arXiv Science⌕ Search

arXiv subjects

Hung-yi Lee

Publications and source records attributed to Hung-yi Lee.

At least 37 records · Page 2Linked to original sources

Causal Tracing of Audio-Text Fusion in Large Audio Language Models

Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains unclear. We adapt causal tracing to investigate the internal information flow of LALMs during audio comprehension. By conducting layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral, we evaluate the causal effects of individual hidden states. Layer-wise analysis identifies different fusion strategies, from progressive integration in DeSTA to abrupt late-stage fusion in Qwen. Token-wise analysis shows that the final sequence token acts as an informational bottleneck where the network decisively retrieves relevant information from the audio. We also observe an attention-like query mechanism at intermediate token positions that triggers the model to pull task-relevant audio context. These findings provide a clear characterization of when and where multi-modal integration occurs within LALMs.

cs.SD↗

Listen, Critique, and Refine: RL-Based Self-Refinement for Instruction-Following Speech Synthesis

Large Audio Language Models (LALMs) can follow diverse instructions to synthesize speech in specified styles. However, complex instructions that require simultaneous control over pitch dynamics, speaking rate, and emotional tone often exceed what a single-pass generation can faithfully realize. While recent reasoning models have shown that intermediate "thinking" tokens improve output quality, this paradigm has been confined to the text modality. In this work, we extend reasoning to the audio token space by training a LALM with reinforcement learning to reason over its own speech output. The model first generates a draft speech as a form of audio-token reasoning, critiques its own generation by reflecting on the acoustic realization in text, and then produces a refined version conditioned on both the first-pass speech and the critique, all within a single model. After RL training, the refined two-hop outputs achieve a relative improvement of 7.15\% on the InstructTTSEval benchmark, demonstrating the model's reflective ability.

eess.AS↗

TurnBench: A Multi-Domain Benchmark for Turn-Taking Dynamics in Spoken Dialogue

Speakers in natural conversation take turns speaking and listening, deciding in real time when to take, hold, or yield the floor. However, turn-taking evaluation remains limited due to the lack of a consistent, linguistically grounded evaluation protocol and hand-annotated data covering diverse conversation types. To address this, we present TurnBench, a multi-domain benchmark that pairs a 30-hour, hand-labeled corpus of dyadic human conversation with a standardized evaluation protocol for end-of-turn and interruption detection. We set conversation type as a controllable experimental variable, covering six distinct interaction styles, and triple-annotate each conversation. Benchmarking 14 heterogeneous turn-taking systems, we find end-of-turn recall stable across types, while interruption false positives are strongly type-dependent and concentrated in backchannel-dense interaction styles. Although in smooth floor transfers human listeners begin speaking a median 151 ms before the current turn ends, no current system performs equivalently without incurring excessive false positives. We release our corpus, a 104-hour training set, and a public leaderboard with an interactive dataset viewer at https://turnbench.sesame.com.

eess.AS↗

Correlation-Guided Encoder Selection for Multi-Encoder Large Audio-Language Models

Multi-encoder fusion extends Large Audio-Language Models (LALMs) beyond speech-centric recognition, but selecting encoders via intuition or exhaustive search often introduces redundant representations and inflates an already constrained compute budget. We propose CUES (Correlation-gUided Encoder Selection), a lightweight heuristic that estimates complementarity through task- and category-level Pearson correlations between encoders' performance profiles, scoring a candidate set from single-encoder evaluations alone--without fusion training during selection. Evaluated on the XARES-LLM benchmark with a frozen SmolLM2-135M backbone (LoRA-adapted) via five-fold cross-validation, CUES consistently identifies the same configuration per track from held-out development splits alone, without using test data for selection. For the broad Track~A suite, CUES selects a cross-family trio (Whisper-medium, mHuBERT-147, and Dasheng-base), achieving a 4.3% relative gain over Whisper-medium (0.771 vs. 0.739). For Track~B text generation, it re-anchors on a focused, speech-only pair (mHuBERT-147 and WavLM-base-plus) and actively abstains from adding a divergent encoder, outperforming mHuBERT-147 by 6.3% (0.589 vs. 0.554). Rather than a failure to scale, this divergence is consistent with a diversity--interference trade-off that CUES navigates per track from correlation signals alone: across the evaluated pool, added cross-family diversity tends toward an inverted-U on broad audio tasks but toward steady degradation on text generation, which favors a focused, speech-anchored set.

eess.AS↗

REDDIT: Forgetting-Resistant Correction of Timestamp Drift in ASR via Replay-Based Distribution Editing

Modern autoregressive ASR systems can emit timestamps as decoded tokens, enabling timestamped transcription without frame-level aligners or inference-time post-processing. We show that these generated timestamps can drift across long non-speech spans: the transcript may remain plausible, but the decoded time axis drifts away from the audio. We study this non-speech-induced timestamp drift with self-built gap and long-gap benchmarks across 15 evaluated timestamp-producing ASR and audio-language systems. Naive timestamp-corrected fine-tuning improves alignment but can severely degrade non-target ASR behavior, exposing a forgetting problem. We propose REDDIT(REplay-based Distribution eDITing), a lightweight two-stage post-training framework that corrects timestamps while avoiding this catastrophic forgetting: it first edits timestamp targets under the model's own replayed decoder context while matching the frozen base distribution on non-timestamp tokens, then applies a short edited-prefix refinement stage. In this framework, we construct correction supervision without human transcripts or human timestamp annotations by combining VAD-trimmed speech spans with inserted non-speech gaps and known concatenation offsets. On Whisper-tiny, 34.9 hours of targeted correction audio used and only 1.6% of model parameters updated, raising long-gap mIoU from 38.7% to 95.0% and reducing mixed-gap out-of-domain AAS from 2752 ms to 223 ms while preserving CV-en MER at 41.3% (versus 524.2% for ordinary SFT decoder tuning).

cs.CL↗

How Contrastive Decoding Enhances Large Audio Language Models

While Contrastive Decoding (CD) has been proposed to enhance Large Audio Language Models (LALMs), it has not been evaluated at scale, and the underlying mechanisms driving its success remain unclear. This study systematically evaluates four distinct CD strategies across diverse LALM architectures. We identify Audio-Aware Decoding and Audio Contrastive Decoding as the most effective methods. However, their impact varies significantly across models. To explain this variability, we profile the baseline error composition of each model and measure how readily contrastive decoding corrects each error type. Our analysis demonstrates that CD reliably rectifies errors in which models falsely claim an absence of audio or resort to uncertainty-driven guessing, but is relatively poor at correcting flawed reasoning or confident misassertions. Crucially, CD's benefit closely tracks the composition of a model's baseline error profile: when errors caused by audio ignorance or uncertainty-driven guessing constitute only a small fraction of a model's errors, gains are marginal or even negative. A token-level analysis reveals the underlying mechanism: when the amateur's output is dominated by hesitation markers, CD's suppression naturally targets uncertainty-driven errors while having limited effect on confident misassertions.

cs.SD↗

MP-Bench: Evaluating Voice Agents as a Multiparty Conversation Participant

Conversational voice agents have advanced significantly, offering increasingly natural human-machine interactions through both cascaded and end-to-end architectures. However, while recent benchmarks extensively evaluate dyadic interactions and passive audio comprehension, they largely overlook a prevalent real-world scenario: multi-party conversations. Evaluating agents in these settings is fundamentally more challenging than in dyadic interactions due to the exponentially greater conversational complexity. For voice agents to integrate seamlessly into human group dynamics, they must not only generate contextually appropriate responses but also demonstrate a nuanced understanding of open turn-taking. To address this gap, we introduce Multiparty Bench (MP-Bench), the first benchmark specifically designed to objectively evaluate conversational speech systems as active participants within multi-party contexts. MP-Bench assesses agent behavior along two primary dimensions: turn-taking awareness and response appropriateness. Additionally, we incorporate comprehension-based question-answering tasks as a complementary evaluation. By benchmarking 12 voice agents, we find that real-time voice agents stay at or below 22% on multiparty comprehension and remain near chance on multiparty turn-taking, exposing an open challenge for real-time voice agents under multiparty scenario.

eess.AS↗

TASTE2: Text-Aligned Speech Modeling and Deployment toward Full-Duplex Voice Interaction

Full-duplex voice interaction requires more than utterance-level conversion. It must process streaming speech, manage turn-taking and interruptions, while preserving pretrained linguistic competence and acoustic paralinguistic cues. We ask whether TASTE (Text-Aligned Speech Tokenization and Embedding) provides a viable path toward this goal. We present TASTE2, which transforms utterance-level TASTE into an incremental dialogue stack. A shared text-token vocabulary removes word-level averaging, while modality-aligned dialogue training predicts one continuous audio latent per text token without interleaving heterogeneous token streams. An incremental Speech Detokenizer enables streaming synthesis through CosyVoice2. After speech and dialogue training, TASTE2 (Merge) reaches 56.3% on LLaMA-Questions against a 57.3% Qwen2.5-7B Instruct text-only reference (98.2% accuracy retention), and TASTE2 (Direct) reaches 53.0% (92.4% retention). We build TASTE2 VoiceBot, which processes user speech incrementally, streams synthesized audio, and stops generation on barge-in. On Full-Duplex-Bench v1.0, TASTE2 and TASTE2 VoiceBot handle interruptions well while maintaining high conversational coherence. Natural conversation remains challenging, and deployed mean time to first audio is 2.701 s on two NVIDIA RTX A6000 after TensorRT acceleration. Finally, to our knowledge, we provide the first systematic characterization of explicit paralinguistic control in a TASTE based model. Fast speaking rate serves as a cross-strategy proof of concept after dialogue SFT, while emotion control is strategy dependent and the remaining attributes stay weak. Together, these results establish TASTE based modeling as a practical route toward full-duplex systems while identifying natural conversation robustness, speech generation latency, and feature general paralinguistic control as open challenges. Explore TASTE2 online.

cs.SD↗

Causal Localization of the Refusal Direction in Audio Language Models

A large audio language model (LALM) attaches a speech front end to a text language model (LM) that is already safety-aligned. When such a model refuses a harmful spoken request, is the refusal carried by the front end, or inherited from the text LM? We test this with causal interventions. At each model's audio-to-LM interface and at tested LM residual layers, we fit a direction separating harmful from benign prompts, ablate its component, and measure the resulting change in the model's first-token refusal margin. Four of the five models are evaluated under held-out category shift. Across five LALMs spanning three backbone families, with three models passing a baseline safety gate, the largest tested effects occur in a mid-to-late LM band, while ablations of the tested interface directions have little effect. On Qwen2.5-Omni, ablating the L16 direction changes the margin by -7.10, versus -0.013 at the projector. The audio pathway is still in use: zeroing the encoder output changes the margin by -4.7. In the same model, the contrast is linearly decodable at an early layer where single-layer ablation has little effect, and a direction fitted on the LM backbone alone transfers to the full audio model. Because harmful and benign prompts also differ in form, we interpret the direction as refusal-linked rather than harmfulness-specific. These interventions localize dependence of the refusal margin, not where refusal is computed. Moving the margin also does not always change what the model writes. Safety audits of these models should use interventions rather than rely on probes alone, and should examine the inherited text LM alongside the audio interface.

cs.MM↗

Tri-PvP: Exposing Modality Bias in Omni-Modal Large Language Models through Perceptual-Propositional Evidence Conflicts

Omni-modal large language models (OLLMs) jointly process vision, audio, and text, yet their modality bias under cross-modal conflict remains underexplored. Existing benchmarks conflate two distinct forms of evidence within a single modality: perceptual signals (e.g., a photograph or recording of a dog) and propositional signals (e.g., the declarative claim "this is a dog"), such that any measured modality bias is inherently confounded with evidence-form bias, precluding clean attribution to either source. To address this, we introduce Tri-PvP, an 8,000-sample tri-modal conflict benchmark crossing vision, audio, and text, where vision and audio each take perceptual or propositional form. Evaluating five OLLMs, we find robust visual bias across most models and evidence-type conditions. Crucially, we reveal a systematic asymmetry in evidence-form bias: models exhibit a stronger bias toward perceptual signal in vision but propositional in audio. Further analyses via layer-wise linear probing and contrastive decoding reveal that modality bias is already linearly decodable from early representation layers and can only be partially mitigated, calling for mitigation strategies beyond surface-level interventions.

cs.CL↗

AdaSearch: Balancing Parametric Knowledge and Search in Large Language Models via Reinforcement Learning

Equipping large language models (LLMs) with search engines via reinforcement learning (RL) promises effective search agents. However, adaptively balancing internal parametric knowledge with external search remains a challenge, as overreliance on search introduces unnecessary cost and risks exposure to noisy or malicious content, while relying solely on parametric knowledge risks hallucination. Prior efforts mitigate search overuse through tool-call reward shaping, which requires heavy reward engineering and conflates necessary and unnecessary search. To address these limitations, we revisit the evaluation of search agents through an F1-based decision metric, revealing that prior methods often overlook readily available parametric knowledge. Motivated by this, we propose AdaSearch, a simple two-stage, outcome-driven RL framework that disentangles problem-solving from the decision to search, making the decision process explicit and interpretable. Extensive experiments demonstrate that AdaSearch significantly improves search-decision quality and reduces unnecessary search calls, with only a small trade-off in QA accuracy relative to always-search.

cs.CL↗

Hearing the Order: Investigating Position Bias in Large Audio-Language Models

Large audio-language models (LALMs) are often used in tasks that involve reasoning over ordered options. An open question is whether their predictions are influenced by the order of answer choices, which would indicate a form of position bias and undermine their reliability. In this paper, we identify and analyze this problem in LALMs. We demonstrate that no model is immune to this bias through extensive experiments on six LALMs across three widely used benchmarks and their spoken counterparts. Shuffling the order of answer options can cause performance fluctuations of up to 24% and even change model rankings, raising concerns about the reliability of current evaluation practices. We also study permutation-based strategies and show that they can mitigate bias in most cases. Our work represents the first systematic investigation of this issue in LALMs, and we hope it raises awareness and motivates further research in this direction.

cs.SD↗

Listen to the Latents: Self-Correcting Speech Recognition in Large Audio Language Models Through Hidden-State Interactions

Recent automatic speech recognition (ASR) systems increasingly integrate large language models (LLMs) to leverage their semantic knowledge, either externally through logit fusion or internally through warm initialization. However, how to effectively combine these two strategies remains underexplored. In this work, we refine warm-initialized LLM-based ASR models by leveraging their own pre-adaptation base LLMs, focusing on LoRA-adapted settings where the base LLM is preserved. To achieve this, we propose Hybrid Search, a targeted correction strategy motivated by two observations. First, interaction features that characterize the relationship between LLM-based ASR hidden states and base-LLM hidden states provide informative signals about a token's degree of semantic dependence. Second, selectively refining targeted tokens with high semantic dependence improves ASR performance far beyond naive global LLM-correction methods including rescoring and late fusion. Our analysis suggests that, even after semantic knowledge transfer through warm initialization, LLM-based ASR models can still leverage their base LLM to further improve inference-time performance.

cs.CL↗

SPAR-K: Scheduled Periodic Alternating Early Exit for Spoken Language Models

Interleaved spoken language models (SLMs) alternately generate text and speech tokens, but decoding at full transformer depth for every step becomes costly, especially due to long speech sequences. We propose SPAR-K, a modality-aware early exit framework designed to accelerate interleaved SLM inference while preserving perceptual quality. SPAR-K introduces a speech alternating-depth schedule: most speech positions exit at a fixed intermediate layer, while periodic full-depth "refresh" steps mitigate distribution shift due to early exit. We evaluate our framework using Step-Audio-2-mini and GLM-4-Voice across four datasets spanning reasoning, factual QA, and dialogue tasks, measuring performance in terms of ASR transcription accuracy and perceptual quality. Experimental results demonstrate that SPAR-K largely preserves question-answering accuracy with a maximum accuracy drop of 0.82\% while reducing average speech decoding depth by up to 11\% on Step-Audio-2-mini and 5\% on GLM-4-Voice, both with negligible changes in MOS and WER and no auxiliary computation overhead. We further demonstrate that confidence-based early exit strategies, widely used in text LLMs, are suboptimal for SLMs, highlighting that the unique statistical nature of speech tokens necessitates a specialized early exit design.

cs.CL↗

Joint Optimization of Tool Creation and Use for Large Language Model Agents

Tool-augmented language models are bounded by the APIs humans bothered to write; existing tool-creation systems patch this by prompting a frozen LLM at inference time, leaving the model that writes a tool decoupled from the one that uses it, with no signal that the schemas it produces are schemas it can invoke. We propose SMITH (Schema-grounded Multi-task Iterative Tool Honing), a reinforcement learning framework that jointly trains tool creation and tool use inside a single policy. Each rollout is either a build task (write a tool from a few examples) or a use task (invoke a pooled tool on a held-out question). Three separate reward axes catch schema, code, and outcome failures independently, so each failure mode contributes its own gradient. A 4B Qwen3 trained with SMITH on 13 procedural reasoning tasks with exact verifiers reaches 79.8 macro-average accuracy on held-out tasks, the best across all evaluated methods and ahead of an untrained 30B-A3B tool-writer. It also reaches 40.4 on TabMWP-Hard and 42.6 on out-of-domain GQA (+7.6 over the best same-backbone inference-time baseline), without any visual or tabular training data. Tools written by our 4B models also lifted the performance of LFM-2.5-350M and Qwen3-30B-A3B under same reasoning tasks.

cs.AI↗

INSPIRE: A Benchmark for Instruction-Aware Speech Retrieval

Existing speech retrieval systems rely on fixed similarity matching and cannot adapt to diverse user intents. We introduce INSPIRE, the first benchmark for instruction-aware speech retrieval, in which natural-language instructions dynamically specify relevance criteria, including semantic content, speaker identity, speaking style, environmental sounds, and their combinations. We evaluate four retrieval paradigms: large audio-language models, cascaded pipelines, self-supervised speech models, and contrastive audio-language models. Our results reveal that no current method robustly handles all retrieval intents. Text-based approaches perform relatively better at semantic retrieval but struggle with paralinguistic attributes, while speech-based models are moderately better at capturing acoustic properties but falter at following instructions. These findings highlight the need for unified architectures capable of instruction-aware speech retrieval.

cs.SD↗

Findings of the First Teaching Monster Challenge: A Benchmark of Pedagogical Content Knowledge in AI Agents

AI agents can now solve problems, answer like subject experts, and generate long-form multimodal content. However, whether they can adapt a lesson to fit a specified learner, which education calls Pedagogical Content Knowledge (PCK), has not been benchmarked. To measure it, we introduce the Teaching Monster Challenge, the first instructional video generation benchmark to treat the learner persona as an explicit evaluation criterion. Each system is given a topic and a learner persona and must generate a complete instructional video. Every video is screened by an LLM-judge, ranked by crowd pairwise voting, and finalized by an expert panel. The first edition shows that today's systems handle the content well but are far weaker at presenting it and adapting it to the learner. The same process exposes a limit of automatic judging. The LLM-judge separates a clear low-performing tail but ranks the strongest systems poorly. The strongest systems receive nearly identical scores from the judge, so its ranking of them does not match human preference. Progress therefore requires not only better teaching systems but also better automatic judges, and we release the benchmark, rubric, and human judgments as a testbed for both.

cs.AI↗

EduPanel: A Three-Agent LLM Judge for Teaching Videos -- Reliability, Complementarity, and Human Trust Calibration

Teaching videos are becoming a major medium for education, creating a growing need for scalable evaluation of their pedagogical quality. Existing automatic judges do not fully address this setting because teaching quality depends on multimodal evidence and should be evaluated with respect to the intended learner rather than as a universal property. We present EduPanel, a rubric-grounded, learner-conditioned LLM judge that decomposes evaluation across specialized agents to produce interpretable assessments for different aspects of teaching quality. Across expert studies, architecture ablations, and learner-persona analyses, EduPanel achieves reliability comparable to a median human expert. In expert evaluation, its feedback improves scoring accuracy (MAE 0.87 to 0.73), while experts remain able to detect unreliable outputs (AUC = 0.77) instead of accepting them blindly. These results suggest that EduPanel can serve as effective assistants for educational evaluation rather than replacements for human experts.

cs.HC↗