Multi-Target Backdoor Attacks Against Speaker Recognition

In this work, we propose a multi-target backdoor attack against speaker identification using position-independent clicking sounds as triggers. Unlike previous single-target approaches, our method targets up to 50 speakers simultaneously, achieving success rates of up to 95.04%. To simulate more realistic attack conditions, we vary the signal-to-noise ratio between speech and trigger, demonstrating a trade-off between stealth and effectiveness. We further extend the attack to the speaker verification task by selecting the most similar training speaker - based on cosine similarity - as a proxy target. The attack is most effective when target and enrolled speaker pairs are highly similar, reaching success rates of up to 90% in such cases.

DualSpeechLM: Towards Unified Speech Understanding and Generation via Dual Speech Token Modeling with Large Language Models

Extending pre-trained Large Language Models (LLMs)'s speech understanding or generation abilities by introducing various effective speech tokens has attracted great attention in the speech community. However, building a unified speech understanding and generation model still faces the following challenges: (1) Due to the huge modality gap between speech tokens and text tokens, extending text LLMs to unified speech LLMs relies on large-scale paired data for fine-tuning, and (2) Generation and understanding tasks prefer information at different levels, e.g., generation benefits from detailed acoustic features, while understanding favors high-level semantics. This divergence leads to difficult performance optimization in one unified model. To solve these challenges, in this paper, we present two key insights in speech tokenization and speech language modeling. Specifically, we first propose an Understanding-driven Speech Tokenizer (USTokenizer), which extracts high-level semantic information essential for accomplishing understanding tasks using text LLMs. In this way, USToken enjoys better modality commonality with text, which reduces the difficulty of modality alignment in adapting text LLMs to speech LLMs. Secondly, we present DualSpeechLM, a dual-token modeling framework that concurrently models USToken as input and acoustic token as output within a unified, end-to-end framework, seamlessly integrating speech understanding and generation capabilities. Furthermore, we propose a novel semantic supervision loss and a Chain-of-Condition (CoC) strategy to stabilize model training and enhance speech generation performance. Experimental results demonstrate that our proposed approach effectively fosters a complementary relationship between understanding and generation tasks, highlighting the promising strategy of mutually enhancing both tasks in one unified model.

QAMRO: Quality-aware Adaptive Margin Ranking Optimization for Human-aligned Assessment of Audio Generation Systems

Evaluating audio generation systems, including text-to-music (TTM), text-to-speech (TTS), and text-to-audio (TTA), remains challenging due to the subjective and multi-dimensional nature of human perception. Existing methods treat mean opinion score (MOS) prediction as a regression problem, but standard regression losses overlook the relativity of perceptual judgments. To address this limitation, we introduce QAMRO, a novel Quality-aware Adaptive Margin Ranking Optimization framework that seamlessly integrates regression objectives from different perspectives, aiming to highlight perceptual differences and prioritize accurate ratings. Our framework leverages pre-trained audio-text models such as CLAP and Audiobox-Aesthetics, and is trained exclusively on the official AudioMOS Challenge 2025 dataset. It demonstrates superior alignment with human evaluations across all dimensions, significantly outperforming robust baseline models.

Munsit at NADI 2025 Shared Task 2: Pushing the Boundaries of Multidialectal Arabic ASR with Weakly Supervised Pretraining and Continual Supervised Fine-tuning

Automatic speech recognition (ASR) plays a vital role in enabling natural human-machine interaction across applications such as virtual assistants, industrial automation, customer support, and real-time transcription. However, developing accurate ASR systems for low-resource languages like Arabic remains a significant challenge due to limited labeled data and the linguistic complexity introduced by diverse dialects. In this work, we present a scalable training pipeline that combines weakly supervised learning with supervised fine-tuning to develop a robust Arabic ASR model. In the first stage, we pretrain the model on 15,000 hours of weakly labeled speech covering both Modern Standard Arabic (MSA) and various Dialectal Arabic (DA) variants. In the subsequent stage, we perform continual supervised fine-tuning using a mixture of filtered weakly labeled data and a small, high-quality annotated dataset. Our approach achieves state-of-the-art results, ranking first in the multi-dialectal Arabic ASR challenge. These findings highlight the effectiveness of weak supervision paired with fine-tuning in overcoming data scarcity and delivering high-quality ASR for low-resource, dialect-rich languages.

MultiAiTutor: Child-Friendly Educational Multilingual Speech Generation Tutor with LLMs

Generative speech models have demonstrated significant potential in personalizing teacher-student interactions, offering valuable real-world applications for language learning in children's education. However, achieving high-quality, child-friendly speech generation remains challenging, particularly for low-resource languages across diverse languages and cultural contexts. In this paper, we propose MultiAiTutor, an educational multilingual generative AI tutor with child-friendly designs, leveraging LLM architecture for speech generation tailored for educational purposes. We propose to integrate age-appropriate multilingual speech generation using LLM architectures, facilitating young children's language learning through culturally relevant image-description tasks in three low-resource languages: Singaporean-accent Mandarin, Malay, and Tamil. Experimental results from both objective metrics and subjective evaluations demonstrate the superior performance of the proposed MultiAiTutor compared to baseline methods.

Revealing the Role of Audio Channels in ASR Performance Degradation

Pre-trained automatic speech recognition (ASR) models have demonstrated strong performance on a variety of tasks. However, their performance can degrade substantially when the input audio comes from different recording channels. While previous studies have demonstrated this phenomenon, it is often attributed to the mismatch between training and testing corpora. This study argues that variations in speech characteristics caused by different recording channels can fundamentally harm ASR performance. To address this limitation, we propose a normalization technique designed to mitigate the impact of channel variation by aligning internal feature representations in the ASR model with those derived from a clean reference channel. This approach significantly improves ASR performance on previously unseen channels and languages, highlighting its ability to generalize across channel and language differences.

Robot can reduce superior's dominance in group discussions with human social hierarchy

This study investigated whether robotic agents that deal with social hierarchical relationships can reduce the dominance of superiors and equalize participation among participants in discussions with hierarchical structures. Thirty doctors and students having hierarchical relationship were gathered as participants, and an intervention experiment was conducted using a robot that can encourage participants to speak depending on social hierarchy. These were compared with strategies that intervened equally for all participants without considering hierarchy and with a no-action. The robots performed follow actions, showing backchanneling to speech, and encourage actions, prompting speech from members with less speaking time, on the basis of the hierarchical relationships among group members to equalize participation. The experimental results revealed that the robot's actions could potentially influence the speaking time among members, but it could not be conclusively stated that there were significant differences between the robot's action conditions. However, the results suggested that it might be possible to influence speaking time without decreasing the satisfaction of superiors. This indicates that in discussion scenarios where experienced superiors are likely to dominate, controlling the robot's backchanneling behavior could potentially suppress dominance and equalize participation among group members.

Transient Noise Removal via Diffusion-based Speech Inpainting

In this paper, we present PGDI, a diffusion-based speech inpainting framework for restoring missing or severely corrupted speech segments. Unlike previous methods that struggle with speaker variability or long gap lengths, PGDI can accurately reconstruct gaps of up to one second in length while preserving speaker identity, prosody, and environmental factors such as reverberation. Central to this approach is classifier guidance, specifically phoneme-level guidance, which substantially improves reconstruction fidelity. PGDI operates in a speaker-independent manner and maintains robustness even when long segments are completely masked by strong transient noise, making it well-suited for real-world applications, such as fireworks, door slams, hammer strikes, and construction noise. Through extensive experiments across diverse speakers and gap lengths, we demonstrate PGDI's superior inpainting performance and its ability to handle challenging acoustic conditions. We consider both scenarios, with and without access to the transcript during inference, showing that while the availability of text further enhances performance, the model remains effective even in its absence. For audio samples, visit: https://mordehaym.github.io/PGDI/

Selection of Layers from Self-supervised Learning Models for Predicting Mean-Opinion-Score of Speech

Self-supervised learning (SSL) models like Wav2Vec2, HuBERT, and WavLM have been widely used in speech processing. These transformer-based models consist of multiple layers, each capturing different levels of representation. While prior studies explored their layer-wise representations for efficiency and performance, speech quality assessment (SQA) models predominantly rely on last-layer features, leaving intermediate layers underexamined. In this work, we systematically evaluate different layers of multiple SSL models for predicting mean-opinion-score (MOS). Features from each layer are fed into a lightweight regression network to assess effectiveness. Our experiments consistently show early-layers features outperform or match those from the last layer, leading to significant improvements over conventional approaches and state-of-the-art MOS prediction models. These findings highlight the advantages of early-layer selection, offering enhanced performance and reduced system complexity.

DeCRED: Decoder-Centric Regularization for Encoder-Decoder Based Speech Recognition

This paper presents a simple yet effective regularization for the internal language model induced by the decoder in encoder-decoder ASR models, thereby improving robustness and generalization in both in- and out-of-domain settings. The proposed method, Decoder-Centric Regularization in Encoder-Decoder (DeCRED), adds auxiliary classifiers to the decoder, enabling next token prediction via intermediate logits. Empirically, DeCRED reduces the mean internal LM BPE perplexity by 36.6% relative to 11 test sets. Furthermore, this translates into actual WER improvements over the baseline in 5 of 7 in-domain and 3 of 4 out-of-domain test sets, reducing macro WER from 6.4% to 6.3% and 18.2% to 16.2%, respectively. On TEDLIUM3, DeCRED achieves 7.0% WER, surpassing the baseline and encoder-centric InterCTC regularization by 0.6% and 0.5%, respectively. Finally, we compare DeCRED with OWSM v3.1 and Whisper-medium, showing competitive WERs despite training on much less data with fewer parameters.