Nima Mesgarani

Xilin Jiang, Yinghao Aaron Li, Adrian Nicolas Florea et al.

It is too early to conclude that Mamba is a better alternative to transformers for speech before comparing Mamba with transformers in terms of both performance and efficiency in multiple speech-related tasks. To reach this conclusion, we propose and evaluate three models for three tasks: Mamba-TasNet for speech separation, ConMamba for speech recognition, and VALL-M for speech synthesis. We compare them with transformers of similar sizes in performance, memory, and speed. Our Mamba or Mamba-transformer hybrid models show comparable or higher performance than their transformer counterparts: Sepformer, Conformer, and VALL-E. They are more efficient than transformers in memory and speed for speech longer than a threshold duration, inversely related to the resolution of a speech token. Mamba for separation is the most efficient, and Mamba for recognition is the least. Further, we show that Mamba is not more efficient than transformer for speech shorter than the threshold duration and performs worse in models that require joint modeling of text and speech, such as cross or masked attention of two inputs. Therefore, we argue that the superiority of Mamba or transformer depends on particular problems and models. Code available at https://github.com/xi-j/Mamba-TasNet and https://github.com/xi-j/Mamba-ASR.

Yinghao Aaron Li, Cong Han, Vinay S. Raghavan et al.

In this paper, we present StyleTTS 2, a text-to-speech (TTS) model that leverages style diffusion and adversarial training with large speech language models (SLMs) to achieve human-level TTS synthesis. StyleTTS 2 differs from its predecessor by modeling styles as a latent random variable through diffusion models to generate the most suitable style for the text without requiring reference speech, achieving efficient latent diffusion while benefiting from the diverse speech synthesis offered by diffusion models. Furthermore, we employ large pre-trained SLMs, such as WavLM, as discriminators with our novel differentiable duration modeling for end-to-end training, resulting in improved speech naturalness. StyleTTS 2 surpasses human recordings on the single-speaker LJSpeech dataset and matches it on the multispeaker VCTK dataset as judged by native English speakers. Moreover, when trained on the LibriTTS dataset, our model outperforms previous publicly available models for zero-shot speaker adaptation. This work achieves the first human-level TTS on both single and multispeaker datasets, showcasing the potential of style diffusion and adversarial training with large SLMs. The audio demos and source code are available at https://styletts2.github.io/.

Junkai Wu, Xulin Fan, Bo-Ru Lu et al. · uw

In recent years, we have observed a rapid advancement in speech language models (SpeechLLMs), catching up with humans' listening and reasoning abilities. SpeechLLMs have demonstrated impressive spoken dialog question-answering (SQA) performance in benchmarks like Gaokao, the English listening test of the college entrance exam in China, which seemingly requires understanding both the spoken content and voice characteristics of speakers in a conversation. However, after carefully examining Gaokao's questions, we find the correct answers to many questions can be inferred from the conversation transcript alone, i.e.\ without speaker segmentation and identification. Our evaluation of state-of-the-art models Qwen-Audio and WavLLM on both Gaokao and our proposed "What Do You Like?" dataset shows a significantly higher accuracy in these context-based questions than in identity-critical questions, which can only be answered reliably with correct speaker identification. The results and analysis suggest that when solving SQA, the current SpeechLLMs exhibit limited speaker awareness from the audio and behave similarly to an LLM reasoning from the conversation transcription without sound. We propose that tasks focused on identity-critical questions could offer a more accurate evaluation framework of SpeechLLMs in SQA.

Yinghao Aaron Li, Cong Han, Nima Mesgarani

Text-to-Speech (TTS) has recently seen great progress in synthesizing high-quality speech owing to the rapid development of parallel TTS systems, but producing speech with naturalistic prosodic variations, speaking styles and emotional tones remains challenging. Moreover, since duration and speech are generated separately, parallel TTS models still have problems finding the best monotonic alignments that are crucial for naturalistic speech synthesis. Here, we propose StyleTTS, a style-based generative model for parallel TTS that can synthesize diverse speech with natural prosody from a reference speech utterance. With novel Transferable Monotonic Aligner (TMA) and duration-invariant data augmentation schemes, our method significantly outperforms state-of-the-art models on both single and multi-speaker datasets in subjective tests of speech naturalness and speaker similarity. Through self-supervised learning of the speaking styles, our model can synthesize speech with the same prosodic and emotional tone as any given reference speech without the need for explicitly labeling these categories.

Yinghao Aaron Li, Cong Han, Xilin Jiang et al.

Large-scale pre-trained language models have been shown to be helpful in improving the naturalness of text-to-speech (TTS) models by enabling them to produce more naturalistic prosodic patterns. However, these models are usually word-level or sup-phoneme-level and jointly trained with phonemes, making them inefficient for the downstream TTS task where only phonemes are needed. In this work, we propose a phoneme-level BERT (PL-BERT) with a pretext task of predicting the corresponding graphemes along with the regular masked phoneme predictions. Subjective evaluations show that our phoneme-level BERT encoder has significantly improved the mean opinion scores (MOS) of rated naturalness of synthesized speech compared with the state-of-the-art (SOTA) StyleTTS baseline on out-of-distribution (OOD) texts.

Sukru Samet Dindar, Riki Shimizu, Xilin Jiang et al.

Empathetic spoken dialogue systems must infer a user's emotional state to respond appropriately, yet everyday speech often carries weak, neutral, or ambiguous affective cues. To address this, we introduce Sympatheia, a speech-to-speech dialogue framework conditioned on affect inferred from the user's speech and, when available, explicit affect specifications provided as a continuous valence--arousal (VA) control signal by a multimodal sensing module or user interface. To train our model, we construct Sympatheia-18k, an emotion-conditioned synthetic spoken dialogue corpus with 12 emotion anchors. This dataset includes an emotional split for learning affective speech behavior, and a neutral split that pairs emotionally neutral queries with multiple emotion-conditioned responses to isolate explicit emotion control in emotionally ambiguous cases. Empirical results show that Sympatheia outperforms speech conversational baselines in generating responses whose semantic content and spoken delivery are both emotionally appropriate. We further show that the same VA interface can integrate emotion estimates from diverse sensing modules, including facial expression, biosignals, and textual affect descriptions, improving response alignment when speech alone provides limited emotional evidence. These results suggest that continuous affect conditioning is an effective practical step for building emotionally adaptive voice assistants.

Yinghao Aaron Li, Xilin Jiang, Jordan Darefsky et al.

The rapid advancement of large language models (LLMs) has significantly propelled the development of text-based chatbots, demonstrating their capability to engage in coherent and contextually relevant dialogues. However, extending these advancements to enable end-to-end speech-to-speech conversation bots remains a formidable challenge, primarily due to the extensive dataset and computational resources required. The conventional approach of cascading automatic speech recognition (ASR), LLM, and text-to-speech (TTS) models in a pipeline, while effective, suffers from unnatural prosody because it lacks direct interactions between the input audio and its transcribed text and the output audio. These systems are also limited by their inherent latency from the ASR process for real-time applications. This paper introduces Style-Talker, an innovative framework that fine-tunes an audio LLM alongside a style-based TTS model for fast spoken dialog generation. Style-Talker takes user input audio and uses transcribed chat history and speech styles to generate both the speaking style and text for the response. Subsequently, the TTS model synthesizes the speech, which is then played back to the user. While the response speech is being played, the input speech undergoes ASR processing to extract the transcription and speaking style, serving as the context for the ensuing dialogue turn. This novel pipeline accelerates the traditional cascade ASR-LLM-TTS systems while integrating rich paralinguistic information from input speech. Our experimental results show that Style-Talker significantly outperforms the conventional cascade and speech-to-speech baselines in terms of both dialogue naturalness and coherence while being more than 50% faster.

Xilin Jiang, Cong Han, Yinghao Aaron Li et al.

In this study, we present a simple multi-channel framework for contrastive learning (MC-SimCLR) to encode 'what' and 'where' of spatial audios. MC-SimCLR learns joint spectral and spatial representations from unlabeled spatial audios, thereby enhancing both event classification and sound localization in downstream tasks. At its core, we propose a multi-level data augmentation pipeline that augments different levels of audio features, including waveforms, Mel spectrograms, and generalized cross-correlation (GCC) features. In addition, we introduce simple yet effective channel-wise augmentation methods to randomly swap the order of the microphones and mask Mel and GCC channels. By using these augmentations, we find that linear layers on top of the learned representation significantly outperform supervised models in terms of both event classification accuracy and localization error. We also perform a comprehensive analysis of the effect of each augmentation method and a comparison of the fine-tuning performance using different amounts of labeled data.

Yinghao Aaron Li, Cong Han, Xilin Jiang et al.

Recent advancements in speech synthesis have leveraged GAN-based networks like HiFi-GAN and BigVGAN to produce high-fidelity waveforms from mel-spectrograms. However, these networks are computationally expensive and parameter-heavy. iSTFTNet addresses these limitations by integrating inverse short-time Fourier transform (iSTFT) into the network, achieving both speed and parameter efficiency. In this paper, we introduce an extension to iSTFTNet, termed HiFTNet, which incorporates a harmonic-plus-noise source filter in the time-frequency domain that uses a sinusoidal source from the fundamental frequency (F0) inferred via a pre-trained F0 estimation network for fast inference speed. Subjective evaluations on LJSpeech show that our model significantly outperforms both iSTFTNet and HiFi-GAN, achieving ground-truth-level performance. HiFTNet also outperforms BigVGAN-base on LibriTTS for unseen speakers and achieves comparable performance to BigVGAN while being four times faster with only $1/6$ of the parameters. Our work sets a new benchmark for efficient, high-quality neural vocoding, paving the way for real-time applications that demand high quality speech synthesis.

Linyang He, Qiyao Yu, Hanze Dong et al.

Mathematical reasoning is a hallmark of human intelligence, and whether large language models (LLMs) can meaningfully perform it remains a central question in artificial intelligence and cognitive science. As LLMs are increasingly integrated into scientific workflows, rigorous evaluation of their mathematical capabilities becomes a practical necessity. Existing benchmarks are limited by synthetic settings and data contamination. We present LiveMathematicianBench, a dynamic multiple-choice benchmark for research-level mathematical reasoning built from recent arXiv papers published after model training cutoffs. By grounding evaluation in newly published theorems, it provides a realistic testbed beyond memorized patterns. The benchmark introduces a thirteen-category logical taxonomy of theorem types (e.g., implication, equivalence, existence, uniqueness), enabling fine-grained evaluation across reasoning forms. It employs a proof-sketch-guided distractor pipeline that uses high-level proof strategies to construct plausible but invalid answer choices reflecting misleading proof directions, increasing sensitivity to genuine understanding over surface-level matching. We also introduce a substitution-resistant mechanism to distinguish answer recognition from substantive reasoning. Evaluation shows the benchmark is far from saturated: Gemini-3.1-pro-preview, the best model, achieves only 43.5%. Under substitution-resistant evaluation, accuracy drops sharply: GPT-5.4 scores highest at 30.6%, while Gemini-3.1-pro-preview falls to 17.6%, below the 20% random baseline. A dual-mode protocol reveals that proof-sketch access yields consistent accuracy gains, suggesting models can leverage high-level proof strategies for reasoning. Overall, LiveMathematicianBench offers a scalable, contamination-resistant testbed for studying research-level mathematical reasoning in LLMs.

Tianjun Zhong, Linyang He, Nima Mesgarani

Recent progress in large language models has renewed interest in how multi-step reasoning is represented internally. While prior work often treats reasoning as a linear chain, many reasoning problems are more naturally modeled as directed acyclic graphs (DAGs), where intermediate conclusions branch, merge, and are reused. Whether such graph structure is reflected in model internals remains unclear. We introduce Reasoning DAG Probing, a framework for testing whether LLM hidden states linearly encode properties of an underlying reasoning DAG and where this structure emerges across layers. We associate each reasoning node with a textual realization and train lightweight probes to predict node depth, pairwise distance, and adjacency from hidden states. Using these probes, we analyze the emergence of DAG structure across layers, reconstruct approximate reasoning graphs, and evaluate controls that disrupt reasoning-relevant structure while preserving surface text. Across reasoning benchmarks, we find that DAG structure is meaningfully encoded in LLM representations, with recoverability peaking in intermediate layers, varying systematically by node depth, edge span, and model scale, and enabling nontrivial recovery of dependency graphs. These findings suggest that LLM reasoning is not purely sequential, but exhibits measurable internal graph structure.

Yinghao Aaron Li, Cong Han, Nima Mesgarani

In recent years, large-scale pre-trained speech language models (SLMs) have demonstrated remarkable advancements in various generative speech modeling applications, such as text-to-speech synthesis, voice conversion, and speech enhancement. These applications typically involve mapping text or speech inputs to pre-trained SLM representations, from which target speech is decoded. This paper introduces a new approach, SLMGAN, to leverage SLM representations for discriminative tasks within the generative adversarial network (GAN) framework, specifically for voice conversion. Building upon StarGANv2-VC, we add our novel SLM-based WavLM discriminators on top of the mel-based discriminators along with our newly designed SLM feature matching loss function, resulting in an unsupervised zero-shot voice conversion system that does not require text labels during training. Subjective evaluation results show that SLMGAN outperforms existing state-of-the-art zero-shot voice conversion models in terms of naturalness and achieves comparable similarity, highlighting the potential of SLM-based discriminators for related applications.

Qinhao Chen, Linyang He, Nima Mesgarani

Existing feature-interpretation pipelines typically operate on uniformly sampled units, but only a small fraction of cross-layer transcoder (CLT) features matter for a target behavior, with the rest resulting in expensive feature explaining and evaluating costs. We introduce the first CLT-native end-to-end framework, PIE, connecting Pruning, automatic Interpretation, and interpretation Evaluation, enabling systematic measurement of behavioral fidelity and downstream interpretability under pruning. To achieve this, we propose Feature Attribution Patching (FAP), a patch-grounded attribution method that scores CLT features by aggregating gradient-weighted write contributions, and FAP-Synergy, a synergy-aware reranking procedure. We evaluate pruning using KL-divergence behavior retention and assess interpretation quality with FADE-style metrics. Across IOI and Doc-String, across budgets $K \in \{50, 100, 200, 400, 800\}$, and across FAP, FAP-Synergy, Activation-Magnitude, and ACDC-style pruning, the FAP family consistently achieves the best or near-best fidelity, with FAP-Synergy providing its clearest gains in strict-budget regimes. On IOI with CLTs for Llama-3.2-1B and Gemma-2-2B, pruning to $K=100$ features matches the KL fidelity that random selection from the active feature set requires $\approx 4$k features to achieve ($\approx 40\times$ compression), enabling $\approx 40\times$ fewer interpretation/evaluation calls while substantially reducing low-quality features.

Xilin Jiang, Sukru Samet Dindar, Vishal Choudhari et al.

Auditory foundation models, including auditory large language models (LLMs), process all sound inputs equally, independent of listener perception. However, human auditory perception is inherently selective: listeners focus on specific speakers while ignoring others in complex auditory scenes. Existing models do not incorporate this selectivity, limiting their ability to generate perception-aligned responses. To address this, we introduce Intention-Informed Auditory Scene Understanding (II-ASU) and present Auditory Attention-Driven LLM (AAD-LLM), a prototype system that integrates brain signals to infer listener attention. AAD-LLM extends an auditory LLM by incorporating intracranial electroencephalography (iEEG) recordings to decode which speaker a listener is attending to and refine responses accordingly. The model first predicts the attended speaker from neural activity, then conditions response generation on this inferred attentional state. We evaluate AAD-LLM on speaker description, speech transcription and extraction, and question answering in multitalker scenarios, with both objective and subjective ratings showing improved alignment with listener intention. By taking a first step toward intention-aware auditory AI, this work explores a new paradigm where listener perception informs machine listening, paving the way for future listener-centered auditory systems. Demo and code available: https://aad-llm.github.io.

Chenxing Li, Jiaming Xu, Nima Mesgarani et al.

Most speech separation methods, trying to separate all channel sources simultaneously, are still far from having enough general- ization capabilities for real scenarios where the number of input sounds is usually uncertain and even dynamic. In this work, we employ ideas from auditory attention with two ears and propose a speaker and direction inferred speech separation network (dubbed SDNet) to solve the cocktail party problem. Specifically, our SDNet first parses out the respective perceptual representations with their speaker and direction characteristics from the mixture of the scene in a sequential manner. Then, the perceptual representations are utilized to attend to each corresponding speech. Our model gener- ates more precise perceptual representations with the help of spatial features and successfully deals with the problem of the unknown number of sources and the selection of outputs. The experiments on standard fully-overlapped speech separation benchmarks, WSJ0- 2mix, WSJ0-3mix, and WSJ0-2&3mix, show the effectiveness, and our method achieves SDR improvements of 25.31 dB, 17.26 dB, and 21.56 dB under anechoic settings. Our codes will be released at https://github.com/aispeech-lab/SDNet.

Gavin Mischler, Yinghao Aaron Li, Stephan Bickel et al.

Recent advancements in artificial intelligence have sparked interest in the parallels between large language models (LLMs) and human neural processing, particularly in language comprehension. While prior research has established similarities in the representation of LLMs and the brain, the underlying computational principles that cause this convergence, especially in the context of evolving LLMs, remain elusive. Here, we examined a diverse selection of high-performance LLMs with similar parameter sizes to investigate the factors contributing to their alignment with the brain's language processing mechanisms. We find that as LLMs achieve higher performance on benchmark tasks, they not only become more brain-like as measured by higher performance when predicting neural responses from LLM embeddings, but also their hierarchical feature extraction pathways map more closely onto the brain's while using fewer layers to do the same encoding. We also compare the feature extraction pathways of the LLMs to each other and identify new ways in which high-performing models have converged toward similar hierarchical processing mechanisms. Finally, we show the importance of contextual information in improving model performance and brain similarity. Our findings reveal the converging aspects of language processing in the brain and LLMs and offer new directions for developing models that align more closely with human cognitive processing.

Siavash Shams, Sukru Samet Dindar, Xilin Jiang et al.

Transformers have revolutionized deep learning across various tasks, including audio representation learning, due to their powerful modeling capabilities. However, they often suffer from quadratic complexity in both GPU memory usage and computational inference time, affecting their efficiency. Recently, state space models (SSMs) like Mamba have emerged as a promising alternative, offering a more efficient approach by avoiding these complexities. Given these advantages, we explore the potential of SSM-based models in audio tasks. In this paper, we introduce Self-Supervised Audio Mamba (SSAMBA), the first self-supervised, attention-free, and SSM-based model for audio representation learning. SSAMBA leverages the bidirectional Mamba to capture complex audio patterns effectively. We incorporate a self-supervised pretraining framework that optimizes both discriminative and generative objectives, enabling the model to learn robust audio representations from large-scale, unlabeled datasets. We evaluated SSAMBA on various tasks such as audio classification, keyword spotting, and speaker identification. Our results demonstrate that SSAMBA outperforms the Self-Supervised Audio Spectrogram Transformer (SSAST) in most tasks. Notably, SSAMBA is approximately 92.7% faster in batch inference speed and 95.4% more memory-efficient than SSAST for the tiny model size with an input token size of 22k. These efficiency gains, combined with superior performance, underscore the effectiveness of SSAMBA's architectural innovation, making it a compelling choice for a wide range of audio processing applications.

Xilin Jiang, Cong Han, Yinghao Aaron Li et al.

In daily life, we encounter a variety of sounds, both desirable and undesirable, with limited control over their presence and volume. Our work introduces "Listen, Chat, and Remix" (LCR), a novel multimodal sound remixer that controls each sound source in a mixture based on user-provided text instructions. LCR distinguishes itself with a user-friendly text interface and its unique ability to remix multiple sound sources simultaneously within a mixture, without needing to separate them. Users input open-vocabulary text prompts, which are interpreted by a large language model to create a semantic filter for remixing the sound mixture. The system then decomposes the mixture into its components, applies the semantic filter, and reassembles filtered components back to the desired output. We developed a 160-hour dataset with over 100k mixtures, including speech and various audio sources, along with text prompts for diverse remixing tasks including extraction, removal, and volume control of single or multiple sources. Our experiments demonstrate significant improvements in signal quality across all remixing tasks and robust performance in zero-shot scenarios with varying numbers and types of sound sources. An audio demo is available at: https://listenchatremix.github.io/demo.

Chao Huang, Yuesheng Ma, Junxuan Huang et al.

Audio source separation is fundamental for machines to understand complex acoustic environments and underpins numerous audio applications. Current supervised deep learning approaches, while powerful, are limited by the need for extensive, task-specific labeled data and struggle to generalize to the immense variability and open-set nature of real-world acoustic scenes. Inspired by the success of generative foundation models, we investigate whether pre-trained text-guided audio diffusion models can overcome these limitations. We make a surprising discovery: zero-shot source separation can be achieved purely through a pre-trained text-guided audio diffusion model under the right configuration. Our method, named ZeroSep, works by inverting the mixed audio into the diffusion model's latent space and then using text conditioning to guide the denoising process to recover individual sources. Without any task-specific training or fine-tuning, ZeroSep repurposes the generative diffusion model for a discriminative separation task and inherently supports open-set scenarios through its rich textual priors. ZeroSep is compatible with a variety of pre-trained text-guided audio diffusion backbones and delivers strong separation performance on multiple separation benchmarks, surpassing even supervised methods.

Linyang He, Ercong Nie, Sukru Samet Dindar et al.

We introduce XCOMPS in this work, a multilingual conceptual minimal pair dataset covering 17 languages. Using this dataset, we evaluate LLMs' multilingual conceptual understanding through metalinguistic prompting, direct probability measurement, and neurolinguistic probing. By comparing base, instruction-tuned, and knowledge-distilled models, we find that: 1) LLMs exhibit weaker conceptual understanding for low-resource languages, and accuracy varies across languages despite being tested on the same concept sets. 2) LLMs excel at distinguishing concept-property pairs that are visibly different but exhibit a marked performance drop when negative pairs share subtle semantic similarities. 3) Instruction tuning improves performance in concept understanding but does not enhance internal competence; knowledge distillation can enhance internal competence in conceptual understanding for low-resource languages with limited gains in explicit task performance. 4) More morphologically complex languages yield lower concept understanding scores and require deeper layers for conceptual reasoning.

Adrian Florea, Xilin Jiang, Nima Mesgarani et al.

Large language models (LLMs) for audio have excelled in recognizing and analyzing human speech, music, and environmental sounds. However, their potential for understanding other types of sounds, particularly biomedical sounds, remains largely underexplored despite significant scientific interest. In this study, we focus on diagnosing cardiovascular diseases using phonocardiograms, i.e., heart sounds. Most existing deep neural network (DNN) paradigms are restricted to heart murmur classification (healthy vs unhealthy) and do not predict other acoustic features of the murmur such as timing, grading, harshness, pitch, and quality, which are important in helping physicians diagnose the underlying heart conditions. We propose to finetune an audio LLM, Qwen2-Audio, on the PhysioNet CirCor DigiScope phonocardiogram (PCG) dataset and evaluate its performance in classifying 11 expert-labeled murmur features. Additionally, we aim to achieve more noise-robust and generalizable system by exploring a preprocessing segmentation algorithm using an audio representation model, SSAMBA. Our results indicate that the LLM-based model outperforms state-of-the-art methods in 8 of the 11 features and performs comparably in the remaining 3. Moreover, the LLM successfully classifies long-tail murmur features with limited training data, a task that all previous methods have failed to classify. These findings underscore the potential of audio LLMs as assistants to human cardiologists in enhancing heart disease diagnosis.

Xilin Jiang, Junkai Wu, Vishal Choudhari et al. · uw

Audio large language models (LLMs) are considered experts at recognizing sound objects, yet their performance relative to LLMs in other sensory modalities, such as visual or audio-visual LLMs, and to humans using their ears, eyes, or both remains unexplored. To investigate this, we systematically evaluate audio, visual, and audio-visual LLMs, specifically Qwen2-Audio, Qwen2-VL, and Qwen2.5-Omni, against humans in recognizing sound objects of different classes from audio-only, silent video, or sounded video inputs. We uncover a performance gap between Qwen2-Audio and Qwen2-VL that parallels the sensory discrepancy between human ears and eyes. To reduce this gap, we introduce a cross-modal distillation framework, where an LLM in one modality serves as the teacher and another as the student, with knowledge transfer in sound classes predicted as more challenging to the student by a heuristic model. Distillation in both directions, from Qwen2-VL to Qwen2-Audio and vice versa, leads to notable improvements, particularly in challenging classes. This work highlights the sensory gap in LLMs from a human-aligned perspective and proposes a principled approach to enhancing modality-specific perception in multimodal LLMs.

Linyang He, Qiaolin Wang, Xilin Jiang et al.

Transformer-based speech language models (SLMs) have significantly improved neural speech recognition and understanding. While existing research has examined how well SLMs encode shallow acoustic and phonetic features, the extent to which SLMs encode nuanced syntactic and conceptual features remains unclear. By drawing parallels with linguistic competence assessments for large language models, this study is the first to systematically evaluate the presence of contextual syntactic and semantic features across SLMs for self-supervised learning (S3M), automatic speech recognition (ASR), speech compression (codec), and as the encoder for auditory large language models (AudioLLMs). Through minimal pair designs and diagnostic feature analysis across 71 tasks spanning diverse linguistic levels, our layer-wise and time-resolved analysis uncovers that 1) all speech encode grammatical features more robustly than conceptual ones.

Xilin Jiang, Qiaolin Wang, Junkai Wu et al.

Internet audio-visual clips convey meaning through time-varying sound and motion, which extend beyond what text alone can represent. To examine whether AI models can understand such signals in human cultural contexts, we introduce AVMeme Exam, a human-curated benchmark of over one thousand iconic Internet sounds and videos spanning speech, songs, music, and sound effects. Each meme is paired with a unique Q&A assessing levels of understanding from surface content to context and emotion to usage and world knowledge, along with metadata such as original year, transcript, summary, and sensitivity. We systematically evaluate state-of-the-art multimodal large language models (MLLMs) alongside human participants using this benchmark. Our results reveal a consistent limitation: current models perform poorly on textless music and sound effects, and struggle to think in context and in culture compared to surface content. These findings highlight a key gap in human-aligned multimodal intelligence and call for models that can perceive contextually and culturally beyond the surface of what they hear and see. Project page: avmemeexam.github.io/public

Yizi Zhang, Linyang He, Chaofei Fan et al.

Speech brain-computer interfaces (BCIs) aim to restore communication for people with paralysis by translating neural activity into text. Most systems use cascaded frameworks that decode phonemes before assembling sentences with an n-gram language model (LM), preventing joint optimization of all stages simultaneously. Here, we introduce an end-to-end Brain-to-Text (BIT) framework that translates neural activity into coherent sentences using a single differentiable neural network. Central to our approach is a cross-task, cross-species pretrained neural encoder, whose representations transfer to both attempted and imagined speech. In a cascaded setting with an n-gram LM, the pretrained encoder establishes a new state-of-the-art (SOTA) on the Brain-to-Text '24 and '25 benchmarks. Integrated end-to-end with audio large language models (LLMs) and trained with contrastive learning for cross-modal alignment, BIT reduces the word error rate (WER) of the prior end-to-end method from 24.69% to 10.22%. Notably, we find that small-scale audio LLMs markedly improve end-to-end decoding. Beyond record-setting performance, BIT aligns attempted and imagined speech embeddings to enable cross-task generalization. Altogether, our approach advances the integration of large, diverse neural datasets, paving the way for an end-to-end decoding framework that supports seamless, differentiable optimization.

Linyang He, Tianjun Zhong, Richard Antonello et al.

Understanding how the human brain progresses from processing simple linguistic inputs to performing high-level reasoning is a fundamental challenge in neuroscience. While modern large language models (LLMs) are increasingly used to model neural responses to language, their internal representations are highly "entangled," mixing information about lexicon, syntax, meaning, and reasoning. This entanglement biases conventional brain encoding analyses toward linguistically shallow features (e.g., lexicon and syntax), making it difficult to isolate the neural substrates of cognitively deeper processes. Here, we introduce a residual disentanglement method that computationally isolates these components. By first probing an LM to identify feature-specific layers, our method iteratively regresses out lower-level representations to produce four nearly orthogonal embeddings for lexicon, syntax, meaning, and, critically, reasoning. We used these disentangled embeddings to model intracranial (ECoG) brain recordings from neurosurgical patients listening to natural speech. We show that: 1) This isolated reasoning embedding exhibits unique predictive power, accounting for variance in neural activity not explained by other linguistic features and even extending to the recruitment of visual regions beyond classical language areas. 2) The neural signature for reasoning is temporally distinct, peaking later (~350-400ms) than signals related to lexicon, syntax, and meaning, consistent with its position atop a processing hierarchy. 3) Standard, non-disentangled LLM embeddings can be misleading, as their predictive success is primarily attributable to linguistically shallow features, masking the more subtle contributions of deeper cognitive processing.

Qiaolin Wang, Xilin Jiang, Linyang He et al. · uw

While large audio-language models (LALMs) have demonstrated state-of-the-art audio understanding, their reasoning capability in complex soundscapes still falls behind large vision-language models (LVLMs). Compared to the visual domain, one bottleneck is the lack of large-scale chain-of-thought audio data to teach LALM stepwise reasoning. To circumvent this data and modality gap, we present SightSound-R1, a cross-modal distillation framework that transfers advanced reasoning from a stronger LVLM teacher to a weaker LALM student on the same audio-visual question answering (AVQA) dataset. SightSound-R1 consists of three core steps: (i) test-time scaling to generate audio-focused chains of thought (CoT) from an LVLM teacher, (ii) audio-grounded validation to filter hallucinations, and (iii) a distillation pipeline with supervised fine-tuning (SFT) followed by Group Relative Policy Optimization (GRPO) for the LALM student. Results show that SightSound-R1 improves LALM reasoning performance both in the in-domain AVQA test set as well as in unseen auditory scenes and questions, outperforming both pretrained and label-only distilled baselines. Thus, we conclude that vision reasoning can be effectively transferred to audio models and scaled with abundant audio-visual data.

Siavash Shams, Richard Antonello, Gavin Mischler et al.

Decoding continuous language from neural signals remains a significant challenge in the intersection of neuroscience and artificial intelligence. We introduce Neuro2Semantic, a novel framework that reconstructs the semantic content of perceived speech from intracranial EEG (iEEG) recordings. Our approach consists of two phases: first, an LSTM-based adapter aligns neural signals with pre-trained text embeddings; second, a corrector module generates continuous, natural text directly from these aligned embeddings. This flexible method overcomes the limitations of previous decoding approaches and enables unconstrained text generation. Neuro2Semantic achieves strong performance with as little as 30 minutes of neural data, outperforming a recent state-of-the-art method in low-data settings. These results highlight the potential for practical applications in brain-computer interfaces and neural decoding technologies.

Linyang He, Ercong Nie, Helmut Schmid et al.

This study investigates the linguistic understanding of Large Language Models (LLMs) regarding signifier (form) and signified (meaning) by distinguishing two LLM assessment paradigms: psycholinguistic and neurolinguistic. Traditional psycholinguistic evaluations often reflect statistical rules that may not accurately represent LLMs' true linguistic competence. We introduce a neurolinguistic approach, utilizing a novel method that combines minimal pair and diagnostic probing to analyze activation patterns across model layers. This method allows for a detailed examination of how LLMs represent form and meaning, and whether these representations are consistent across languages. We found: (1) Psycholinguistic and neurolinguistic methods reveal that language performance and competence are distinct; (2) Direct probability measurement may not accurately assess linguistic competence; (3) Instruction tuning won't change much competence but improve performance; (4) LLMs exhibit higher competence and performance in form compared to meaning. Additionally, we introduce new conceptual minimal pair datasets for Chinese (COMPS-ZH) and German (COMPS-DE), complementing existing English datasets.

Xilin Jiang, Yinghao Aaron Li, Nima Mesgarani

Lifelong audio feature extraction involves learning new sound classes incrementally, which is essential for adapting to new data distributions over time. However, optimizing the model only on new data can lead to catastrophic forgetting of previously learned tasks, which undermines the model's ability to perform well over the long term. This paper introduces a new approach to continual audio representation learning called DeCoR. Unlike other methods that store previous data, features, or models, DeCoR indirectly distills knowledge from an earlier model to the latest by predicting quantization indices from a delayed codebook. We demonstrate that DeCoR improves acoustic scene classification accuracy and integrates well with continual self-supervised representation learning. Our approach introduces minimal storage and computation overhead, making it a lightweight and efficient solution for continual learning.

Yinghao Aaron Li, Ali Zare, Nima Mesgarani

We present an unsupervised non-parallel many-to-many voice conversion (VC) method using a generative adversarial network (GAN) called StarGAN v2. Using a combination of adversarial source classifier loss and perceptual loss, our model significantly outperforms previous VC models. Although our model is trained only with 20 English speakers, it generalizes to a variety of voice conversion tasks, such as any-to-many, cross-lingual, and singing conversion. Using a style encoder, our framework can also convert plain reading speech into stylistic speech, such as emotional and falsetto speech. Subjective and objective evaluation experiments on a non-parallel many-to-many voice conversion task revealed that our model produces natural sounding voices, close to the sound quality of state-of-the-art text-to-speech (TTS) based voice conversion methods without the need for text labels. Moreover, our model is completely convolutional and with a faster-than-real-time vocoder such as Parallel WaveGAN can perform real-time voice conversion.

Cong Han, Yi Luo, Chenda Li et al.

Leveraging additional speaker information to facilitate speech separation has received increasing attention in recent years. Recent research includes extracting target speech by using the target speaker's voice snippet and jointly separating all participating speakers by using a pool of additional speaker signals, which is known as speech separation using speaker inventory (SSUSI). However, all these systems ideally assume that the pre-enrolled speaker signals are available and are only evaluated on simple data configurations. In realistic multi-talker conversations, the speech signal contains a large proportion of non-overlapped regions, where we can derive robust speaker embedding of individual talkers. In this work, we adopt the SSUSI model in long recordings and propose a self-informed, clustering-based inventory forming scheme for long recording, where the speaker inventory is fully built from the input signal without the need for external speaker signals. Experiment results on simulated noisy reverberant long recording datasets show that the proposed method can significantly improve the separation performance across various conditions.

Yi Luo, Cong Han, Nima Mesgarani

Despite the recent progress on neural network architectures for speech separation, the balance between the model size, model complexity and model performance is still an important and challenging problem for the deployment of such models to low-resource platforms. In this paper, we propose two simple modules, group communication and context codec, that can be easily applied to a wide range of architectures to jointly decrease the model size and complexity without sacrificing the performance. A group communication module splits a high-dimensional feature into groups of low-dimensional features and captures the inter-group dependency. A separation module with a significantly smaller model size can then be shared by all the groups. A context codec module, containing a context encoder and a context decoder, is designed as a learnable downsampling and upsampling module to decrease the length of a sequential feature processed by the separation module. The combination of the group communication and the context codec modules is referred to as the GC3 design. Experimental results show that applying GC3 on multiple network architectures for speech separation can achieve on-par or better performance with as small as 2.5% model size and 17.6% model complexity, respectively.

Yi Luo, Nima Mesgarani

Various neural network architectures have been proposed in recent years for the task of multi-channel speech separation. Among them, the filter-and-sum network (FaSNet) performs end-to-end time-domain filter-and-sum beamforming and has shown effective in both ad-hoc and fixed microphone array geometries. In this paper, we investigate multiple ways to improve the performance of FaSNet. From the problem formulation perspective, we change the explicit time-domain filter-and-sum operation which involves all the microphones into an implicit filter-and-sum operation in the latent space of only the reference microphone. The filter-and-sum operation is applied on a context around the frame to be separated. This allows the problem formulation to better match the objective of end-to-end separation. From the feature extraction perspective, we modify the calculation of sample-level normalized cross correlation (NCC) features into feature-level NCC (fNCC) features. This makes the model better matches the implicit filter-and-sum formulation. Experiment results on both ad-hoc and fixed microphone array geometries show that the proposed modification to the FaSNet, which we refer to as iFaSNet, is able to significantly outperform the benchmark FaSNet across all conditions with an on par model complexity.

Yi Luo, Zhuo Chen, Cong Han et al.

Modules in all existing speech separation networks can be categorized into single-input-multi-output (SIMO) modules and single-input-single-output (SISO) modules. SIMO modules generate more outputs than input, and SISO modules keep the numbers of input and output the same. While the majority of separation models only contain SIMO architectures, it has also been shown that certain two-stage separation systems integrated with a post-enhancement SISO module can improve the separation quality. Why performance improvements can be achieved by incorporating the SISO modules? Are SIMO modules always necessary? In this paper, we empirically examine those questions by designing models with varying configurations in the SIMO and SISO modules. We show that comparing with the standard SIMO-only design, a mixed SIMO-SISO design with a same model size is able to improve the separation performance especially under low-overlap conditions. We further validate the necessity of SIMO modules and show that SISO-only models are still able to perform separation without sacrificing the performance. The observations allow us to rethink the model design paradigm and present different views on how the separation is performed.

Yi Luo, Cong Han, Nima Mesgarani

Model size and complexity remain the biggest challenges in the deployment of speech enhancement and separation systems on low-resource devices such as earphones and hearing aids. Although methods such as compression, distillation and quantization can be applied to large models, they often come with a cost on the model performance. In this paper, we provide a simple model design paradigm that explicitly designs ultra-lightweight models without sacrificing the performance. Motivated by the sub-band frequency-LSTM (F-LSTM) architectures, we introduce the group communication (GroupComm), where a feature vector is split into smaller groups and a small processing block is used to perform inter-group communication. Unlike standard F-LSTM models where the sub-band outputs are concatenated, an ultra-small module is applied on all the groups in parallel, which allows a significant decrease on the model size. Experiment results show that comparing with a strong baseline model which is already lightweight, GroupComm can achieve on par performance with 35.6 times fewer parameters and 2.3 times fewer operations.

Yi Luo, Nima Mesgarani

Many recent source separation systems are designed to separate a fixed number of sources out of a mixture. In the cases where the source activation patterns are unknown, such systems have to either adjust the number of outputs or to identify invalid outputs from the valid ones. Iterative separation methods have gain much attention in the community as they can flexibly decide the number of outputs, however (1) they typically rely on long-term information to determine the stopping time for the iterations, which makes them hard to operate in a causal setting; (2) they lack a "fault tolerance" mechanism when the estimated number of sources is different from the actual number. In this paper, we propose a simple training method, the auxiliary autoencoding permutation invariant training (A2PIT), to alleviate the two issues. A2PIT assumes a fixed number of outputs and uses auxiliary autoencoding loss to force the invalid outputs to be the copies of the input mixture, and detects invalid outputs in a fully unsupervised way during inference phase. Experiment results show that A2PIT is able to improve the separation performance across various numbers of speakers and effectively detect the number of speakers in a mixture.

Cong Han, Yi Luo, Nima Mesgarani

Deep learning speech separation algorithms have achieved great success in improving the quality and intelligibility of separated speech from mixed audio. Most previous methods focused on generating a single-channel output for each of the target speakers, hence discarding the spatial cues needed for the localization of sound sources in space. However, preserving the spatial information is important in many applications that aim to accurately render the acoustic scene such as in hearing aids and augmented reality (AR). Here, we propose a speech separation algorithm that preserves the interaural cues of separated sound sources and can be implemented with low latency and high fidelity, therefore enabling a real-time modification of the acoustic scene. Based on the time-domain audio separation network (TasNet), a single-channel time-domain speech separation system that can be implemented in real-time, we propose a multi-input-multi-output (MIMO) end-to-end extension of TasNet that takes binaural mixed audio as input and simultaneously separates target speakers in both channels. Experimental results show that the proposed end-to-end MIMO system is able to significantly improve the separation performance and keep the perceived location of the modified sources intact in various acoustic scenes.

Yi Luo, Zhuo Chen, Nima Mesgarani et al.

An important problem in ad-hoc microphone speech separation is how to guarantee the robustness of a system with respect to the locations and numbers of microphones. The former requires the system to be invariant to different indexing of the microphones with the same locations, while the latter requires the system to be able to process inputs with varying dimensions. Conventional optimization-based beamforming techniques satisfy these requirements by definition, while for deep learning-based end-to-end systems those constraints are not fully addressed. In this paper, we propose transform-average-concatenate (TAC), a simple design paradigm for channel permutation and number invariant multi-channel speech separation. Based on the filter-and-sum network (FaSNet), a recently proposed end-to-end time-domain beamforming system, we show how TAC significantly improves the separation performance across various numbers of microphones in noisy reverberant separation tasks with ad-hoc arrays. Moreover, we show that TAC also significantly improves the separation performance with fixed geometry array configuration, further proving the effectiveness of the proposed paradigm in the general problem of multi-microphone speech separation.

Yi Luo, Enea Ceolini, Cong Han et al.

Beamforming has been extensively investigated for multi-channel audio processing tasks. Recently, learning-based beamforming methods, sometimes called \textit{neural beamformers}, have achieved significant improvements in both signal quality (e.g. signal-to-noise ratio (SNR)) and speech recognition (e.g. word error rate (WER)). Such systems are generally non-causal and require a large context for robust estimation of inter-channel features, which is impractical in applications requiring low-latency responses. In this paper, we propose filter-and-sum network (FaSNet), a time-domain, filter-based beamforming approach suitable for low-latency scenarios. FaSNet has a two-stage system design that first learns frame-level time-domain adaptive beamforming filters for a selected reference channel, and then calculate the filters for all remaining channels. The filtered outputs at all channels are summed to generate the final output. Experiments show that despite its small model size, FaSNet is able to outperform several traditional oracle beamformers with respect to scale-invariant signal-to-noise ratio (SI-SNR) in reverberant speech enhancement and separation tasks. Moreover, when trained with a frequency-domain objective function on the CHiME-3 dataset, FaSNet achieves 14.3\% relative word error rate reduction (RWERR) compared with the baseline model. These results show the efficacy of FaSNet particularly in reverberant and noisy signal conditions.

Yi Luo, Nima Mesgarani

Single-channel, speaker-independent speech separation methods have recently seen great progress. However, the accuracy, latency, and computational cost of such methods remain insufficient. The majority of the previous methods have formulated the separation problem through the time-frequency representation of the mixed signal, which has several drawbacks, including the decoupling of the phase and magnitude of the signal, the suboptimality of time-frequency representation for speech separation, and the long latency in calculating the spectrograms. To address these shortcomings, we propose a fully-convolutional time-domain audio separation network (Conv-TasNet), a deep learning framework for end-to-end time-domain speech separation. Conv-TasNet uses a linear encoder to generate a representation of the speech waveform optimized for separating individual speakers. Speaker separation is achieved by applying a set of weighting functions (masks) to the encoder output. The modified encoder representations are then inverted back to the waveforms using a linear decoder. The masks are found using a temporal convolutional network (TCN) consisting of stacked 1-D dilated convolutional blocks, which allows the network to model the long-term dependencies of the speech signal while maintaining a small model size. The proposed Conv-TasNet system significantly outperforms previous time-frequency masking methods in separating two- and three-speaker mixtures. Additionally, Conv-TasNet surpasses several ideal time-frequency magnitude masks in two-speaker speech separation as evaluated by both objective distortion measures and subjective quality assessment by human listeners. Finally, Conv-TasNet has a significantly smaller model size and a shorter minimum latency, making it a suitable solution for both offline and real-time speech separation applications.

Yi Luo, Nima Mesgarani

Robust speech processing in multi-talker environments requires effective speech separation. Recent deep learning systems have made significant progress toward solving this problem, yet it remains challenging particularly in real-time, short latency applications. Most methods attempt to construct a mask for each source in time-frequency representation of the mixture signal which is not necessarily an optimal representation for speech separation. In addition, time-frequency decomposition results in inherent problems such as phase/magnitude decoupling and long time window which is required to achieve sufficient frequency resolution. We propose Time-domain Audio Separation Network (TasNet) to overcome these limitations. We directly model the signal in the time-domain using an encoder-decoder framework and perform the source separation on nonnegative encoder outputs. This method removes the frequency decomposition step and reduces the separation problem to estimation of source masks on encoder outputs which is then synthesized by the decoder. Our system outperforms the current state-of-the-art causal and noncausal speech separation algorithms, reduces the computational cost of speech separation, and significantly reduces the minimum required latency of the output. This makes TasNet suitable for applications where low-power, real-time implementation is desirable such as in hearable and telecommunication devices.

Hassan Akbari, Himani Arora, Liangliang Cao et al.

In this study, we propose a deep neural network for reconstructing intelligible speech from silent lip movement videos. We use auditory spectrogram as spectral representation of speech and its corresponding sound generation method resulting in a more natural sounding reconstructed speech. Our proposed network consists of an autoencoder to extract bottleneck features from the auditory spectrogram which is then used as target to our main lip reading network comprising of CNN, LSTM and fully connected layers. Our experiments show that the autoencoder is able to reconstruct the original auditory spectrogram with a 98% correlation and also improves the quality of reconstructed speech from the main lip reading network. Our model, trained jointly on different speakers is able to extract individual speaker characteristics and gives promising results of reconstructing intelligible speech with superior word recognition accuracy.

Yi Luo, Zhuo Chen, Nima Mesgarani

Despite the recent success of deep learning for many speech processing tasks, single-microphone, speaker-independent speech separation remains challenging for two main reasons. The first reason is the arbitrary order of the target and masker speakers in the mixture permutation problem, and the second is the unknown number of speakers in the mixture output dimension problem. We propose a novel deep learning framework for speech separation that addresses both of these issues. We use a neural network to project the time-frequency representation of the mixture signal into a high-dimensional embedding space. A reference point attractor is created in the embedding space to represent each speaker which is defined as the centroid of the speaker in the embedding space. The time-frequency embeddings of each speaker are then forced to cluster around the corresponding attractor point which is used to determine the time-frequency assignment of the speaker. We propose three methods for finding the attractors for each source in the embedding space and compare their advantages and limitations. The objective function for the network is standard signal reconstruction error which enables end-to-end operation during both training and test phases. We evaluated our system using the Wall Street Journal dataset WSJ0 on two and three speaker mixtures and report comparable or better performance than other state-of-the-art deep learning methods for speech separation.

Zhuo Chen, Yi Luo, Nima Mesgarani

Despite the overwhelming success of deep learning in various speech processing tasks, the problem of separating simultaneous speakers in a mixture remains challenging. Two major difficulties in such systems are the arbitrary source permutation and unknown number of sources in the mixture. We propose a novel deep learning framework for single channel speech separation by creating attractor points in high dimensional embedding space of the acoustic signals which pull together the time-frequency bins corresponding to each source. Attractor points in this study are created by finding the centroids of the sources in the embedding space, which are subsequently used to determine the similarity of each bin in the mixture to each source. The network is then trained to minimize the reconstruction error of each source by optimizing the embeddings. The proposed model is different from prior works in that it implements an end-to-end training, and it does not depend on the number of sources in the mixture. Two strategies are explored in the test time, K-means and fixed attractor points, where the latter requires no post-processing and can be implemented in real-time. We evaluated our system on Wall Street Journal dataset and show 5.49\% improvement over the previous state-of-the-art methods.

Yi Luo, Zhuo Chen, John R. Hershey et al.

Deep clustering is the first method to handle general audio separation scenarios with multiple sources of the same type and an arbitrary number of sources, performing impressively in speaker-independent speech separation tasks. However, little is known about its effectiveness in other challenging situations such as music source separation. Contrary to conventional networks that directly estimate the source signals, deep clustering generates an embedding for each time-frequency bin, and separates sources by clustering the bins in the embedding space. We show that deep clustering outperforms conventional networks on a singing voice separation task, in both matched and mismatched conditions, even though conventional networks have the advantage of end-to-end training for best signal approximation, presumably because its more flexible objective engenders better regularization. Since the strengths of deep clustering and conventional network architectures appear complementary, we explore combining them in a single hybrid network trained via an approach akin to multi-task learning. Remarkably, the combination significantly outperforms either of its components.