arXiv subjects

Zhizheng Wu

Publications and source records attributed to Zhizheng Wu.

At least 19 recordsLinked to original sources

arXiv2026

E3-System: Expressive EmoData, Explicitly Controllable, Easy-to-Use TTS

Recent zero-shot Text-to-Speech (TTS) systems can clone unseen speakers, yet controllable expressiveness remains limited by a gap between semantic intent and acoustic realization. Text descriptions offer a natural control interface but provide weak acoustic grounding, whereas speech style prompts provide concrete prosody but entangle style with the prompt speaker's timbre. To bridge this gap, we propose E3-System, a unified intent-to-speech framework for expressive zero-shot TTS that couples three requirements: Expressive style modeling, Explicit timbre-style control, and Easy-to-use natural-language interaction. Given a target speaker and either a natural-language style description or a speech style prompt, E3-System generates expressive speech while preserving the target timbre. The key idea is to close the full control loop: building a large expressive style space, grounding user intent into usable acoustic style references, and synthesizing speech with disentangled timbre and style. To realize this loop, EmoData provides a thousand-hour multilingual emotional speech dataset, Style-RAG grounds text descriptions into high-quality acoustic style prompts from these data, and Dual-TTS transfers the grounded style to arbitrary target speakers through VC-augmented timbre-style disentanglement. Experiments show that E3-System achieves strong speech quality, speaker preservation, and explicit expressive controllability.

eess.AS↗

arXiv2026

DuplexAgent-RSI: Recursive Harness Improvement for Full-Duplex Voice Agent Collaboration

Voice agents are converging on a collaboration pattern: a full-duplex interaction model stays on the live channel as the entry to the conversation, while search, reasoning, and coding are handled through asynchronous delegation. A duplex model supports continuous listening and speaking, but complex reasoning and tool use may exceed its capabilities. A coding agent can plan and execute extended tasks, but its sequential interface is a poor fit for live conversation. Combining them requires a harness that coordinates task acceptance, progress, cancellation, replacement, and result delivery while keeping the conversation responsive. Existing harnesses often rely on coupled heuristics, making them difficult to improve systematically from evidence. We present DuplexAgent, a full-duplex collaboration system whose harness expresses this workflow as six editable modules, and Duplex-Harness-RSI, a closed loop that revises them from interaction traces. A simulator automatically generates timed test conversations, runs the system, and produces failure traces that identify the collaboration modules requiring repair. Reasoning LLMs and coding agents in the delegation pool also serve the improvement loop: the Exam Planner selects the next tests from observed weaknesses and the repair archive, and the Harness Editor proposes targeted module changes. The capabilities that serve the user thus also improve the system's coordination. Experiments on intelligence, agentic, and duplex benchmarks show that DuplexAgent combines continuous interaction with difficult reasoning and complex task execution, achieving stronger spoken-knowledge and executable-tool scores than the compared delegated systems while maintaining strong interruption response. A harness ablation further shows that this modular, verifiable loop outperforms the initial harness and repeated editing that lacks its diagnosis and repair archive.

cs.AI↗

arXiv2026

AuraSE: Low-Hallucination Generative Speech Enhancement via Multimodal Flow Matching and Inference Policy Optimization

Generative speech enhancement models can produce cleaner and more natural-sounding speech than conventional discriminative approaches, but may hallucinate by changing speech content or speaker identity, even with transcript conditioning. We present AuraSE, a flow-matching framework that addresses hallucination through complementary modality and inference designs. First, a double-stream-to-single-stream multimodal Diffusion Transformer (MMDiT) allows transcript and acoustic representations to interact while preserving a dedicated pathway for the degraded input. Second, we find that the best decoder configuration, governed by guidance scale, sampling temperature, and step count, varies substantially across utterances. This observation motivates Inference Policy Optimization (IPO), an online, on-policy preference optimization method. IPO generates multiple candidates from the current model under different inference configurations, ranks them with a multi-objective reward, and learns from their relative preferences. AuraSE-IPO ranks first on 11 of 12 metrics across the synthetic test sets and obtains the highest DNSMOS and blind-listening scores among the evaluated systems on the real DNS blind test set. At deployment, it uses a fixed $10$-step ODE decoder without classifier-free guidance (CFG) or per-utterance configuration search.

cs.SD↗

arXiv2026

AURAL: Adaptive Latent Reasoning with Joint Chunk for Speech Language Models

Model intelligence and fast response jointly shape the quality of interaction with speech language models, yet remain difficult to achieve together. Explicit chain-of-thought (CoT) improves reasoning and audio understanding, but generating intermediate reasoning tokens delays responses. Describing fine-grained acoustic cues further lengthens CoT and increases latency. Latent reasoning can reduce this overhead, yet existing methods often trail CoT and remain limited by single-path supervision and reasoning budgets that do not adapt to problem difficulty. We introduce AURAL, which models a distribution over multiple plausible reasoning continuations in latent space and jointly predicts chunks of future states to reduce sequential forward passes and reasoning latency. To provide initial supervision for latent reasoning, we construct AuralReason-683K: 683K bilingual speech utterances (about 1,000 hours) with concise CoT for emotion recognition, empathetic dialogue, and general reasoning. AURAL-RL then explores beyond these traces, rewarding concise reasoning that yields high-quality answers and adapting reasoning effort to each problem. Across two backbones, AURAL-RL achieves performance comparable to CoT-RL, with larger gains over the respective supervised checkpoints on most metrics. Analysis further shows that harder questions elicit more latent reasoning steps. On Qwen2.5-Omni, it reduces time to the first answer token by 11.8x, from 1.22 to 0.10 s, versus 0.05 s for direct answering.

cs.CL↗

arXiv2026

Interpreting and Evaluating Dynamic-Rate Speech Codec Boundaries

Dynamic-frame-rate neural speech codecs replace a uniform frame grid with variable-duration tokens, making boundary placement part of the representation itself. Yet it is unclear what these boundaries encode and whether interpretable boundaries are also useful for neural speech reconstruction. This work combines boundary interpretation and controlled reconstruction analysis by comparing predicted boundaries with linguistic and acoustic references. We find that boundary meaning depends on the underlying speech representation. For the ASR-oriented SenseVoice and Whisper encoders, shallow layers emphasize phonetic, voicing, and acoustic transitions, whereas deeper layers shift toward syllable and subword structure. In a comparison of six dynamic-frame-rate algorithms and a uniform (fixed-frame-rate) baseline, higher-level linguistic alignment is associated with lower pooling distortion and better reconstruction from semantic tokens. Frame-rate-matched boundary tests make the distinction concrete: a syllable-derived partition improves over both Uniform and Similarity, while a phoneme-derived partition does not improve over Uniform. We infer that for semantically rich speech representations, useful codec boundaries are best understood as allocation decisions organized around the syllable scale.

cs.SD↗

arXiv2026

CHILDES-Aligned: A Curated Children's Speech Dataset via Multi-Model Timestamp Ensembling

CHILDES is a large-scale child speech corpus containing long-form recordings of naturalistic child-adult interactions, making it a valuable resource for studying child speech and language development. However, utterance-level timestamps provided in this corpus are often noisy, incomplete, or misaligned with the audio. As a result, utterances cannot always be reliably localized within long recordings, which limits the direct use of these data for training and evaluating speech models. In this work, we propose BEACON (Boundary Estimation via Alignment CONsensus), an ensemble timestamp-curation framework that refines utterance-level timestamps by aggregating knowledge from multiple off-the-shelf ASR models. Specifically, each model's word-level timestamp predictions are first aligned to provided human transcripts, and the final utterance time boundaries are determined by a consensus voting strategy. The framework is corpus-agnostic and applies to any long-form recording paired with a trusted transcript whose timestamps are unreliable or missing, offering a general recipe for timestamp curation. Leveraging this pipeline, we curate and release a 413-hour general-purpose child-speech dataset with corrected utterance-level timestamps, together with a 283-hour quality-controlled subset for ASR training. Fine-tuning on this subset yields up to an average 19.5% relative WER reduction on four out-of-domain child-speech benchmarks.

eess.AS↗

arXiv2026

EvoAudio: Recursive Self-Improvement for Audio Understanding

Audio language models understand what is said far better than how it sounds. Closing this gap takes more than data. Detailed acoustic annotation is costly, labels from stronger models inherit their errors and limits, and fixed data cannot adapt as the learner improves. We therefore propose EvoAudio, a recursive self-improvement system for audio understanding. To our knowledge, it is the first to evolve the model, waveforms, questions, and difficulty in one closed loop. EvoAudio uses the current model's performance to set the focus and difficulty of the next training data. A library of audio tools then constructs questions whose answers follow from how the audio was made, providing verifiable supervision without new human annotation. Reinforcement learning updates the model, and validation decides whether it enters the next evolution round. Across 13 rounds, EvoAudio improves five models with different audio encoders and language backbones on MMSU, MMAU-Pro, and MMAR. It achieves the highest average for every backbone, raising overall performance by up to 6.3 points. The improvement unfolds over successive rounds, with each stronger model starting the next round.

cs.SD↗

arXiv2026

Spoken Language Models that Think Aloud

While Chain-of-Thought (CoT) reasoning has improved the capability of language models, directly applying it to Spoken Language Models (SLMs) may introduce long silent intervals under the serial "think-then-speak" paradigm, disrupting real-time spoken interaction. To address this issue, we propose an asynchronous think-aloud framework for reasoning-based SLMs within the Thinker-Talker architecture. The framework maintains a primary reasoning stream for logical deduction and a lightweight think-aloud stream that generates short, task-grounded progress utterances conditioned on the user input and the evolving reasoning state. A dynamic balance strategy coordinates the two streams at runtime, triggering additional think-aloud speech to avoid silent gaps and canceling pending utterances when the final response becomes ready. Experiments on spoken reasoning and question-answering benchmarks show that our approach substantially reduces user-audible silence during reasoning while maintaining answer accuracy comparable to that of a serial "think-then-speak" baseline, demonstrating the potential of asynchronous think-aloud for responsive interaction in SLMs.

cs.CL↗

arXiv2026

CycleSpeech: Reciprocal Alignment for Instruction-Controlled Speech Synthesis and Paralinguistic Understanding

Instruction-controlled speech synthesis and paralinguistic understanding are often trained independently, leaving reciprocal feedback between the two tasks underexplored. We introduce CycleSpeech, a framework that connects generation and understanding through a shared, structured voice profile that serves as a common target for supervision and reciprocal feedback. The forward cycle assesses whether synthesized speech expresses the intended attributes by comparing recovered and target profiles. The backward cycle evaluates whether profiles inferred from real speech can guide reconstruction of the source speaking style. To support both directions, we construct a bilingual dataset of 20,046 examples pairing instructions, target speech, speaker references, and structured profiles. Building on joint supervised fine-tuning, CycleGRPO alternates policy updates using reciprocal rewards grounded in profile consistency and speaking-style reconstruction. Fixed target profiles anchor feedback from the evolving counterpart. This procedure requires neither human preference annotations nor an additional preference-trained reward model. Evaluations on Chinese and English benchmarks show improved instruction adherence and profile recovery while maintaining competitive synthesis quality. Compared with Step-Audio-2-mini, CycleSpeech improves instruction-match accuracy by 4.50 and 10.06 percentage points in Chinese and English, respectively. Controlled ablations further support the contribution of cycle feedback to generation control. These results support structured voice profiles as an interface for reciprocal training between speech generation and paralinguistic understanding. An online demo is available at https://cyclespeech.github.io.

cs.SD↗

arXiv2026

CPR: Combining global composing, local performing and full-sequence refining in piano rendering with continuous autoregressive modelling

Prompt-conditioned piano MIDI-to-Music rendering aims to faithfully render target notes while reproducing the timbre of a reference recording. Existing approaches primarily follow two paradigms: autoregressive (AR) modeling and flow matching (or diffusion). Discrete-codec AR models provide causal temporal modeling, but quantization can discard acoustic detail. Flow matching better preserves acoustic structure in the cost of full-sequence attention costs and worse semantic structure. Continuous autoregressive models operate directly on continuous representations. It not only combines the condition-following ability of AR models and distribution-modeling capacity of flow matching but also bypasses the quantization bottleneck with lower computational costs. Building on this principle, we present Composer--Performer--Refiner (CPR) framework. Composer autoregressively predicts continuous hidden states, Performer generates 24kHz acoustic latents through local flow matching and Refiner then upsamples the waveform to 48 kHz. We further introduce Bottlenecked Representation Alignment (BREPA) and Modality--Time RoPE (MT-RoPE) to strengthen musical semantic structure in Composer hidden states and temporal alignments across modalities. Codes are available at https://github.com/FEAfeatherTHER/CPR_official

cs.SD↗

arXiv2026

CoReLoop: Parameter-Efficient Controlled Recurrent Refinement for Audio Deepfake Detection

Generalizing to unseen attacks remains challenging for audio deepfake detectors, and collecting training data covering all potential attacks is impractical. We explore recurrent refinement in an already-trained SSL-based detector without additional data or changes to its original parameters. However, directly recycling encoder outputs as inputs degrades detection in our diagnostic. We propose CoReLoop, which makes this reuse effective by adapting recurrent inputs to the frozen encoder, controlling state updates, and aligning refined outputs with the frozen classifier. By training only lightweight refinement modules and loop-specific low-rank adapters on the original data, CoReLoop enables additional refinement while preserving the detector's original first-pass prediction. On 14 cross-domain test sets, the 24-layer model reduces pooled equal error rate (EER) from 4.85% to 3.74% with two passes, with approximately 10M trainable parameters out of 598M. To selectively apply this refinement, an optional halting head chooses the depth for each utterance, achieving 3.73% pooled EER with an average of 1.18 passes.

cs.SD↗

arXiv2026

GenTraceBench: A Benchmark for Tracing Audio Deepfakes Across Pre- and Post-training Stages

Modern text-to-speech (TTS) systems are rarely deployed as unchanged pre-trained models. They are often adapted through supervised fine-tuning (SFT) or preference optimization such as DPO and GRPO. This raises a practical question for audio deepfake forensics: do fingerprints learned from a foundation generator remain valid after adaptation? We present GenTraceBench, a controlled benchmark spanning five TTS architectures, 16 pre-/post-training variants, and 49,728 utterances generated with fixed texts and speaker prompts. Under a train-on-foundation, test-on-adapted protocol, we evaluate binary detection, closed-set attribution, and open-set verification. DPO and GRPO generally preserve fingerprints, whereas some SFT and pre-training-data changes cause substantial drift; effect sizes vary across three forensic backbones. Repeated training runs confirm the largest W2V-BERT attribution drop, while a data-mixture control with comparable speech quality shows that composition change need not cause drift. In W2V-BERT verification, multi-shot enrollment reduces EER for the SFT condition from 44.4% to 11.0%, whereas the SingNet-only condition remains at or above 45% EER.

cs.SD↗

arXiv2026

Multi-Dimensional Prosody Judgment For Live Streaming Speech Synthesis

Evaluating live streaming speech synthesis (TTS) requires assessing fine-grained, highly expressive prosody such as emotion, intonation, and energy which traditional MOS predictors fail to capture. While proprietary Large Language Models (LLMs) like Gemini can evaluate these aspects, they are too costly for massive inference and reinforcement learning feedback. To address this, we first introduce Live-ProsodyJudge (LPJ), a cost-effective pairwise evaluator distilled from Gemini into Qwen3-Omni. However, we identify a critical flaw in standard multi-dimensional evaluation: verdict coupling. The judge tends to lazily align all individual dimension scores with its overall preference, collapsing a rich multi-dimensional rubric into a single preference bit. To resolve this, we further propose Decoupled-Live-ProsodyJudge (D-LPJ). D-LPJ eliminates the overall verdict target to prevent blind following, masks uncertain pair-dimensions during Supervised Fine-Tuning(SFT), and introduces a novel span-local GRPO strategy that applies normalized advantages strictly to their corresponding rationale spans. Evaluated on highly curated human-annotated test sets, 10 sample balanced-order LPJ achieves higher point accuracy than a single Gemini call, while D-LPJ successfully produces independent,decoupled dimension judgments. Furthermore, in a Best-of-8 TTS candidate selection tournament, the LPJ-selected utterance falls within the human top-3 in 85.29% of high-confidence cases, demonstrating its efficacy for fine-grained TTS preference optimization.

cs.SD↗

arXiv2026

FlexiSLM: A Spoken Language Model with Dynamic and Controllable Frame Rates

Spoken language models (SLMs) extend LLMs to speech input and output, but existing systems use fixed frame rates (e.g., 25 or 12.5 Hz), overlooking speech's time-varying information density and limiting inference-time quality-speed tradeoffs. Recent dynamic-frame-rate audio tokenizers enable very low average frame rates and controllability, yet had not been applied to SLMs. We introduce FlexiSLM, the first SLM with dynamic, controllable frame rates, using pretrained FlexiCodec for dynamic speech output tokens. It integrates this representation into a multi-task speech-to-speech SLM, extends it with input-side frame compression, and adds direct frame-rate conditioning for accurate control during inference. FlexiSLM outperforms fixed-frame-rate 7B models, including Qwen2.5-Omni and Kimi-Audio, at 12.5 and 6.25 Hz; it can be steered down to 4.0 Hz, and at 6.25 Hz roughly halves inference time relative to 12.5 Hz while retaining strong speech-to-speech quality. Audio samples: https://flexislm.github.io; code and data: https://github.com/AmphionTeam/FlexiSLM.

cs.SD↗

arXiv2026

ParaBridge: Bridging Paralinguistic Perception and Dialogue Behavior in Speech Language Models

Speech carries more information than just words: a child's voice, a fearful tone, or a noisy background should all lead a sufficiently competent spoken-dialogue assistant to different replies. Current Speech Language Models (SLMs) can recognize such paralinguistic cues but often ignore them in open-ended dialogue. We observe that a simple paralinguistic instruction scaffold at the inference stage narrows this perception-behavior gap, suggesting that the relevant cues are already latent in the model. Such scaffolds, however, remain brittle under multi-turn context and competing instructions. Therefore, we propose \textbf{ParaBridge}, an on-policy self-distillation method that turns a brittle inference-time scaffold into stable model behavior. During training, the scaffold serves only as a temporary privileged view; the scaffold-free model rolls out its own response, while the scaffolded view supplies dense, full-vocabulary next-token targets along its trajectory. This supervision teaches when non-lexical cues should affect the reply without the need for curated dialogues, human labels, or external reward models. On Qwen3-Omni-thinking, ParaBridge raises scaffold-free VoxSafeBench SAR from $14.6\%$ to $40.3\%$ and improves EchoMind average rating from $3.27$ to $3.92$. It also preserves general ability, with MMAU-Pro, VoiceBench, and GPQA all within $0.4$ points of the original model. Beyond the training distribution, ParaBridge generalizes to unseen paralinguistic cues, transfers from safety-oriented training to empathy-oriented dialogue, and works on a different SLM backbone.

cs.CL↗

arXiv2026

RecurTrace: Adaptive Latent Reasoning with Loop-Time Memory

Repeating a small block of middle layers increases a language model's effective inference depth without adding parameters or generating extra tokens, and recent work shows that this latent recurrence improves reasoning. However, two design choices limit these gains. Each iteration sees only the previous output and cannot directly access earlier computations. Moreover, a fixed loop count wastes depth on easy inputs while leaving hard ones with too little computation. We introduce RecurTrace, which addresses both limitations using the loop's own trajectory. Specifically, Loop Memory Attention lets each looped layer attend to its own states from previous iterations along the loop-time axis, so the model can revisit earlier computations instead of relying on the latest state alone. A halting head then reads the loop state and predicts whether to continue, with supervision from an oracle that identifies when additional depth still reduces loss. In a controlled MathQA comparison on the same looped backbone, RecurTrace achieves 56.9% accuracy with an average of 2.0 loops, exceeding the best fixed loop depth by 2.2 points at matched compute. By comparison, ACT and PonderNet collapse to one loop, and CALM reaches only 54.1% with 5.6 loops, while the stronger LoopUS-Conf and TaH-Mismatch baselines reach 55.3% at 3.2 loops and 55.7% at 2.1 loops. Finally, RecurTrace improves generation accuracy over same-budget fine-tuned baselines at 0.6B, 1.7B, 4B, and 8B, with the gain growing with model size from 0.6 to 3.4 points.

cs.LG↗

arXiv2026

Frequency-Aware Self-Supervised Music Representation Learning

Self-supervised learning (SSL) has emerged as an essential paradigm for music information retrieval (MIR). While current SSL models achieve state-of-the-art performance across various MIR tasks, they typically treat audio as 1D sequences, either operating on time-domain waveforms or on flattened time-frequency-domain spectrograms. This discards the rich spatial and structural information in time-frequency representations and overlooks a fundamental intuition in music production. In particular, music is naturally represented as time-frequency grids in MIDI-based workflows, a structure that tightly corresponds to 2D spectrograms and inherently makes many MIR tasks trivial. Motivated by this intuition, we propose PupuM2D, a musical Masked Modeling Duo (M2D) model that is trained directly on 2D music spectrograms. Instead of applying masked language modeling (MLM) to 1D sequences, PupuM2D learns robust representations by predicting the latent embeddings of masked 2D music spectrogram patches from unmasked contexts. To optimally adapt such a framework to music signals, we also apply domain-specific modifications to model architecture, training scheme, and inference paradigm, with comprehensive ablation studies showing their effectiveness. Evaluations on the MARBLEv2 benchmark show that PupuM2D outperforms the sequence-based SSL models across multiple MIR tasks in linear probing. Additionally, case studies of the attention maps also confirm that PupuM2D captures musically meaningful patterns within the 2D time-frequency domain. Codes and checkpoints are available at: https://www.yichenggu.com/PupuM2D/.

cs.SD↗

arXiv2026

VoxPrivacy: A Benchmark for Evaluating Interactional Privacy of Speech Language Models

As Speech Language Models (SLMs) transition from personal devices to shared, multi-user environments such as smart homes, a new challenge emerges: the model is expected to distinguish between users to manage information flow appropriately. Without this capability, an SLM could reveal one user's confidential schedule to another, a privacy failure we term interactional privacy. Thus, the ability to generate speaker-aware responses becomes essential for SLM safe deployment. Current SLM benchmarks test dialogue ability but overlook speaker identity. Multi-speaker benchmarks check who said what without assessing whether SLMs adapt their responses. Privacy benchmarks focus on globally sensitive data (e.g., bank passwords) while neglecting contextual privacy-sensitive information (e.g., a user's private appointment). To address this gap, we introduce VoxPrivacy, the first benchmark designed to evaluate interactional privacy in SLMs. VoxPrivacy spans three tiers of increasing difficulty, from following direct secrecy commands to proactively protecting privacy. Our evaluation of nine SLMs on a 32-hour bilingual dataset reveals a widespread vulnerability: most open-source models perform close to random chance (around 50% accuracy) on conditional privacy decisions, while even strong closed-source systems fall short on proactive privacy inference. We further validate these findings on Real-VoxPrivacy, a human-recorded subset, confirming that failures observed on synthetic data persist in real speech. Finally, we demonstrate a viable path forward: by fine-tuning on a new 4,000-hour training set, we improve privacy-preserving abilities while maintaining robustness. To support future work, we release the VoxPrivacy benchmark, the large-scale training set, and the fine-tuned model to foster the development of safer and more context-aware SLMs.

eess.AS↗