Human infants, with only a few hundred hours of speech exposure, acquire basic units of new languages, highlighting a striking efficiency gap compared to the data-hungry self-supervised speech models. To address this gap, this paper introduces SpidR-Adapt for rapid adaptation of speech units to new languages using minimal unlabeled data. We cast such low-resource speech representation learning as a meta-learning problem and construct a multi-task adaptive pre-training (MAdaPT) protocol which formulates the adaptation process as a bi-level optimization framework. To enable scalable meta-training under this framework, we propose a novel heuristic solution, first-order bi-level optimization (FOBLO), avoiding heavy computation costs. Finally, we stabilize meta-training by using a robust initialization through interleaved supervision which alternates self-supervised and supervised objectives. Empirically, SpidR-Adapt achieves rapid gains in phonemic discriminability (ABX) and downstream spoken language modeling scores (sWUGGY, sBLIMP, tSC), surpassing in-domain toplines after training on less than 1h of target-language audio and delivering 100\times greater data efficiency than standard multi-task training.. These findings highlight a practical, architecture-agnostic path toward biologically inspired, data-efficient representations. We open-source the training code and model checkpoints at https://github.com/facebookresearch/spidr-adapt.
论文检索
输入标题、作者或关键词，从 2,480 篇学术成果中精准定位
Learning speech representations that are useful for a variety of downstream tasks has received considerable attention, due to the outstanding properties of Self-Supervised Learning (SSL) trained models. Despite advancements in modelling methods, understanding the difference in task performance on representations is limited. Mainly motivated by the no-free-lunch theorem and speech production, this work investigates changes in task performance in sparse speech representations, providing interpretability analysis under the Information Bottleneck (IB) framework. Autoencoders with varying sparsity levels were trained using three SSL features, and evaluated on six tasks of SUPERB: Speech Enhancement (SE), Speaker Identification (SID), Speech Emotion Recognition (SER), Phone Recognition (PR), Automatic Speech Recognition (ASR) and Slot Filling (SF). Experiments show that: 1) different tasks manifest different degrees of sensitivity to the sparsity levels; 2) the optimal sparsity level for task performance varies; 3) the choice of SSL features has a limited impact on most tasks but with an exception of PR; 4) overall PR and ASR require more preservation of relevant information about the labels, while SID and SER demand more compression of irrelevant information, where the input quality can shift this trade-off to some degree. These findings can contribute to the design of a universal sparse speech representation learner.
UniSonate: A Unified Model for Speech, Music, and Sound Effect Generation with Text Instructions
PDF ↗Generative audio modeling has largely been fragmented into specialized tasks, text-to-speech (TTS), text-to-music (TTM), and text-to-audio (TTA), each operating under heterogeneous control paradigms. Unifying these modalities remains a fundamental challenge due to the intrinsic dissonance between structured semantic representations (speech/music) and unstructured acoustic textures (sound effects). In this paper, we introduce UniSonate, a unified flow-matching framework capable of synthesizing speech, music, and sound effects through a standardized, reference-free natural language instruction interface. To reconcile structural disparities, we propose a novel dynamic token injection mechanism that projects unstructured environmental sounds into a structured temporal latent space, enabling precise duration control within a phoneme-driven Multimodal Diffusion Transformer (MM-DiT). Coupled with a multi-stage curriculum learning strategy, this approach effectively mitigates cross-modal optimization conflicts. Extensive experiments demonstrate that UniSonate achieves state-of-the-art performance in instruction-based TTS (WER 1.47%) and TTM (SongEval Coherence 3.18), while maintaining competitive fidelity in TTA. Crucially, we observe positive transfer, where joint training on diverse audio data significantly enhances structural coherence and prosodic expressiveness compared to single-task baselines.
High-quality, large-scale audio captioning is crucial for advancing audio understanding, yet current automated methods often generate captions that lack fine-grained detail and contextual accuracy, primarily due to their reliance on limited unimodal or superficial multimodal information. Drawing inspiration from human auditory perception, which adeptly integrates cross-modal cues and performs sophisticated auditory scene analysis, we introduce a novel two-stage automated pipeline. This pipeline first employs specialized pretrained models to extract diverse contextual cues (e.g., speech, music, general sounds, and visual information from associated video). A large language model (LLM) then synthesizes these rich, multimodal inputs to generate detailed and context-aware audio captions. Key contributions of this work include: (1) the proposed scalable method for fine-grained audio caption generation; (2) FusionAudio, a new large-scale dataset comprising 1.2 million such detailed captions, combined with 6 million QA pairs; and (3) enhanced audio models developed using FusionAudio, specifically a CLAP-based audio encoder with superior audio-text alignment and instruction following. This paper paves the way for more nuanced and accurate automated understanding of complex audio environments.
Multimodal foundation models that integrate audio, vision, and language achieve strong performance on reasoning and generation tasks, yet their robustness to adversarial manipulation remains poorly understood. We study a realistic and underexplored threat model: **untargeted, audio-only adversarial attacks** on trimodal audio–video–language models. We analyze six complementary attack objectives that target different stages of multimodal processing, including audio encoder representations, cross-modal attention, hidden states, and output likelihoods. Across four state-of-the-art models and multiple benchmarks, we show that audio-only perturbations can induce severe multimodal failures, achieving up to **96% attack success rate.** We further show that attacks can be successful at low perceptual distortions (LPIPS \leq 0.08, SI-SNR \geq 0 dB) and benefit more from extended optimization than increased data scale. We evaluate the feasibility of these attacks under physically realistic conditions by incorporating room impulse response (RIR) modeling, showing that audio-only perturbations remain effective under environmental transformations and thus highlight the practical risk of single-modality attacks in real-world multimodal systems. Transferability across models and encoders remains limited, while speech recognition systems such as Whisper primarily respond to perturbation magnitude, achieving **>97% attack success** under severe distortion. These results expose a previously overlooked single-modality attack surface in multimodal systems and motivate defenses that enforce cross-modal consistency. Our project website is available at https://aafiya-h.github.io/soundbreak/.
The rise of Large Audio-Language Models (LAMs) brings both potential and risks, as their audio outputs may contain harmful or unethical content. However, current research lacks a systematic, quantitative evaluation of LAM safety, especially against jailbreak attacks, which are challenging due to the temporal and semantic nature of speech. To bridge this gap, we introduce AJailBench, the first benchmark specifically designed to evaluate jailbreak vulnerabilities in LAMs. We begin by constructing -Base, a dataset of 1,495 adversarial audio prompts spanning 10 policy-violating categories. Using this dataset, we evaluate several state-of-the-art LAMs and reveal that none exhibit consistent robustness across attacks. To further strengthen jailbreak testing and simulate more realistic attack conditions, we propose a method to generate dynamic adversarial variants. Our Audio Perturbation Toolkit (APT) applies targeted distortions across time, frequency, and amplitude domains. To preserve the original jailbreak intent, we enforce a semantic consistency constraint and employ Bayesian optimization to efficiently search for perturbations that are both subtle and highly effective. This results in AJailBench-APT+, an extended dataset of optimized adversarial audio samples. Our findings demonstrate that even small, semantically preserved perturbations can significantly reduce the safety performance of leading LAMs, underscoring the need for more robust and semantically aware defense mechanisms. We release AJailBench to facilitate future research: https://anonymous.4open.science/r/AudioJailbreak-4262/
Accented speech remains a persistent challenge for automatic speech recognition (ASR), as most models are trained on data dominated by a few high-resource English varieties, leading to substantial performance degradation for other accents. Accent-agnostic approaches improve robustness yet struggle with heavily accented or unseen varieties, while accent-specific methods rely on limited and often noisy labels. We introduce MoE-CTC, a Mixture-of-Experts architecture with intermediate CTC supervision that jointly promotes expert specialization and generalization. During training, accent-aware routing encourages experts to capture accent-specific patterns, which gradually transitions to label-free routing for inference. Each expert is equipped with its own CTC head to align routing with transcription quality, and a routing-augmented loss further stabilizes optimization. Experiments on the MCV-Accent benchmark demonstrate consistent gains across both seen and unseen accents in low- and high-resource conditions, achieving up to 29.3% relative WER reduction over strong FastConformer baselines.
LLM-ForcedAligner: A Non-Autoregressive and Accurate LLM-Based Forced Aligner for Multilingual and Long-Form Speech
PDF ↗Forced alignment (FA) predicts start and end timestamps for words or characters in speech, but existing methods are language-specific and prone to cumulative temporal shifts. The multilingual speech understanding and long-sequence processing abilities of speech large language models (SLLMs) make them promising for FA in multilingual, crosslingual, and long-form speech settings. However, directly applying the next-token prediction paradigm of SLLMs to FA results in hallucinations and slow inference. To bridge the gap, we propose LLM-ForcedAligner, reformulating FA as a slot-filling paradigm: timestamps are treated as discrete indices, and special timestamp tokens are inserted as slots into the transcript. Conditioned on the speech embeddings and the transcript with slots, the SLLM directly predicts the time indices at slots. During training, causal attention masking with non-shifted input and label sequences allows each slot to predict its own timestamp index based on itself and preceding context, with loss computed only at slot positions. Dynamic slot insertion enables FA at arbitrary positions. Moreover, non-autoregressive inference is supported, avoiding hallucinations and improving speed. Experiments across multilingual, crosslingual, and long-form speech scenarios show that LLM-ForcedAligner achieves a 69% 78% relative reduction in accumulated averaging shift compared with prior methods. The checkpoint and inference code will be released later.
Detecting hallucinations in Retrieval-Augmented Generation remains a challenge. Prior approaches attribute hallucinations to a binary conflict between internal knowledge stored in FFNs and the retrieved context. However, this perspective is incomplete, failing to account for the impact of other components of the LLM, such as the user query, previously generated tokens, the self token, and the final LayerNorm adjustment. To comprehensively capture the impact of these components on hallucination detection, we propose TPA which mathematically attributes each token’s probability to seven distinct sources: Query, RAG Context, Past Token, Self Token, FFN, Final LayerNorm, and Initial Embedding. This attribution quantifies how each source contributes to the generation of the next token. Specifically, we aggregate these attribution scores by Part-of-Speech (POS) tags to quantify the contribution of each model component to the generation of specific linguistic categories within a response. By leveraging these patterns, such as detecting anomalies where Nouns rely heavily on LayerNorm, TPA effectively identifies hallucinated responses. Extensive experiments show that TPA achieves state-of-the-art performance.
What Do Prosody and Text Convey? Characterizing How Meaningful Information is Distributed Across Multiple Channels
PDF ↗Prosody—the melody of speech—conveys critical information often not captured by the words or text of a message.In this paper, we propose an information-theoretic approach to quantify how much is conveyed by prosody that is not recoverable from text alone, and, crucially, what prosody conveys.Our approach applies large speech and language models to estimate the mutual information between a particular dimension of an utterance’s meaning (e.g., its emotion) and any of its communication channels (e.g., audio or text).We then use this approach to quantify the information conveyed by audio and text about sarcasm, emotion, and questionhood, using speech from television and podcasts.We find that for sarcasm and emotion, the audio channel, and by implication the prosodic channel, transmits over an order of magnitude more information about these features than the text channel alone, at least when long-term context beyond the current sentence is unavailable.For questionhood, prosody provides comparatively less additional information.We conclude by outlining a program applying our approach to more dimensions of meaning, communication channels, and languages.
TED-TTS: Training-Free Intra-Utterance Emotion and Duration Control for Text-to-Speech Synthesis
PDF ↗While controllable Text-to-Speech (TTS) has achieved notable progress, most existing methods remain limited to inter-utterance-level control, making fine-grained intra-utterance expression challenging due to their reliance on non-public datasets or complex multi-stage training. In this paper, we propose TED-TTS, a training-free controllable framework for pretrained zero-shot TTS to enable intra-utterance emotion and duration expression. Specifically, we propose a segment-aware emotion conditioning strategy that combines causal masking with monotonic stream alignment filtering to isolate emotion conditioning and schedule mask transitions, enabling smooth intra-utterance emotion shifts while preserving global semantic coherence. Based on this, we further propose a segment-aware duration steering strategy to combine local duration embedding steering with global EOS logit modulation, allowing local duration adjustment while ensuring globally consistent termination. To eliminate the need for segment-level manual prompt engineering, we construct a 30,000-sample multi-emotion and duration-annotated text dataset to enable LLM-based automatic prompt construction. Extensive experiments demonstrate that our training-free method not only achieves state-of-the-art intra-utterance consistency in multi-emotion and duration control, but also maintains baseline-level speech quality of the underlying TTS model. Audio samples are available at https://aclanonymous111.github.io/TED-TTS-DemoPage/.
Comparative Analysis of the Intrinsic Metrics for Tokenizers and their effect on Downstream Tasks for Hindi and Marathi
PDF ↗Various studies have pointed out that the performance of language models is poor in non-English or non-European languages. One of the factors affecting this performance is the effectiveness and suitability of the tokenization scheme used in the model. Indic scripts require multiple Unicode codepoints to represent a single visual unit to be encoded in the standard UTF-8 scheme. This paper investigates the effect of multiple tokenizers that use UTF-8 text input on the downstream performance of pretrained language models for Hindi and Marathi, languages written in \textit{Devanāgari} script. We present the intrinsic performance of the tokenizers using Fertility, Rényi Efficiency and Percentile Frequency, and report the extrinsic performance of monolingual and multilingual models on question-answering tasks, using an automated parts-of-speech and sentence similarity based evaluation framework, and on word-level tasks such as grapheme-to-phoneme conversion and transliteration. We propose a grapheme cluster tokenizer for the script which shows performance better than or competitive with other popular tokenizers. We also find that the Rényi Efficiency metric is highly correlated to downstream performance on question answering.
Audio carries richer information than text, including emotion, speaker traits, and environmental context, while also enabling lower-latency processing compared to speech-to-text pipelines. However, recent multimodal information retrieval research has predominantly focused on images, largely overlooking audio, especially in the setting of interleaved audio-text contextual retrieval. In this work, we introduce the Audio-Text Interleaved contextual Retrieval (ATIR) task, where queries can alternate between audio and text modalities. We construct an ATIR benchmark by integrating several Automatic Speech Recognition (ASR), QA, and retrieval datasets, ultimately unifying four types of contextual retrieval tasks. This benchmark substantially addresses the limitations of existing audio retrieval datasets in semantic retrieval. To study this task, we evaluate several off-the-shelf retrievers and train our ATIR model based on a Multimodal Large Language Model (MLLM). We further propose a novel token compression mechanism, which is orthogonal to existing compression methods, to mitigate the challenge of excessive audio tokens in MLLM-based ATIR models. Experimental results show that our ATIR model achieves significant improvements over strong baselines.
Rectifying the Emotional Flow: Aligning Priors and Dynamic Guidance for High-Arousal Text-to-Speech
PDF ↗While diffusion and flow-matching models have advanced TTS, generating high-arousal emotions remains a persistent challenge due to the trade-off between stability and expressiveness. Existing systems often suffer from linguistic collapse when pursuing high intensity or fail to meet target emotional levels under stable settings. In this work, we identify that standard Gaussian initialization inevitably introduces a neutral prosody bias, while uniform Classifier-Free Guidance often distorts the acoustic manifold, leading to artifacts. To address this, we propose an inference framework that rectifies the emotional trajectory. An Emotion-Rectified Noise Prior injects a semantic gradient at initialization to align sampling with the target emotional manifold, and Likelihood-Inverse Guidance adaptively schedules guidance via a conditional/unconditional likelihood ratio, strengthening guidance only when the trajectory drifts toward a neutral fallback. Extensive experiments demonstrate that our method effectively resolves the stability bottleneck in high-intensity scenarios, achieving superior linguistic accuracy and emotional fidelity without model retraining. Audio samples are available at https://showtts.github.io/emotionTTS/.
The global deployment of Large Language Models (LLMs) underscores the urgent need to evaluate their cultural alignment. However, assessing genuine "cultural awareness" across modalities (text, vision, speech) and languages remains a significant challenge. To comprehensively investigate this domain, we propose MMAC, a systematic framework that encompasses a tri-modally aligned cultural benchmark creation pipeline and a five-dimensional evaluation protocol to assess cross-country awareness disparities, evaluate cross-lingual and cross-modal consistency, and verify cultural knowledge generalization and grounding validity. Given the prevailing Western cultural bias in current models, we focus on 8 Asian countries as our dataset foundation to more acutely reveal potential cultural deficiencies in LLMs. Our dataset, MMAC-bench, features 27,000 human-curated questions across 10 languages. Crucially, it is the first dataset aligned at the input level across text, image, and speech, enabling direct cross-modal transfer tests. Each question consists of multiple-choice options accompanied by open-ended generated explanations, where 79% require multi-step reasoning grounded in cultural context, moving beyond simple memorization. We probe the causes of modal divergence, offering insights into fostering culturally robust MLLMs.
Many recent studies have shown that the perception of speech can be decoded from brain signals and subsequently reconstructed as continuous language. However, there is a lack of neurological basis for how the semantic information embedded within brain signals can be used more effectively to guide language reconstruction. Predictive coding theory suggests the human brain naturally engages in continuously predicting future words that span multiple timescales. This implies that the decoding of brain signals could potentially be associated with a predictable future. To explore the predictive coding theory within the context of language reconstruction, this paper proposes PredFT (FMRI-to-Text decoding with Predictive coding). PredFT consists of a main network and a side network. The side network obtains brain predictive representation from related regions of interest (ROIs) with a self-attention module. The representation is then fused into the main network for continuous language decoding. Experiments on two naturalistic language comprehension fMRI datasets show that PredFT outperforms current decoding models on several evaluation metrics.
SpeakerSleuth: Can Large Audio-Language Models Judge Speaker Consistency across Multi-turn Dialogues?
PDF ↗Large Audio-Language Models (LALMs) as judges have emerged as a prominent approach for evaluating speech generation quality, yet their ability to assess speaker consistency across multi-turn dialogues remains unexplored. We present SpeakerSleuth, a benchmark evaluating whether LALMs can reliably judge speaker consistency across multi-turn dialogues through three tasks reflecting real-world requirements. We construct 1,818 human-verified evaluation instances across four diverse datasets spanning synthetic and real speech, with controlled acoustic difficulty. Evaluating twelve widely-used LALMs, we find that models struggle to reliably detect acoustic inconsistencies. For instance, given audio samples of the same speaker’s turns, some models overpredict inconsistency, whereas others are overly lenient. Models further struggle to identify the exact turns that are problematic. When other interlocutors’ turns are provided as textual context, performance degrades dramatically as models prioritize textual coherence over acoustic cues, failing to detect even obvious gender switches for a speaker. On the other hand, models perform substantially better in comparing and ranking acoustic variants, demonstrating inherent acoustic discrimination capabilities. These findings expose a significant bias in LALMs: they tend to prioritize text over acoustics, revealing fundamental modality imbalances that need to be addressed to build reliable audio-language judges. Our code and data are available at https://github.com/holi-lab/SpeakerSleuth.
SEE: Signal Embedding Energy for Quantifying Noise Interference in Large Audio Language Models
PDF ↗Large Audio Language Models (LALMs) have been widely applied in real-time scenarios, such as in-car assistants and online meeting comprehension. In practice, audio inputs are often corrupted by device and environmental noise, leading to performance degradation. However, existing LALM studies on noise lack quantitative analysis and rely mainly on intuition and empirical observation, thus failing to understand practical robustness. To address this issue, we introduce Signal Embedding Energy (SEE), a method for quantifying the impact of noise intensity on LALM inputs, enabling the differentiation of LALM robustness in real-world deployments. SEE introduces a perspective based on structured activation subspaces derived from the model’s internal representations, which more accurately captures its perception of noise than raw audio features. Across experiments, SEE exhibits a strong correlation with LALM performance, achieving a correlation of 0.98. Surprisingly, traditional audio denoising methods are only marginally effective for LALMs, and, in some cases, even increase SEE and impair performance. This suggests a mismatch between speech-centric denoising objectives and the noise sensitivity of modern LALMs. Therefore, we propose a mitigation strategy derived from SEE to denoise LALM inputs, outperforming existing denoising methods. This paper introduces a novel metric for noise quantification in LALMs, providing guidance for robustness improvements in real-world deployments.
Codec-based language models (LMs) have revolutionized text-to-speech (TTS). However, standard codecs entangle timbre and prosody, which hinders independent control in continuation-based LMs. To tackle this challenge, we propose DisCo-Speech, a zero-shot controllable TTS framework featuring a disentangled speech codec (DisCodec) and an LM-based generator. The core component DisCodec employs a two-stage design: 1) tri-factor disentanglement to separate speech into content, prosody, and timbre subspaces via parallel encoders and hybrid losses; and 2) fusion and reconstruction that merges content and prosody into unified content-prosody tokens suitable for LM prediction, while jointly optimizing reconstruction to address the disentanglement-reconstruction trade-off. This allows the LM to perform prosodic continuation from a style prompt while the decoder injects target timbre, enabling flexible zero-shot control. Experiments demonstrate that DisCo-Speech achieves competitive voice cloning and superior zero-shot prosody control. By resolving the core entanglement at the codec level, DisCo-Speech provides a robust foundation for controllable speech synthesis. Audio samples are available at: https://disco-speech.github.io/DisCo-demo/. Code and weights will be released at: https://github.com/disco-speech/DisCo-Speech upon acceptance.
Multi-speaker automatic speech recognition (MASR) aims to predict ”who spoke when and what” from multi-speaker speech, a key technology for multi-party dialogue understanding. However, most existing approaches decouple temporal modeling and speaker modeling when addressing ”when” and ”who”: some inject speaker cues before encoding (e.g., speaker masking), which can cause irreversible information loss; others fuse identity by mixing speaker posteriors after encoding, which may entangle acoustic content with speaker identity. This separation is brittle under rapid turn-taking and overlapping speech, often leading to degraded performance. To address these limitations, we propose TellWhisper, a unified framework that jointly models speaker identity and temporal within the speech encoder. Specifically, we design TS\text{-}RoPE, a time-speaker rotary positional encoding: time coordinates are derived from frame indices, while speaker coordinates are derived from speaker activity and pause cues. By applying region-specific rotation angles, the model explicitly captures per-speaker continuity, speaker-turn transitions, and state dynamics, enabling the attention mechanism to simultaneously attend to ”when” and ”who”. Moreover, to estimate frame-level speaker activity, we develop Hyper\text{-}SD, which casts speaker classification in hyperbolic space to enhance inter-class separation and refine speaker-activity estimates. Extensive experiments demonstrate the effectiveness of the proposed approach. The project webpage is available at https://walker-hyf.github.io/TellWhisper.