Large Audio-Language Models show consistent performance gains across speech and audio benchmarks, yet high scores may not reflect true auditory perception. If a model can answer questions without processing the acoustic signal, the benchmark fails as a measure of auditory understanding. We present a diagnostic framework using two axes: text prior, which measures answerability from text and general knowledge alone, and audio reliance, which assesses actual dependency on the acoustic signal. Evaluating eight LALMs across three benchmarks, we find that models retain 60-72% of their full audio scores even without any audio input. Moreover, among items that require audio, only 3.0-4.2% need the complete audio clip; the majority can be resolved using localized fragments. These findings challenge the assumption that benchmark performance equals robust audio understanding, and we conclude with practical guidelines for improving evaluation reliability and benchmark design.

Automatic chord recognition (ACR) extracts time-aligned chord labels from music audio recordings. Despite recent advances, ACR still struggles with oversegmentation, data scarcity, and imbalance, especially in recognizing complex chords such as non-triads, which are unpopular in existing datasets. To address these challenges, we reformulate ACR as a segment-level sequence-to-sequence prediction task, where chord sequences are predicted auto-regressively rather than frame by frame. This design mitigates excessive segmentation by detecting chord changes only at segment boundaries. We further introduce two types of token representations and an encoder pre-training method, both specifically designed for time-aligned chord modeling. Experimental results show that our model improves performance in both chord recognition and segmentation, with notable gains for complex and infrequent chord types. These findings demonstrate the effectiveness of segment-level sequence modeling, structured tokenization, and representation learning for advancing chord recognition systems.

Fine-grained local timing control is still absent from modern text-to-speech systems: existing approaches typically provide only utterance-level duration or global speaking-rate control, while precise token-level timing manipulation remains unavailable. To the best of our knowledge, MAGIC-TTS is the first TTS model with explicit local timing control over token-level content duration and pause. MAGIC-TTS is enabled by explicit token-level duration conditioning, carefully prepared high-confidence duration supervision, and training mechanisms that correct zero-value bias and make the model robust to missing local controls. On our timing-control benchmark, MAGIC-TTS substantially improves token-level duration and pause following over spontaneous synthesis. Even when no timing control is provided, MAGIC-TTS maintains natural high-quality synthesis. We further evaluate practical local editing with a scenario-based benchmark covering navigation guidance, guided reading, and accessibility-oriented code reading. In this setting, MAGIC-TTS realizes a reproducible uniform-timing baseline and then moves the edited regions toward the requested local targets with low mean bias. These results show that explicit fine-grained controllability can be implemented effectively in a high-quality TTS system and can support realistic local timing-editing applications.

RuASD (Russian AntiSpoofing Dataset) is a dedicated, reproducible benchmark for Russian-language speech anti-spoofing designed to evaluate both in-domain discrimination and robustness to deployment-style distribution shifts. It combines a large spoof subset synthesized using 37 modern Russian-capable TTS and voice-cloning systems with a bona fide subset curated from multiple heterogeneous open Russian speech corpora, enabling systematic evaluation across diverse data sources. To emulate typical dissemination and channel effects in a controlled and reproducible manner, RuASD includes configurable simulations of platform and transmission distortions, including room reverberation, additive noise/music, and a range of speech-codec transcodings implemented via a unified processing chain. We benchmark a diverse set of publicly available anti-spoofing countermeasures spanning lightweight supervised architectures, graph-attention models, SSL-based detectors, and large-scale pretrained systems, and report reference results on both clean and simulated conditions to characterize robustness under realistic perturbation pipelines. The dataset is publickly available at \href{https://huggingface.co/datasets/MTUCI/RuASD}{\underline{Hugging Face}} and \href{https://modelscope.cn/datasets/lab260/RuASD}{\underline{ModelScope}}.

Deploying Automatic Speech Recognition (ASR) models on memory-constrained edge devices requires aggressive low-bit weight quantization. Layer-wise post-training quantization is practical and effective, but it suffers from cross-layer error accumulation. Existing compensation methods typically use a single global strength for all layers, which is ill-suited to encoder-decoder ASR models whose acoustic encoder and linguistic decoder exhibit markedly different sensitivities to quantization noise. We propose FADE, a diagnostic-driven framework that assigns each layer an adaptive compensation coefficient by combining two complementary signals: an intrinsic vulnerability score from weight geometry and a calibration reliability score from the data-driven solution. The resulting layer-wise coefficient balances local quantization fidelity against cross-layer error correction, enabling tailored compensation without retraining or hyperparameter search. Experiments on Whisper, Moonshine, and Qwen3-ASR across four benchmarks show that FADE consistently improves mean Word Error Rate over strong baselines at both 3- and 4-bit precision while substantially reducing run-to-run variance.

With the development of large-scale diffusion-based and language-modeling-based generative models, impressive progress has been achieved in text-to-audio generation. Despite producing high-quality outputs, existing text-to-audio models mainly aim to generate semantically aligned sound and fall short of controlling fine-grained acoustic characteristics of specific sounds. As a result, users who need specific sound content may find it difficult to generate the desired audio clips. In this paper, we present DreamAudio for customized text-to-audio generation (CTTA). Specifically, we introduce a new framework that is designed to enable the model to identify auditory information from user-provided reference concepts for audio generation. Given a few reference audio samples containing personalized audio events, our system can generate new audio samples that include these specific events. In addition, two types of datasets are developed for training and testing the proposed systems. The experiments show that DreamAudio generates audio samples that are highly consistent with the customized audio features and aligned well with the input text prompts. Furthermore, DreamAudio offers comparable performance in general text-to-audio tasks. We also provide a human-involved dataset containing audio events from real-world CTTA cases as the benchmark for customized generation tasks.

Anglicisms are a challenge in German speech recognition. Due to their irregular pronunciation compared to native German words, automatically generated pronunciation dictionaries often include faulty phoneme sequences for Anglicisms. In this work, we propose a multitask sequence-to-sequence approach for grapheme-to-phoneme conversion to improve the phonetization of Anglicisms. We extended a grapheme-to-phoneme model with a classifier to distinguish Anglicisms from native German words. With this approach, the model learns to generate pronunciations differently depending on the classification result. We used our model to create supplementary Anglicism pronunciation dictionaries that are added to an existing German speech recognition model. Tested on a dedicated Anglicism evaluation set, we improved the recognition of Anglicisms compared to a baseline model, reducing the word error rate by 1 % and the Anglicism error rate by 3 %. We show that multitask learning can help solving the challenge of Anglicisms in German speech recognition.

With the rapid advancement of speech generation technologies, the threat posed by speech deepfakes in real-time communication (RTC) scenarios has intensified. However, existing detection studies mainly focus on offline simulations and struggle to cope with the complex distortions introduced during RTC transmission, including unknown speech enhancement processes (e.g., noise suppression) and codec compression. To address this challenge, we present the first large-scale speech deepfake dataset tailored for RTC scenarios, termed \textit{RTCFake}, totaling approximately 600 hours. The dataset is constructed by transmitting speech through multiple mainstream social media and conferencing platforms (e.g., Zoom), enabling precise pairing between offline and online speech. In addition, we propose a phoneme-guided consistency learning (PCL) strategy that enforces models to learn platform-invariant semantic structural representations. In this paper, the RTCFake dataset is divided into training, development, and evaluation sets. The evaluation set further includes both unseen RTC platforms and unseen complex noise conditions, thereby providing a more realistic and challenging evaluation benchmark for speech deepfake detection. Furthermore, the proposed PCL strategy achieves significant improvements in both cross-platform generalization and noise robustness, offering an effective and generalizable modeling paradigm. The \textit{RTCFake} dataset is provided in the {https://huggingface.co/datasets/JunXueTech/RTCFake}.

Recent large audio language models (LALMs) demonstrate remarkable capabilities in processing extended multi-modal sequences, yet incur high inference costs. Token compression is an effective method that directly reduces redundant tokens in the sequence. Existing compression methods usually assume that all attention heads in LALMs contribute equally to various audio tasks and calculate token importance by averaging scores across all heads. However, our analysis demonstrates that attention heads exhibit distinct behaviors across diverse audio domains. We further reveal that only a sparse subset of attention heads actively responds to audio, with completely different performance when handling semantic and acoustic tasks. In light of this observation, we propose HeadRouter, a head-importance-aware token pruning method that perceives the varying importance of attention heads in different audio tasks to maximize the retention of crucial tokens. HeadRouter is training-free and can be applied to various LALMs. Extensive experiments on the AudioMarathon and MMAU-Pro benchmarks demonstrate that HeadRouter achieves state-of-the-art compression performance, exceeding the baseline model even when retaining 70% of the audio tokens and achieving 101.8% and 103.0% of the vanilla average on Qwen2.5-Omni-3B and Qwen2.5-Omni-7B, respectively.

Real-time text-driven joint audio-video avatar generation requires jointly synthesizing portrait video and speech with high fidelity and precise synchronization, yet existing audio-visual diffusion models remain too slow for interactive use and often degrade noticeably after aggressive acceleration. We present Hallo-Live, a streaming framework for joint audio-visual avatar generation that combines asynchronous dual-stream diffusion with human-centric preference-guided distillation. To reduce articulation lag in causal generation, we introduce Future-Expanding Attention, which allows each video block to access synchronous audio together with a short horizon of future phonetic cues. To mitigate the quality loss of few-step distillation, we further propose Human-Centric Preference-Guided DMD (HP-DMD), which reweights training samples using rewards from visual fidelity, speech naturalness, and audio-visual synchronization. On two NVIDIA H200 GPUs, Hallo-Live runs at 20.38 FPS with 0.94 seconds latency, yielding 16.0x higher throughput and 99.3x lower latency than the teacher model Ovi. Despite this speedup, it retains strong generation quality, reaching comparable VideoAlign overall score and Sync Confidence score while outperforming other accelerated baselines in the overall quality-efficiency trade-off. Qualitative results further show robust generalization across photorealistic, multi-speaker, and stylized scenarios. To the best of our knowledge, Hallo-Live is the first framework to combine streaming dual-stream diffusion with preference-guided distillation for real-time, text-driven audio-visual generation.

Joint audio-video generation models have shown that unified generation yields stronger cross-modal coherence than cascaded approaches. However, existing models couple modalities throughout denoising via pervasive attention, treating high-level semantics and low-level details in a fully entangled manner. This is suboptimal for talking head synthesis: while audio and facial motion are semantically correlated, their low-level realizations (acoustic signals and visual textures) follow distinct rendering processes. Enforcing joint modeling across all levels causes unnecessary entanglement and reduces efficiency. We propose Talker-T2AV, an autoregressive diffusion framework where high-level cross-modal modeling occurs in a shared backbone, while low-level refinement uses modality-specific decoders. A shared autoregressive language model jointly reasons over audio and video in a unified patch-level token space. Two lightweight diffusion transformer heads decode the hidden states into frame-level audio and video latents. Experiments on talking portrait benchmarks show Talker-T2AV outperforms dual-branch baselines in lip-sync accuracy, video quality, and audio quality, achieving stronger cross-modal consistency than cascaded pipelines.

Machine generation of symbolic music and digital audio are hot topics but there have been relatively few digital musical instruments that integrate generative AI. Present musical AI tools are not artist centred and do not support experimentation or integrating into musical instruments or practices. This work introduces an inexpensive generative AI instrument platform based on a single board computer that connects via MIDI to other musical devices. The platform uses artist-collected datasets with models trained on a regular computer. This paper asks what the design space of intelligent musical instruments might look like when accessible and portable AI systems are available for artistic exploration. I contribute five examples of instruments created and tested through a two-year first-person artistic research process. These show that (re)mapping can replace retraining for discovering AI interaction, that fast input interleaving is a new co-creative strategy, that small-data AI models can be a transportable design resource, and that cheap hardware can lower barriers to inclusion. This work could enable artists to explore new interaction and performance schemes with intelligent musical instruments.

We test whether Speech Articulatory Coding (SPARC) features can linearly predict surface electromyography (sEMG) envelopes across aloud, mimed, and subvocal speech in twenty-four subjects. Using elastic-net multivariate temporal response function (mTRF) with sentence-level cross-validation, SPARC yields higher prediction accuracy than phoneme one-hot representations on nearly all electrodes and in all speech modes. Aloud and mimed speech perform comparably, and subvocal speech remains above chance, indicating detectable articulatory activity. Variance partitioning shows a substantial unique contribution from SPARC and a minimal unique contribution from phoneme features. mTRF weight patterns reveal anatomically interpretable relationships between electrode sites and articulatory movements that remain consistent across modes. This study focuses on representation/encoding analysis (not end-to-end decoding) and supports SPARC as a robust and interpretable intermediate target for sEMG-based silent-speech modeling.

Recent progress in multimodal models has spurred rapid advances in audio understanding, generation, and editing. However, these capabilities are typically addressed by specialized models, leaving the development of a truly unified framework that can seamlessly integrate all three tasks underexplored. While some pioneering works have explored unifying audio understanding and generation, they often remain confined to specific domains. To address this, we introduce Audio-Omni, the first end-to-end framework to unify generation and editing across general sound, music, and speech domains, with integrated multi-modal understanding capabilities. Our architecture synergizes a frozen Multimodal Large Language Model for high-level reasoning with a trainable Diffusion Transformer for high-fidelity synthesis. To overcome the critical data scarcity in audio editing, we construct AudioEdit, a new large-scale dataset comprising over one million meticulously curated editing pairs. Extensive experiments demonstrate that Audio-Omni achieves state-of-the-art performance across a suite of benchmarks, outperforming prior unified approaches while achieving performance on par with or superior to specialized expert models. Beyond its core capabilities, Audio-Omni exhibits remarkable inherited capabilities, including knowledge-augmented reasoning generation, in-context generation, and zero-shot cross-lingual control for audio generation, highlighting a promising direction toward universal generative audio intelligence. The code, model, and dataset will be publicly released on https://zeyuet.github.io/Audio-Omni.

Recent advances in AudioLLMs have enabled spoken dialogue systems to move beyond turn-based interaction toward real-time full-duplex communication, where the agent must decide when to speak, yield, or interrupt while the user is still talking. Existing full-duplex approaches either rely on voice activity cues, which lack semantic understanding, or on ASR-based modules, which introduce latency and degrade under overlapping speech and noise. Moreover, available datasets rarely capture realistic interaction dynamics, limiting evaluation and deployment. To mitigate the problem, we propose \textbf{FastTurn}, a unified framework for low-latency and robust turn detection. To advance latency while maintaining performance, FastTurn combines streaming CTC decoding with acoustic features, enabling early decisions from partial observations while preserving semantic cues. We also release a test set based on real human dialogue, capturing authentic turn transitions, overlapping speech, backchannels, pauses, pitch variation, and environmental noise. Experiments show FastTurn achieves higher decision accuracy with lower interruption latency than representative baselines and remains robust under challenging acoustic conditions, demonstrating its effectiveness for practical full-duplex dialogue systems.

Aligning text-to-speech (TTS) system outputs with human feedback through preference optimization has been shown to effectively improve the robustness and naturalness of language model-based TTS models. Current approaches primarily require paired desirable and undesirable samples at the utterance level. However, such pairs are often limited in TTS output data, and utterance-level formulation prevents fine-grained token-level optimization needed for accurate pronunciation alignment. In this study, we propose TKTO that eliminates the need for paired data, enabling a more data-efficient training paradigm, and directly targets token-level units, automatically providing fine-grained alignment signals without token-level annotations. TKTO improves the challenging Japanese TTS accuracy by 39% and reduces CER by 54%, automatically assigning 12.8 times stronger reward to targeted tokens.

Large audio-language models (LALMs) unify speech and text processing, but their robustness in noisy real-world settings remains underexplored. We investigate how irrelevant audio, such as silence, synthetic noise, and environmental sounds, affects text reasoning tasks where audio is unnecessary. Across three text-based benchmarks, we find that even non-informative audio reduces accuracy and increases prediction volatility; the severity of interference scales with longer durations, higher amplitudes, and elevated decoding temperatures. Silence, often assumed neutral, destabilizes outputs as strongly as synthetic noise. While larger models show greater resilience, vulnerabilities persist across all evaluated systems. We further test mitigation strategies and find that prompting shows limited effectiveness, whereas self-consistency improves stability at the cost of increased computation. Our results reveal cross-modal interference as a key robustness challenge and highlight the need for efficient fusion strategies that preserve reasoning performance in the presence of irrelevant inputs.

Text-guided sound separation enables flexible audio editing, assistive listening, and open-domain source extraction, but systems such as AudioSep remain too expensive for low-latency edge or codec-mediated deployment. Existing neural audio codec separators are efficient, yet largely restricted to fixed stems or closed taxonomies. We introduce CodecSep, a prompt-driven universal sound separation framework that extracts sources directly in neural audio codec latent space. CodecSep combines a frozen DAC backbone with a lightweight FiLM-conditioned Transformer masker driven by CLAP text embeddings, enabling open-vocabulary separation while preserving codec-native efficiency. Across dnr-v2 and five open-domain benchmarks, CodecSep consistently improves over AudioSep in SI-SDR, remains competitive in ViSQOL, and achieves clear gains in human MOS-LQS. Controlled analyses show that fine-grained prompts outperform coarse labels, and that explicit latent masking is substantially more effective than decoder-style latent generation in codec space. Qualitative diagnostics show that neural audio codec latents retain source-dependent structure, which CodecSep exploits mainly through channel-wise source-conditioned modulation. CodecSep also provides a practical code-stream deployment path. When audio is transmitted as neural audio codec codes, CodecSep maps codes to embeddings, separates directly in codec space, and outputs waveforms or re-quantized codes, avoiding the decode-separate-re-encode loop. In this regime, CodecSep requires only 1.35 GMACs end-to-end: about 54 times less compute than AudioSep in the same pipeline and 25 times lower separator-only compute, with much lower latency and memory. More broadly, CodecSep offers a blueprint for codec-native downstream audio processing.

We present a system for non-intrusive prediction of speech quality in noisy and enhanced speech, developed for Track 3 of the VoiceMOS 2024 Challenge. The task required estimating the ITU-T P.835 metrics SIG, BAK, and OVRL without reference signals and with only 100 subjectively labeled utterances for training. Our approach uses wav2vec 2.0 with a two-stage transfer learning strategy: initial fine-tuning on automatically labeled noisy data, followed by adaptation to the challenge data. The system achieved the best performance on BAK prediction (LCC=0.867) and a very close second place in OVRL (LCC=0.711) in the official evaluation. Post-challenge experiments show that adding artificially degraded data to the first fine-tuning stage substantially improves SIG prediction, raising correlation with ground truth scores from 0.207 to 0.516. These results demonstrate that transfer learning with targeted data generation is effective for predicting P.835 scores under severe data constraints.

With advancements in large audio-language models (LALMs), which enhance large language models (LLMs) with auditory capabilities, these models are expected to demonstrate universal proficiency across various auditory tasks. While numerous benchmarks have emerged to assess LALMs' performance, they remain fragmented and lack a structured taxonomy. To bridge this gap, we conduct a comprehensive survey and propose a systematic taxonomy for LALM evaluations, categorizing them into four dimensions based on their objectives: (1) General Auditory Awareness and Processing, (2) Knowledge and Reasoning, (3) Dialogue-oriented Ability, and (4) Fairness, Safety, and Trustworthiness. We provide detailed overviews within each category and highlight challenges in this field, offering insights into promising future directions. To the best of our knowledge, this is the first survey specifically focused on the evaluations of LALMs, providing clear guidelines for the community. We will release the collection of the surveyed papers and actively maintain it to support ongoing advancements in the field.

Audio-text retrieval enables semantic alignment between audio content and natural language queries, supporting applications in multimedia search, accessibility, and surveillance. However, current state-of-the-art approaches struggle with long, noisy, and weakly labeled audio due to their reliance on contrastive learning and large-batch training. We propose a novel multimodal retrieval framework that refines audio and text embeddings using a cross-modal embedding refinement module combining transformer-based projection, linear mapping, and bidirectional attention. To further improve robustness, we introduce a hybrid loss function blending cosine similarity, $\mathcal{L}_{1}$, and contrastive objectives, enabling stable training even under small-batch constraints. Our approach efficiently handles long-form and noisy audio (SNR 5 to 15) via silence-aware chunking and attention-based pooling. Experiments on benchmark datasets demonstrate improvements over prior methods.

Human-imitated speech poses a greater challenge than AI-generated speech for both human listeners and automatic detection systems. Unlike AI-generated speech, which often contains artifacts, over-smoothed spectra, or robotic cues, imitated speech is produced naturally by humans, thereby preserving a higher degree of naturalness that makes imitation-based speech forgery significantly more challenging to detect using conventional acoustic or cepstral features. To overcome this challenge, this study proposes an auditory perception-based Spectro-Temporal Modulation (STM) representation framework for human-imitated speech detection. The STM representations are derived from two cochlear filterbank models: the Gammatone Filterbank (GTFB), which simulates frequency selectivity and can be regarded as a first approximation of cochlear filtering, and the Gammachirp Filterbank (GCFB), which further models both frequency selectivity and level-dependent asymmetry. These STM representations jointly capture temporal and spectral fluctuations in speech signals, corresponding to changes over time in the spectrogram and variations along the frequency axis related to human auditory perception. We also introduce a Segmental-STM representation to analyze short-term modulation patterns across overlapping time windows, enabling high-resolution modeling of temporal speech variations. Experimental results show that STM representations are effective for human-imitated speech detection, achieving accuracy levels close to those of human listeners. In addition, Segmental-STM representations are more effective, surpassing human perceptual performance. The findings demonstrate that perceptually inspired spectro-temporal modeling is promising for detecting imitation-based speech attacks and improving voice authentication robustness.

Statistical modeling of popular music presents a unique challenge due to the complexity of song structures, which cannot be easily analyzed using conventional statistical tools. However, recent advances in data science have shown that converting non-standard data objects into real vector-valued embeddings enables meaningful statistical analysis. In this work, we demonstrate an approach based on logistic principal component analysis to construct embeddings from global song features, allowing for standard multivariate analysis. We apply this method to a corpus of Lennon and McCartney songs from 1962-1966, using embeddings derived from chords, melodic notes, chord and pitch transitions, and melodic contours. Our analysis explores how these song embeddings cluster by Beatles album, how songwriting styles evolved over time, and whether Lennon and McCartney's compositions exhibited convergence or divergence. This embedding-based approach offers a powerful framework for statistically examining musical structure and stylistic development in popular music.

Recent advances in speech-aware language models have coupled strong acoustic encoders with large language models, enabling systems that move beyond transcription to produce richer outputs. Among these, word-level timestamp prediction is critical for applications such as captioning, media search, and multimodal synchronization, yet it is often handled by external alignment tools. In this work, we extend an existing speech-aware language model to predict timestamps directly alongside transcripts. We introduce a set of novel lightweight training strategies that improve alignment robustness while preserving recognition quality. Experiments across multiple datasets show that these strategies not only enhance timestamp accuracy, but also yield gains in overall ASR performance. Together, they demonstrate an efficient and unified approach to speech recognition with precise timestamp prediction.

The Filtered-x Normalized Least Mean Square (FxNLMS) algorithm suffers from slow convergence and a risk of divergence, although it can achieve low steady-state errors after sufficient adaptation. In contrast, the Generative Fixed-Filter Active Noise Control (GFANC) method offers fast response speed, but its lack of adaptability may lead to large steady-state errors. This paper proposes a hybrid GFANC-FxNLMS algorithm to leverage the complementary advantages of both approaches. In the hybrid GFANC-FxNLMS algorithm, GFANC provides a frame-level control filter as an initialization for FxNLMS, while FxNLMS performs continuous adaptation at the sampling rate. Small variations in the GFANC-generated filter may repeatedly reinitialize FxNLMS, interrupting its adaptation process and destabilizing the system. An online clustering module is introduced to avoid unnecessary re-initializations and improve system stability. Simulation results show that the proposed algorithm achieves fast response, very low steady-state error, and high stability, requiring only one pre-trained broadband filter.

As Large Language Models (LLMs) expand beyond text, integrating speech as a native modality has given rise to SpeechLLMs, which directly process spoken language and enable speech-to-text translation (ST) and other downstream tasks, bypassing traditional transcription-based pipelines. Whether this integration improves ST quality over established cascaded architectures, however, remains an open question. We present Hearing to Translate, the first comprehensive test suite rigorously benchmarking 6 state-of-the-art SpeechLLMs against 16 strong direct and cascade systems that couple leading speech foundation models (SFM), with multilingual LLMs. Our analysis spans 16 benchmarks, 13 language pairs, and 9 challenging conditions, including disfluent, noisy, and long-form speech. Across this extensive evaluation, we find that cascaded systems remain the most reliable solution overall, but most recent SpeechLLMs can match or even outperform cascades in various settings while SFMs lag behind both, highlighting that integrating an LLM, either within the model or in a pipeline, is essential for high-quality speech translation.

Omni-modal large language models (OLLMs) offer a promising end-to-end solution for slide-enhanced speech recognition due to their inherent multimodal capabilities. However, we found a fundamental issue faced by OLLMs: \textit{Visual Interference}, where models show a bias towards visible text over auditory signals, causing them to hallucinate slide content that was never spoken. To address this, we propose Visually-Anchored Policy Optimization (VAPO), which aims to reshape models' inference process to follow the human-like ``Look-then-Listen'' inference chain. Specifically, we design a temporally decoupled policy: the model first extracts visual priors in a <think> block to serve as semantic anchors, then generates the transcription in an <answer> block. The policy is optimized via multi-objective reinforcement learning. Furthermore, we introduce SlideASR-Bench, a comprehensive benchmark designed to address the scarcity of entity-rich data, comprising a large-scale synthetic corpus for training and a challenging real-world test set for evaluation. We conduct extensive evaluations demonstrating that VAPO effectively eliminates visual interference and achieves state-of-the-art performance on SlideASR-Bench and public datasets, significantly reducing entity recognition errors in specialized domains.

Fine-tuning speech representation models can enhance performance on specific tasks but often compromises their cross-task generalization ability. This degradation is often caused by excessive changes in the representations, making it difficult to retain information learned during pre-training. Existing approaches, such as regularizing weight changes during fine-tuning, may fail to maintain sufficiently high feature similarity with the pre-trained model, and thus could possibly lose cross-task generalization. To address this issue, we propose Speech-FT, a novel two-stage fine-tuning framework designed to maintain cross-task generalization while benefiting from fine-tuning. Speech-FT first applies fine-tuning specifically designed to reduce representational drift, followed by weight-space interpolation with the pre-trained model to restore cross-task generalization. Extensive experiments on HuBERT, wav2vec 2.0, DeCoAR 2.0, and WavLM Base+ demonstrate that Speech-FT consistently improves performance across a wide range of supervised, unsupervised, and multitask fine-tuning scenarios. Moreover, Speech-FT achieves superior cross-task generalization compared to fine-tuning baselines that explicitly constrain weight changes, such as weight-space regularization and LoRA fine-tuning. Our analysis reveals that Speech-FT maintains higher feature similarity to the pre-trained model compared to alternative strategies, despite allowing larger weight-space updates. Notably, Speech-FT achieves significant improvements on the SUPERB benchmark. For example, when fine-tuning HuBERT on automatic speech recognition, Speech-FT is able to reduce phone error rate from 5.17% to 3.94%, lower word error rate from 6.38% to 5.75%, and increase speaker identification accuracy from 81.86% to 84.11%. Speech-FT provides a simple yet powerful solution for further refining speech representation models after pre-training.

In this work, we propose an approach to music source separation that uses a generative diffusion model as a last-stage refinement on top of a deterministic separator, progressively enhancing the separated sources through iterative denoising. While the diffusion refinement yields measurable quality gains, it requires iterative steps at inference, increasing computational cost. To speed up the inference process, we apply consistency distillation, reducing inference to a single step while maintaining quality; with two or more steps, the distilled model even surpasses the diffusion-based approach. Crucially, our method is architecture-agnostic: we demonstrate state-of-the-art results when applied to both a custom U-Net-based separator on Slakh2100 and the state-of-the-art BS-RoFormer model on MUSDB18, showing that the refinement generalizes across backbone architectures. Sound examples are available at: https://consistency-separation.github.io/.

Target Speaker Extraction (TSE) aims to extract the clean speech of the target speaker in an audio mixture, eliminating irrelevant background noise and speech. While prior work has explored various auxiliary cues including pre-recorded speech, visual information, and spatial information, the acquisition and selection of such strong cues are infeasible in many practical scenarios. Differently, in this paper, we condition the TSE algorithm on semantic cues extracted from limited and unaligned text contents, such as condensed points from a presentation slide. This method is particularly useful in scenarios like meetings, poster sessions, or lecture presentations, where acquiring other cues in real time may be challenging. To this end, we design two different networks. Specifically, our proposed Text Prompt Extractor Network (TPE) fuses audio features with content-based semantic cues to facilitate time-frequency mask generation to filter out extraneous noise. The experimental results show the efficacy in accurately extracting the target speaker's speech by utilizing semantic cues derived from limited and unaligned text, resulting in SI-SDRi of 12.16 dB, SDRi of 12.66 dB, PESQi of 0.830 and STOIi of 0.150.