This work presents a framework for compressing self-supervised models for speaker diarization through structured pruning guided by knowledge distillation. We investigate pruning objectives that target reducing both model parameters and computational complexity, where knowledge distillation enables compact models and pruning removes unnecessary parameters to improve hardware efficiency. We further analyze alternative pruning strategies, showing that a simple overall pruning approach provides the best balance between efficiency and accuracy. Compared to the original unpruned model, our method achieves up to 80% model size reduction and 4x faster inference without performance degradation. Comprehensive experiments across eight public diarization datasets demonstrate that the pruned models consistently match or surpass the performance of their uncompressed counterparts. Furthermore, we show strong out-of-domain generalization on the CHiME-6 dataset, achieving accuracy comparable to the top systems in the CHiME-7 challenge without any domain adaptation. These results highlight that structured pruning, when guided by distillation, can yield efficient and generalizable diarization systems suitable for real-world applications.

Speaker-attributed automatic speech recognition (ASR) in multi-speaker environments remains a significant challenge, particularly when systems conditioned on speaker embeddings fail to generalize to unseen speakers. In this work, we propose Diarization-Conditioned Whisper (DiCoW), a novel approach to target-speaker ASR that leverages speaker diarization outputs as conditioning information. DiCoW extends the pre-trained Whisper model by integrating diarization labels directly, eliminating reliance on speaker embeddings and reducing the need for extensive speaker-specific training data. Our method introduces frame-level diarization-dependent transformations (FDDT) and query-key biasing (QKb) techniques to refine the model's focus on target speakers while effectively handling overlapping speech. By leveraging diarization outputs as conditioning signals, DiCoW simplifies the workflow for multi-speaker ASR, improves generalization to unseen speakers and enables more reliable transcription in real-world multi-speaker recordings. Additionally, we explore the integration of a connectionist temporal classification (CTC) head to Whisper and demonstrate its ability to improve transcription efficiency through hybrid decoding. Notably, we show that our approach is not limited to Whisper; it also provides similar benefits when applied to the Branchformer model. We validate DiCoW on real-world datasets, including AMI and NOTSOFAR-1 from CHiME-8 challenge, as well as synthetic benchmarks such as Libri2Mix and LibriCSS, enabling direct comparisons with previous methods. Results demonstrate that DiCoW enhances the model's target-speaker ASR capabilities while maintaining Whisper's accuracy and robustness on single-speaker data.

Modern speaker verification (SV) systems rely on speaker embeddings that are effective but difficult to interpret or query in natural language. Most existing speech-text corpora target controllable synthesis or utterance-level captioning, offering limited speaker-level supervision for in-the-wild speaker recognition. This paper introduces SpeakerCard-1M, a bilingual speaker resource for evidence-grounded SV, derived from VoxCeleb1/2 and CN-Celeb1/2, where the “-1M” suffix refers to the 1.78M utterance-level captions contained in the release. We adopt a tool-first, LLM-last approach in which ten acoustic probes produce field-level evidence, the evidence is aggregated into speaker profiles under a schema that separates relatively stable traits from utterance-level states, and bilingual Speaker Cards are rendered by a constrained LLM that sees only the structured fields. The release includes 56.7k Speaker Card records over 10.2k speakers, 1.78M utterance-level captions, and speaker-ID-disjoint hard-negative triplets. We further define two SV-oriented cross-modal protocols, bidirectional Speaker-Text Retrieval (T2S-R / S2T-R) and Attribute-Conditioned Verification (AC-Verify), and compare a dual-encoder baseline against recent audio language models under a zero-shot forced-choice setting. Joint audio-text training costs only 0.31

Multi-channel speaker verification (SV), employing numerous microphones for capturing enrollment and/or test recordings, gained attention for its benefits in far-field scenarios. While some studies approach the problem by designing multi-channel embedding extractors, we focus on building and thoroughly analyzing a framework integrating beamforming pre-processing paired with single-channel embedding extraction. This strategy benefits from accommodating both multi-channel and single-channel inputs. Furthermore, it provides human-interpretable intermediate output — enhanced speech — that can be independently evaluated and related to SV performance. We first focus on the front-end, taking advantage of deep-learning source separation for direct or indirect mask estimation required by the beamformer. We alternate single-channel network architectures, subsequently extended to multi-channel ones by reference channel attention (RCA). We also analyze the impact of beamformer and network output fusion. Finally, we show improvements brought by end-to-end fine-tuning the entire architecture facilitated by our newly designed multi-channel corpus, MultiSV2, extending our previous MultiSV dataset.

We propose a speaker-attributed (SA) Whisper-based model for multi-talker speech recognition that combines target-speaker modeling with serialized output training (SOT). Our approach leverages a Diarization-Conditioned Whisper (DiCoW) encoder to extract target-speaker embeddings, which are concatenated into a single representation and passed to a shared decoder. This enables the model to transcribe overlapping speech as a serialized output stream with speaker tags and timestamps. In contrast to target-speaker ASR systems such as DiCoW, which decode each speaker separately, our approach performs joint decoding, allowing the decoder to condition on the context of all speakers simultaneously. Experiments show that the model outperforms existing SOT-based approaches and surpasses DiCoW on multi-talker mixtures (e.g., LibriMix).

Abstract Reliable assessment of the abilities of large audio language models (LALMs) is essential to advancing the state of the art. As benchmarks rapidly evolve to incorporate complex reasoning and subjective tasks, they increasingly necessitate open-ended responses from LALMs. We present Open-ended Response Correctness Assessment (ORCA)—a reliable and lightweight model-based approach for answer correctness and disagreement modeling. We employ a three-stage annotation pipeline combining human judgment, structured feedback, and human-AI correction, yielding 9,663 annotations across 3,699 question-answer pairs from 15 LALMs on three audio understanding and reasoning benchmarks (achieving a Krippendorff’s alpha of 0.82). Our experiments employing curriculum learning show that ORCA models achieve a Spearman correlation of 0.91 with average human correctness ratings on seen benchmarks and generalize to unseen benchmarks with a score of 0.85, outperforming several LLM judge baselines including Gemini 2.5 Flash. Furthermore, we demonstrate that ORCA’s predicted variance correlates strongly with human disagreement, allowing it to effectively identify problematic benchmark items.

While low-latency interaction is critical for spoken dialogue, cascaded architectures are often bottlenecked by reactive turn-completion detection. We propose Endpoint Anticipation, shifting from reactive detection to proactive forecasting of end-of-turn signals. Our speech-based model anticipates endpoints upto 2.56 seconds in advance, enabling speculative execution of LLM and TTS pipelines on partial context. We introduce metrics to quantify the trade-off between realized latency reduction and computational redundancy. Evaluation across conversational and task-oriented datasets shows our model consistently outperforms competitive VAP-based baselines. Integration with the Unmute framework demonstrates a 505 ms average latency reduction with a 28.4

Although large-scale self-supervised learning (SSL) models like WavLM have achieved state-of-the-art performance in speech processing, their significant size impedes deployment on resource-constrained devices. While structured pruning is a key technique for model compression, existing methods typically separate it from task-specific fine-tuning. This multi-stage approach struggles to create optimal architectures tailored for diverse downstream tasks. In this work, we introduce a unified framework that integrates structured pruning into the downstream fine-tuning process. Our framework unifies these steps, jointly optimizing for task performance and model sparsity in a single stage. This allows the model to learn a compressed architecture specifically for the end task, eliminating the need for complex multi-stage pipelines and knowledge distillation. Our pruned models achieve up to a 70% parameter reduction with negligible performance degradation on large-scale datasets, achieving equal error rates of 0.7%, 0.8%, and 1.6% on Vox1-O, -E, and -H, respectively. Furthermore, our approach demonstrates improved generalization in low-resource scenarios, reducing overfitting and achieving a state-of-the-art 3.7% EER on ASVspoof5.

Despite achieving state-of-the-art accuracy in speaker verification (SV), large-scale self-supervised learning (SSL) speech models remain difficult to deploy on edge devices because of their computational and memory demands. Existing compression approaches improve efficiency through pruning and quantization, but usually optimize them sequentially and thus overlook their interaction. In this paper, we propose HPQ, a framework that jointly optimizes differentiable structured pruning and learnable quantization in a single fine-tuning stage. By integrating $L_{0}$ regularization with Learned Step Size Quantization (LSQ) into a unified objective, HPQ enables the network to co-adapt its architecture to quantization noise. Experiments on VoxCeleb demonstrate that HPQ establishes a new Pareto frontier: an 8-bit, 70% sparse WavLM model achieves a $13\times$ reduction in model size and a $15\times$ reduction in bit-operations, with only a 0.23% absolute EER degradation compared with the full-precision baseline. We further show that larger backbones and more diverse training data improve robustness under aggressive compression across both WavLM and W2V-BERT.

We propose diarization-conditioned spoken language models (SLMs), a strategy for extending SLMs to far-field multi-talker audio. Rather than adapting the decoder via Serialized Output Training, which risks catastrophic forgetting, we condition the acoustic encoder on diarization masks to extract target-speaker representations, keeping the decoder frozen. We instantiate this as Dixtral, integrating a Diarization Conditioned Whisper (DiCoW) encoder into the Voxtral SLM. On AMI, NOTSOFAR-1, LibriSpeechMix, and Mixer6, Dixtral outperforms Gemini 3.0 Flash, VibeVoice, and Voxtral Mini Transcribe V2 on speaker-attributed transcription by 29.0%, 19.8%, and 16.0% absolute cpWER respectively. On a novel long-form multi-speaker QA benchmark, zero-shot Dixtral matches Gemini on far-field content understanding, and when fine-tuned surpasses both Gemini and Voxtral operating on close-talk across all tasks.

Speaker-attributed automatic speech recognition (ASR) in multi-speaker environments remains a major challenge. While some approaches achieve strong performance when fine-tuned on specific domains, few systems generalize well across out-of-domain datasets. Our prior work, Diarization-Conditioned Whisper (DiCoW), leverages speaker diarization outputs as conditioning information and, with minimal fine-tuning, demonstrated strong multilingual and multi-domain performance. In this paper, we address a key limitation of DiCoW: ambiguity in Silence-Target-Non-target-Overlap (STNO) masks, where two or more fully overlapping speakers may have nearly identical conditioning despite differing transcriptions. We introduce SE-DiCoW (Self-Enrolled Diarization-Conditioned Whisper), which uses diarization output to locate an enrollment segment anywhere in the conversation where the target speaker is most active. This enrollment segment is used as fixed conditioning via cross-attention at each encoder layer. We further refine DiCoW with improved data segmentation, model initialization, and augmentation. Together, these advances yield substantial gains: SE-DiCoW reduces macro-averaged tcpWER by 52.4

Recent breakthroughs in multi-talker ASR (MT-ASR) and speaker diarization (SD) rely on synthetic data to mitigate the scarcity of large-scale conversational recordings, yet the impact of specific simulation choices remains poorly understood. To mind the gap between simulated mixtures and real-world interactions, we present a study of synthetic data generation for leading MT-ASR (DiCoW) and SD (Sortformer) systems. By introducing FastMSS, a highly efficient open-source simulator, we analyze turn-taking dynamics, source domain, acoustic augmentation, and data mixing strategies. Our findings reveal that optimal simulation recipes are highly task-dependent: increasing speech overlap benefits ASR but degrades diarization. Furthermore, broad source diversity consistently outperforms exact domain matching. Ultimately, synthetic-only training approaches real-data baselines, and combining simulated data with real recordings yields substantial gains over real-only training across both tasks.

This work presents a framework for compressing self-supervised models for speaker diarization through structured pruning guided by knowledge distillation. We investigate pruning objectives that target both model parameters and computational complexity, and analyze alternative strategies, showing that a simple overall pruning approach provides the best balance between efficiency and accuracy. Our method achieves up to 80% model size reduction and 4x faster inference without performance degradation. Comprehensive experiments across eight public diarization datasets demonstrate that the pruned models consistently match or surpass the performance of their uncompressed counterparts. Furthermore, we show strong out-of-domain generalization on the CHiME-6 dataset, achieving accuracy comparable to the top systems in the CHiME-7 challenge without any domain adaptation. These results highlight that structured pruning, when guided by distillation, can yield efficient and generalizable diarization systems suitable for real-world applications.

Self-supervised learning (SSL) models like WavLM can be effectively utilized when building speaker diarization systems but are often large and slow, limiting their use in resource constrained scenarios. Previous studies have explored compression techniques, but usually for the price of degraded performance at high pruning ratios. In this work, we propose to compress SSL models through structured pruning by introducing knowledge distillation. Different from the existing works, we emphasize the importance of fine-tuning SSL models before pruning. Experiments on far-field single-channel AMI, AISHELL-4, and AliMeeting datasets show that our method can remove redundant parameters of WavLM Base+ and WavLM Large by up to 80% without any performance degradation. After pruning, the inference speeds on a single GPU for the Base+ and Large models are 4.0 and 2.6 times faster, respectively. Our source code is publicly available.

Self-supervised learning (SSL) models for speaker verification (SV) have gained significant attention in recent years. However, existing SSL-based SV systems often struggle to capture local temporal dependencies and generalize across different tasks. In this paper, we propose context-aware multi-head factorized attentive pooling (CA-MHFA), a lightweight framework that incorporates contextual information from surrounding frames. CA-MHFA leverages grouped, learnable queries to effectively model contextual dependencies while maintaining efficiency by sharing keys and values across groups. Experimental results on the VoxCeleb dataset show that CA-MHFA achieves EERs of 0.42%, 0.48%, and 0.96% on Vox1-0, Voxl-E, and Voxl-H, respectively, outperforming complex models like WavLM-TDNN with fewer parameters and faster convergence. Additionally, CA-MHFA demonstrates strong generalization across multiple SSL models and tasks, including emotion recognition and anti-spoofing, highlighting its robustness and versatility.(1)

Self-supervised learning (SSL) models have significantly advanced speech processing tasks, and several benchmarks have been proposed to validate their effectiveness. However, previous benchmarks have primarily focused on single-speaker scenarios, with less exploration of target-speaker tasks in noisy, multi-talker conditions—a more challenging yet practical case. In this paper, we introduce the Target-Speaker Speech Processing Universal Performance Benchmark (TS-SUPERB), which includes four widely recognized target-speaker processing tasks that require identifying the target speaker and extracting information from the speech mixture. In our benchmark, the speaker embedding extracted from enrollment speech is used as a clue to condition downstream models. The benchmark result reveals the importance of evaluating SSL models in target speaker scenarios, demonstrating that performance cannot be easily inferred from related single-speaker tasks. Moreover, by using a unified SSL-based target speech encoder, consisting of a speaker encoder and an extractor module, we also investigate joint optimization across TS tasks to leverage mutual information and demonstrate its effectiveness. 1

End-to-end spoken dialogue state tracking (DST) is made difficult by the tandem of having to handle speech input and data scarcity. Combining speech foundation encoders and large language models has been proposed in recent work as to alleviate some of this difficulty. Although this approach has been shown to result in strong spoken DST models, achieving state-of-the-art performance in realistic multi-turn DST, it struggles to generalize across domains and requires annotated spoken DST training data for each domain of interest. However, collecting such data for every target domain is both costly and difficult. Noting that textual DST data is more easily obtained for various domains, in this work, we propose jointly training on available spoken DST data and written textual data from other domains as a way to achieve cross-domain generalization. We conduct experiments which show the efficacy of our proposed method for getting good cross-domain DST performance without relying on spoken training data from the target domains.

We present a two-speaker automatic speech recognition (ASR) system that combines DiCoW – a diarization-conditioned variant of Whisper – with DiariZen, a diarization pipeline built on top of Pyannote. We first evaluate both systems in out-of-domain (OOD) multilingual scenarios without any fine-tuning. In this scenario, DiariZen consistently outperforms the baseline Pyannote diarization model, demonstrating strong generalization. Despite being fine-tuned on English-only data for target-speaker ASR, DiCoW retains solid multilingual performance, indicating that encoder modifications preserve Whisper's multilingual capabilities. We then fine-tune both DiCoW and DiariZen on the MLC-SLM challenge data. The fine-tuned DiariZen continues to outperform the fine-tuned Pyannote baseline, while DiCoW sees further gains from domain adaptation. Our final system achieves a micro-average tcpWER/CER of 16.75 ranks second in Task 2 of the MLC-SLM challenge. Lastly, we identify several labeling inconsistencies in the training data – such as missing speech segments and incorrect silence annotations – which can hinder diarization fine-tuning. We propose simple mitigation strategies to address these issues and improve system robustness.

This paper describes the BUT submission to the ESDD 2026 Challenge, specifically focusing on Track 1: Environmental Sound Deepfake Detection with Unseen Generators. To address the critical challenge of generalizing to audio generated by unseen synthesis algorithms, we propose a robust ensemble framework leveraging diverse Self-Supervised Learning (SSL) models. We conduct a comprehensive analysis of general audio SSL models (including BEATs, EAT, and Dasheng) and speech-specific SSLs. These front-ends are coupled with a lightweight Multi-Head Factorized Attention (MHFA) back-end to capture discriminative representations. Furthermore, we introduce a feature domain augmentation strategy based on distribution uncertainty modeling to enhance model robustness against unseen spectral distortions. All models are trained exclusively on the official EnvSDD data, without using any external resources. Experimental results demonstrate the effectiveness of our approach: our best single system achieved Equal Error Rates (EER) of 0.00%, 4.60%, and 4.80% on the Development, Progress (Track 1), and Final Evaluation sets, respectively. The fusion system further improved generalization, yielding EERs of 0.00%, 3.52%, and 4.38% across the same partitions.

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