empirical results highlight a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. These findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning.

IndieFake Dataset: A Benchmark Dataset for Audio Deepfake Detection

Advancements in audio deepfake technology offers benefits like AI assistants, better accessibility for speech impairments, and enhanced entertainment. However, it also poses significant risks to security, privacy, and trust in digital communications. Detecting and mitigating these threats requires comprehensive datasets. Existing datasets lack diverse ethnic accents, making them inadequate for many real-world scenarios. Consequently, models trained on these datasets struggle to detect audio deepfakes in diverse linguistic and cultural contexts such as in South-Asian countries. Ironically, there is a stark lack of South-Asian speaker samples in the existing datasets despite constituting a quarter of the worlds population. This work introduces the IndieFake Dataset (IFD), featuring 27.17 hours of bonafide and deepfake audio from 50 English speaking Indian speakers. IFD offers balanced data distribution and includes speaker-level characterization, absent in datasets like ASVspoof21 (DF). We evaluated various baselines on IFD against existing ASVspoof21 (DF) and In-The-Wild (ITW) datasets. IFD outperforms ASVspoof21 (DF) and proves to be more challenging compared to benchmark ITW dataset. The dataset will be publicly available upon acceptance.

USAD: Universal Speech and Audio Representation via Distillation

Self-supervised learning (SSL) has revolutionized audio representations, yet models often remain domain-specific, focusing on either speech or non-speech tasks. In this work, we present Universal Speech and Audio Distillation (USAD), a unified approach to audio representation learning that integrates diverse audio types - speech, sound, and music - into a single model. USAD employs efficient layer-to-layer distillation from domain-specific SSL models to train a student on a comprehensive audio dataset. USAD offers competitive performance across various benchmarks and datasets, including frame and instance-level speech processing tasks, audio tagging, and sound classification, achieving near state-of-the-art results with a single encoder on SUPERB and HEAR benchmarks.

MLLP-VRAIN UPV system for the IWSLT 2025 Simultaneous Speech Translation Translation task

This work describes the participation of the MLLP-VRAIN research group in the shared task of the IWSLT 2025 Simultaneous Speech Translation track. Our submission addresses the unique challenges of real-time translation of long-form speech by developing a modular cascade system that adapts strong pre-trained models to streaming scenarios. We combine Whisper Large-V3-Turbo for ASR with the multilingual NLLB-3.3B model for MT, implementing lightweight adaptation techniques rather than training new end-to-end models from scratch. Our approach employs document-level adaptation with prefix training to enhance the MT model's ability to handle incomplete inputs, while incorporating adaptive emission policies including a wait-$k$ strategy and RALCP for managing the translation stream. Specialized buffer management techniques and segmentation strategies ensure coherent translations across long audio sequences. Experimental results on the ACL60/60 dataset demonstrate that our system achieves a favorable balance between translation quality and latency, with a BLEU score of 31.96 and non-computational-aware StreamLAAL latency of 2.94 seconds. Our final model achieves a preliminary score on the official test set (IWSLT25Instruct) of 29.8 BLEU. Our work demonstrates that carefully adapted pre-trained components can create effective simultaneous translation systems for long-form content without requiring extensive in-domain parallel data or specialized end-to-end training.

Dense SAE Latents Are Features, Not Bugs

Sparse autoencoders (SAEs) are designed to extract interpretable features from language models by enforcing a sparsity constraint. Ideally, training an SAE would yield latents that are both sparse and semantically meaningful. However, many SAE latents activate frequently (i.e., are \emph{dense}), raising concerns that they may be undesirable artifacts of the training procedure. In this work, we systematically investigate the geometry, function, and origin of dense latents and show that they are not only persistent but often reflect meaningful model representations. We first demonstrate that dense latents tend to form antipodal pairs that reconstruct specific directions in the residual stream, and that ablating their subspace suppresses the emergence of new dense features in retrained SAEs -- suggesting that high density features are an intrinsic property of the residual space. We then introduce a taxonomy of dense latents, identifying classes tied to position tracking, context binding, entropy regulation, letter-specific output signals, part-of-speech, and principal component reconstruction. Finally, we analyze how these features evolve across layers, revealing a shift from structural features in early layers, to semantic features in mid layers, and finally to output-oriented signals in the last layers of the model. Our findings indicate that dense latents serve functional roles in language model computation and should not be dismissed as training noise.

I Know You're Listening: Adaptive Voice for HRI

While the use of social robots for language teaching has been explored, there remains limited work on a task-specific synthesized voices for language teaching robots. Given that language is a verbal task, this gap may have severe consequences for the effectiveness of robots for language teaching tasks. We address this lack of L2 teaching robot voices through three contributions: 1. We address the need for a lightweight and expressive robot voice. Using a fine-tuned version of Matcha-TTS, we use emoji prompting to create an expressive voice that shows a range of expressivity over time. The voice can run in real time with limited compute resources. Through case studies, we found this voice more expressive, socially appropriate, and suitable for long periods of expressive speech, such as storytelling. 2. We explore how to adapt a robot's voice to physical and social ambient environments to deploy our voices in various locations. We found that increasing pitch and pitch rate in noisy and high-energy environments makes the robot's voice appear more appropriate and makes it seem more aware of its current environment. 3. We create an English TTS system with improved clarity for L2 listeners using known linguistic properties of vowels that are difficult for these listeners. We used a data-driven, perception-based approach to understand how L2 speakers use duration cues to interpret challenging words with minimal tense (long) and lax (short) vowels in English. We found that the duration of vowels strongly influences the perception for L2 listeners and created an "L2 clarity mode" for Matcha-TTS that applies a lengthening to tense vowels while leaving lax vowels unchanged. Our clarity mode was found to be more respectful, intelligible, and encouraging than base Matcha-TTS while reducing transcription errors in these challenging tense/lax minimal pairs.