voice assistant engineers

Abstract: Accurate prediction of human Mean Opinion Scores (MOS) is crucial for evaluating synthetic speech quality in text-to-speech (TTS) and voice conversion (VC). Most MOS predictors output only point estimates and ignore uncertainty, limiting reliability in model selection and deployment. We propose ConformalMOS, a framework that augments MOS prediction with conformal prediction to produce statistically valid prediction intervals with guaranteed coverage under exchangeability, alongside point estimates. During training, MOS labels are modeled in an ordinal-aware manner by converting one-hot targets into Gaussian-smoothed distributions, preserving score ordering. This improves uncertainty calibration, particularly when softmax confidence is overconfident on out-of-distribution speech. On BVCC datasets, ConformalMOS achieves competitive accuracy (MSE = 0.08) while delivering empirically validated coverage, enhancing reliability for real-world TTS and VC deployment.

Abstract: Reinforcement learning can improve text-to-speech synthesis through feedback on generated speech. However, sentence-level rewards do not identify local errors, while signed regression updates can lead to excessive fitting or repulsion of sampled outputs. We propose LoCA (Local Credit Assignment) for autoregressive flow-matching text-to-speech. LoCA uses ASR-based text alignment to locate content errors in generated speech and redistribute sentence-level advantages across acoustic frames. To improve training stability, we introduce reference-relative positive and negative thresholds that limit continued fitting and repulsion. Experiments on VoxCPM2 show that LoCA reduces sample-weighted content error on a multilingual CV3 evaluation subset by 20.5% and 14.3% relative to the pretrained backbone and GROW, respectively, while improving speaker similarity. LoCA also improves content accuracy on the Seed-TTS-Eval hard subset. Ablation results support the contributions of local credit and threshold-based regression control.

Abstract: Preference optimization promises to improve text-to-speech (TTS) systems by learning which speech outputs are better. External expert speech seems a natural source of preferred responses, especially when student generations contain errors. Yet demonstration quality does not establish whether the student can benefit from that supervision, or whether better target fitting will produce better speech. We examine this gap across two TTS students using real and synthetic expert speech, comparing preference objectives and direct versus bridge supervision while tracking fixed-sequence likelihoods and generation metrics. With expert supervision held fixed, objectives differ markedly in their sensitivity to preference-score scaling and the resulting generation quality. Likelihood diagnostics characterize these responses, but cannot establish their benefit: expert likelihood can increase while intelligibility worsens, or decrease while it improves. Bridge supervision highlights a distinction between mitigating degradation and improving the student: it can recover performance lost under direct training without consistently surpassing the pretrained base. Route benefits vary across objectives, students, and expert sources; neither route consistently dominates. Together, these findings distinguish demonstration quality, target fitting, and generated-speech quality as complementary criteria for assessing expert supervision.

Abstract: Proactive dialogue agents must guide multi-turn interactions toward task goals while adapting to users whose intentions and attitudes evolve. Existing planners improve system side action selection through search, learned policies, or user aware planning, but are often trained and evaluated against fixed or overly compliant simulators; as a result, high simulated success can reflect soft user bias rather than robust dialogue ability. We introduce MASP, a Mentalized Adversarial Self Play framework that treats proactive dialogue as a partially observable game between a system and a rationally resisting user. MASP represents the user's evolving mental state with a BDI based latent state, separates reward side teacher mentalization from policy side student mentalization, and alternates self play updates for the system and user policies. This turns robustness to rational resistance into a training objective rather than only an evaluation condition. Across diverse proactive dialogue benchmarks, compared with the widely used SFT and Planner-only approaches, MASP substantially improves success rate (SR) and rationality constrained success (RCSR), reduces average turns (AT), and performs better against hard users who make the task goal more difficult to achieve.

Abstract: Long-term memory remains a fundamental bottleneck for large language model (LLM) agents across sessions. Existing systems typically adopt a write-then-retrieve paradigm, storing context-dependent dialogue fragments and retrieving them via nearest-neighbor search in a continuous embedding space. This design yields unstable access units and systematically misses logically related but lexically distant evidence. We propose MemCode, a long-term agent memory framework that couples retrieval-oriented writing with discrete associative retrieval. MemCode rewrites dialogue into self-contained atomic memory units (AMUs), maps AMUs into a learnable codebook with isomorphic semantic quantization (ISQ), and uses codebook co-occurrence topology as a retrieval path complementary to vector search. For long-term memory stores, MemCode separates online atomic writes from asynchronous offline consolidation. On LongMemEval-S and LoCoMo with GPT-4o-mini and Qwen3-30B-A3B-Instruct-2507 backbones, MemCode achieves the best overall accuracy while keeping per-dialogue memory construction within roughly 40~130 seconds, making it 3~21 times faster than LightMem and over an order of magnitude faster than A-MEM and MemoryOS.

Abstract: Code-switching (CS) speech, characterized by language alternation within a single utterance, is prevalent in multilingual communities and poses challenges to automatic speech recognition (ASR) due to rapid language transitions. We propose LLM-HB, a language-aware large language model-guided hotword biasing framework for code-switching ASR (CS-ASR). The framework incorporates a mixture-of-experts (MoE) adaptor and an auxiliary language prediction head to better capture and differentiate multilingual representations, while hotwords are injected through LLM-based prompting. Experiments on the ASRU2019 code-switching challenge test set demonstrate that LLM-HB achieves a 20.30% relative reduction in mixed error rate (MER) over the baseline, reaching 5.85% MER under a 15-distractor hotword setting. To the best of our knowledge, this is the first work to integrate LLM-guided biasing into code-switching ASR. We further release the extracted ASRU2019 hotword list to support future research.

Abstract: Speech representation learning requires a discrete modeling interface while preserving high waveform fidelity. Existing approaches rarely achieve both: discrete tokens ease modeling but lose fine-grained acoustic information, while continuous latents preserve fidelity but require specialized generative objectives. We introduce SemiCodec, a novel Semi-discrete neural speech Codec with synchronized discrete tokens and continuous latents. Direct joint optimization may create a reconstruction shortcut through the continuous pathway, leaving the discrete tokens insufficiently grounded. To address this issue, we propose a Preserve-then-Ground training paradigm that first establishes a high-fidelity synthesis space and then grounds SSL-guided discrete tokens within it. Experiments show that SemiCodec outperforms fully discrete codecs in reconstruction quality and intelligibility while remaining competitive with continuous representations in acoustic fidelity and speaker preservation. Audio samples are available at https://pennyw1-se.github.io/SemiCodec-Demo-Page/

Abstract: Speech large language models (LLMs) are often built using pretrained acoustic encoders as an audio interface to an LLM decoder. However, LLMs scale poorly with input sequence length, while acoustic encoders tend to produce substantially longer sequences than typical textual inputs. This motivates learning to compress encoder outputs while preserving task-oriented information. To this end, we introduce ASSET, a reinforcement learning (RL)-based framework that learns to adaptively compress encoder outputs by first partitioning them into variable-duration segments and then pooling each segment into a continuous audio token. We use group-relative policy optimization (GRPO) to train a segmenter with task performance feedback and a token-rate penalty. In automatic speech recognition (ASR) experiments, ASSET achieves performance comparable to a transcript-assisted character-boundary oracle with approximately 25% fewer audio tokens on English LibriSpeech, while improving on a phone-boundary oracle at a similar token rate on Mandarin AISHELL. Exploratory experiments on non-ASR tasks further illustrate how learned segmentation changes when the framework is adapted to tasks beyond transcription.

Abstract: Large audio-language models (LALMs) exploit multimodal evidence, yet task-irrelevant audio can alter text-reasoning decisions when listening is unnecessary. Aggregate Accuracy can hide this paired drift because audio-induced repairs and damages may cancel. Paired drift analysis and targeted interventions identify architecture-specific, intervention-sensitive late audio pathways as actionable control points. We introduce ICAP-Gate, which applies mechanism-guided, task-conditioned control to each model's pathway. Across four LALMs, two reasoning benchmarks, and environmental-sound and natural-speech interference, ICAP-Gate has lower point estimates for Influence Rate and Answer Flip than ungated inference in all 16 full-split model--condition evaluations. Fixed suppression degrades automatic speech recognition (ASR) across all four models, whereas ICAP-Gate matches ungated ASR performance by preserving the pathway for explicit audio-demand instructions. ICAP-Gate has lower paired-drift point estimates than mitigation prompting in all four evaluated settings and provides competitive stabilization relative to eight-sample Self-Consistency while using one generation per query; in controlled ARC measurements, Self-Consistency incurs $7.0$--$9.2\times$ ungated latency. These results establish selective modality influence control as a design principle for robust multimodal reasoning.

Abstract: Speech quality predictors are increasingly used as rewards, yet no agreed measure of their hackability exists. The usual measurement has two flaws. First, the perturbation reaches the predictor through a processing chain -- here a neural codec -- that shifts the score on its own, which scoring against the raw input charges to the attack. Referencing the unperturbed round trip instead changes measured hackability by up to a factor of four (0.31 to 0.08 for one defence). Second, one trained attacker is a sample, not a measurement: five attackers differing only in random seed reach success rates from 0.00 to 0.38 against one fixed predictor, so a defence claim needs the worst case over several. Under this protocol, four published predictors differ widely: NISQA is hacked on 90% of utterances, SSL-MOS on 21%, DNSMOS on 14% and UTMOS on 6%. We then audit a closed attack-detect-patch loop. It hardens the predictor only in its own attack space, by less than the spread between attackers; a random-perturbation baseline matches it; and it costs up to 0.30 system SRCC out of domain. Enhancers post-trained against patched predictors hack them far less (PESQ -0.03 versus -0.23). Code, preregistration and run outputs are released.

Abstract: Vision-language-action models (VLAs) are strikingly sensitive to instruction phrasing and do not inherit the language robustness of the vision-language models they are built on. A one-word edit can move success by tens of points: $π_{0.5}$ turns on a LIBERO stove 100% of the time for "switch on the stove" and 2% for "switch on the hot plate", and a $π_0$ checkpoint finetuned with rephrase augmentation still shows swings of up to 61 points. We characterize this sensitivity with statistically tested single-edit swings and an oracle phrase search, which shows that phrasing alone nearly closes the 21-point gap between in-distribution and out-of-distribution tasks. We then reduce it without modifying the policy. Because the sensitivity is systematic, it can be expressed as explicit rules: we score many phrasings of a few training tasks, have a large language model distill the evidence into ten to twenty rephrasing rules, and at deployment rewrite each incoming instruction once under these rules. The rules improve the frozen $π_0$ by 16 to 27% relative on twelve held-out tasks across adversarial, VLM-generated, and human-generated phrasings, with gains concentrated on out-of-distribution tasks. The pipeline replicates on $π_{0.5}$ and LIBERO, lifting in-finetune success from 93.6% to 97.8%. The method requires no retraining and no per-step verification, and applies zero-shot to unseen tasks and instructions. Project website: https://sttawm.github.io/rephrase-before-you-act

Abstract: Instruction text-to-speech (ITTS) systems systematically encode acoustic gender skews from descriptive style prompts, such as occupations or personas, even when demographic attributes are left unspecified. Calibrating the implicit gender distribution of the synthesized voices, across heterogeneous architectures and without model retraining or overriding explicit user prompts, is an open problem. In this work, we propose \textit{model-adaptive steering}, a training-free bias calibration method that steers post-encoder conditioning representations using a group deviation vector paired with a coarse-to-fine strength search on a development set. A deterministic lexical gate bypasses intervention whenever explicit gender keywords are detected, preserving intended prompt semantics. Evaluated on a 12,800-prompt held-out benchmark across four ITTS models (three architectures), the proposed method reduces aggregate calibration error from 11.5--28.1 to 0.8--5.9 percentage points (e.g., shifting Parler-TTS Mini from 78.1\% and PromptTTS++ from 27.3\% female to 50.8--52.4\%). Output rates stay within 0.9 pp across anchor set sizes at fixed operating points, while UTMOS decreases by at most 0.08 and WER degrades by at most 3.3 pp.

Abstract: Unlike autoregressive models, discrete diffusion-based models for zero-shot text-to-speech generate speech tokens in parallel and can revisit earlier predictions. Mask-and-replace training extends mask-only training by randomly replacing some tokens, and its gains are commonly attributed to self-correction, the ability to revise previously generated tokens. However, exposure to randomly perturbed context during training may itself improve generation, raising the question of whether these gains require inference-time token revision. To investigate this question, we use DeMaR, which combines mask-and-replace training with confidence-ranked mask-only sampling while preserving the total training corruption probability. Trained from scratch on LibriTTS, DeMaR achieves lower word error rates (WER) than autoregressive and mask-only diffusion baselines using the same speech tokenizer. This advantage persists when each token remains unchanged after first being unmasked. Matched training conditions on two heterogeneous speech tokenizers show that both noisy-context augmentation and replacement supervision improve WER under this restriction. These findings demonstrate training-side benefits of replacement beyond enabling inference-time token revision.

Abstract: Prefill-only decision models inspired by the Jev model score every candidate in a menu during a single forward pass and never decode, which makes one call one to two orders of magnitude cheaper than a same-scale generative language model. We show that this read-out structure comes with a testable property. When an intervention changes only the candidate menu and leaves the input text fixed, the post-intervention accuracy is already determined by the cached first-pass distribution. The estimator restricts the pass-1 probabilities to the menu, renormalizes, and reads off the argmax; it uses no labels and no second forward pass. Across seven model families, ten datasets and two task types, menu-only interventions are predicted to within 4.2 points, and for one family the prediction is exact. A probability-level variant of the same estimator errs by 21.0 points, so the property lives in the ranking rather than in the probabilities and is not recovered by calibration. Same-scale generative language models do not share the property. On those models the same estimator errs by 1.6 to 15.8 points and degrades as the model grows. The property turns inference-time compute into a decision that can be made before deployment. Uniform extra passes buy calibration but almost no accuracy; at matched cost a confidence cascade outperforms every scheme that re-asks the same model, and curating the menu beats enlarging the model, with a 0.8B model on a curated 5-candidate menu reaching 95.4% on CLINC150 against 80.0% for a 4B model on the full 150-label menu.Code and data are available at https://github.com/rlisml/jev-cascade.

Abstract: Frozen speech foundation models (SFMs) make downstream adaptation efficient: the backbone can stay fixed while a task learns layer fusion and a lightweight classifier. Full adversarial fine-tuning is a standard route to robustness, but generating adversarial examples and updating the backbone for every task sacrifices that efficiency. We ask whether robustness can instead be learned before future tasks are known. For a frozen backbone and linear classifier, robustness can be understood through the interaction between representation stability and decision-boundary margin. This leads directly to our design: we stabilize representations across the hidden layers, rather than only the final layer, while preserving clean representations; after clean adaptation selects the layer mixture, we keep it fixed and enlarge only the classifier margin, without downstream adversarial examples. We evaluate Wav2Vec2, HuBERT, and WavLM Large on four tasks under adaptive 30 dB attacks. Across 12 backbone-task pairs, hierarchical robustification improves robust accuracy by 46.4 pp, while margin refinement adds 4.0 pp for 1.1 pp of clean accuracy. Code and configurations are available at https://github.com/arefmousavi/hierarchical-robust-sfm.

Abstract: Knowledge distillation (KD) improves low-resource acoustic learning by enriching one-hot supervision with the softened predictive distribution of a fixed teacher network. However, a teacher trained with limited or imbalanced annotations may produce a biased distribution whose components are not uniformly reliable. Although this distribution can still encode useful knowledge, direct full-distribution matching may also transfer teacher-induced biases, thereby distorting the student's decision boundary and degrading its generalization performance. To address this limitation, we propose Boundary-Anchored Mass-Partitioned Distillation (BA-MPD), a logit-based distillation objective composed of Boundary-Anchored Correction (BAC) and Mass-Partitioned Distillation (MPD). BAC addresses missing ground-truth labels in the set of the teacher's top predictions by swapping the true label for the lowest-ranked entry of the set, thus keeping the mass and uncertainty of the set unchanged. MPD then distills this corrected distribution through separate losses that enforce relational consistency within the set, balance the mass between high- and low-confidence groups, and weight lower-confidence dependencies. Ultimately, BAC and MPD together suppress harmful ranking errors and noisy low-confidence details, while retaining all useful teacher information. Experiments on two acoustic benchmarks under multiple label budgets show that BA-MPD consistently improves over supervised-learning baselines and vanilla KD while remaining competitive with strong logit-based KD baselines. Cross-budget results further show that BA-MPD remains effective when the teacher and student models use mismatched label budgets, demonstrating its ability to exploit imperfect teachers across supervision gaps. Implementation available at https://github.com/ShuanglinLi/BA-MPD.

Abstract: Synthetic dialogues generated by LLM pipelines now serve as complete conversational-recommendation benchmarks: an LLM listener talks to an LLM recommender, and the track logged next in the conversation becomes the official label for each turn. These policy-selected labels make large-scale evaluation reproducible, but they are proxies for what the simulated user asked. We audit the one place where label and request are directly comparable: turns where the user asks for an exact song by name. In the RecSys Challenge 2026 TalkPlay benchmark, using visible dialogue and catalog metadata alone, we find that the official label contradicts the user's exact-song request in half of the audited development turns. This matters beyond one benchmark: naming the desired item is the dominant intent in real music search, where deployed systems avoid substituting an alternative for an exactly named item, on the premise that it costs satisfaction. A small training-time supplement closes most of the gap: adding catalog-resolved request-satisfying targets to a small fraction of training turns yields a 53.3% relative gain in nDCG@20 on the 43 conflict turns while leaving the official metric intact, verified against a matched control that detects the same requests but trains only on official labels.

Abstract: Speculative decoding accelerates autoregressive generation by using a smaller drafter to propose tokens for batched verification by a larger target. However, conventional speculative decoding couples drafting to the target's evolving verified prefix, serializing drafting and verification. We ask whether this dependency is necessary for source-conditioned generation. Our key observation is that, for audio language models, the input audio and user request can provide useful speculative candidates without following the target's evolving text prefix. We propose AS$^2$D (Audio Speculative Speculative Decoding), which enables target-decoupled drafting: an audio-conditioned drafter follows its own generation history while the target independently verifies and corrects ready candidates. Without usable candidates, the target advances alone. Thus, target feedback determines which candidates are committed but no longer determines when the drafter can make progress, enabling drafting and verification to proceed concurrently while retaining target-side verification and correction. We implement AS$^2$D in MNN for Android and evaluate two target models across four phones, seven datasets, and three tasks covering 12.2 hours of audio. Across four phones, AS$^2$D improves pooled ASR throughput by 42-76% over target-only decoding, while only 5.7% of evaluation windows are slower than target-only, compared with 58.1-63.0% for speculative baselines. For ASR, AS$^2$D reaches 97.33-98.20% of a hindsight per-window oracle's pooled throughput over the evaluated drafter/budget catalog. Native on-demand execution with a 7B target achieves up to 78% higher throughput than target-only. These results show that source-conditioned audio generation can relax the conventional dependence of speculative drafting on the target's evolving output prefix, exposing substantial parallelism for efficient inference.

Abstract: Text-to-speech models loop, truncate and lose count on text that repeats a phrase many times. We show that repetition itself is what breaks them, not the length that comes with it. Every repeated sentence in our test set is paired with a control of matched sentence and word count in which no word ever repeats back-to-back. Six models from three architectures render the controls almost perfectly and fail the repeated twins: 94.3% against 18.2% exactly right at k >= 6. The gap survives greedy decoding, repetition-penalty sweeps, four independent speech recognisers and 420 analysis specifications without once reversing sign; a held-out fourth architecture lands within a point of its predicted gap, and one of two non-autoregressive baselines shows the same failure. Varying the period of the text shows the failure grows smoothly with periodicity, half of it surviving when no word is adjacent to itself.

Abstract: Transcribing domain-specific entities and rare proper nouns remains a major challenge in automatic speech recognition (ASR). In this paper, we propose BaLEEN (Biasing with Latent Encoded Entities), a lightweight, hypernetwork-based framework for dynamic contextual adaptation without fine-tuning the underlying ASR model. BaLEEN encodes variable-length contextual keywords using a pretrained language model, compresses them into a fixed sequence of latent vectors via a Perceiver bottleneck, and injects context-dependent bias vectors directly into the intermediate encoder representations of the ASR model. Because both the language model and the backbone ASR model remain entirely frozen during training, BaLEEN operates as a plug-and-play adapter that incurs zero computational overhead at inference time when context biases are precomputed. We evaluate our method on a CTC-based ASR model using a Wikipedia-derived corpus with annotated named entities and synthetic speech. Experimental results demonstrate that BaLEEN reduces keyword miss rate by 8.7% on the test set relative to the unbiased baseline while simultaneously improving overall word error rate by 21% and character error rate by 28%.

Abstract: We present Cardamom, a micro-dialectal Arabic speech dataset designed to support fine-grained evaluation and adaptation of automatic speech recognition (ASR) systems. Community-curated by native speakers familiar with the represented varieties, Cardamom contains approximately 40 hours of transcribed YouTube speech spanning 21 micro-dialects across Egypt, Jordan, Lebanon, Mauritania, Palestine, and Saudi Arabia. Each segment is annotated with one or more operational micro-dialect labels, code-switching information, and utterance-level perceived gender, enabling analysis of sub-country variation that is obscured by conventional country-level labels. We describe the collection and annotation process, motivate the micro-dialect inventory linguistically, and benchmark four multilingual ASR systems in zero-shot and adapted settings. The strongest zero-shot system obtains 43.47% aggregate WER, with particularly high error rates on Mauritanian and Lebanese varieties; adaptation on Cardamom reduces its WER to 35.21%. Audio-based identification experiments further show that the annotations provide a learnable prediction target, with a dedicated classifier reaching 85.57% accuracy on 21-way micro-dialect identification. Cardamom provides a resource for studying localized dialectal variation and developing Arabic speech systems with broader regional coverage.

Abstract: Large audio-language models (LALMs) produce fluent responses about audio but often hallucinate by making plausible yet ungrounded claims. Existing audio hallucination benchmarks mainly measure response correctness, leaving it unclear whether an LALM hallucinates or simply fails to understand the audio. We challenge correctness-based evaluation by defining two hallucination categories: (i) context, where claims are not grounded in the audio; and (ii) knowledge, where claims about audio-related topics lack support from externally verifiable facts. We introduce MISHAP-Bench, a comprehensive benchmark with 12,000 challenging open-ended question-audio pairs and a rigorous evaluation pipeline covering both categories. To evaluate open-ended responses, we propose a groundedness judge that uses reference rubrics and judge prompts guided by human annotations. We extensively evaluate ten state-of-the-art LALMs and show that hallucination remains substantial. Even a frontier model such as Gemini 3.7 Flash reaches a hallucination rate of 36.5%. We further adapt and benchmark four mitigation methods from multiple domains for LALMs. Despite some improvements, effective hallucination mitigation remains an open challenge. Finally, we call on the community to evaluate hallucination and benchmark mitigation methods with MISHAP-Bench.

Abstract: Recent advances in speech deepfake detection (SDD) have leveraged the Mixture of Experts (MoE) to enhance generalization capacity. However, existing gating networks often overlook the acoustic and temporal cues of deepfakes. In this work, we propose a novel domain-adaptive dual-gating MoE (DADGMoE) framework for SDD under unseen attack types and acoustic conditions. Our innovative dual-gating mechanism leverages Sinc-layer-based filters to process both low-level acoustic signals (raw waveforms) and high-level speech representations from a large self-supervised learning (SSL) model. It further incorporates domain prototypes to guide expert routing based on implicit deepfake patterns. The lightweight affine experts process the routed inputs. Experiments show that our DADGMoE significantly outperforms the baseline, achieving up to a 40.8% relative EER reduction on challenging out-of-dataset benchmarks. This framework demonstrates superior generalization capabilities and efficient design.

Abstract: Whisper is still what runs in production: one permissively licensed checkpoint, 99 languages, no per-language tuning. But it writes sentences nobody said. On 42 clips of pure room tone, whisper-large-v3 emits words on 61.9% of them and emits something on 100%. The usual response is to distil a student from Whisper pseudo-labels, filtering the hallucinations out of the corpus first, which deletes the evidence while leaving the behaviour in place: the student is fitted to the subset where the teacher was right, and inherits a failure mode absent from its own training data. The second response, adding non-speech audio so the model learns to stay quiet, already ships and works, but teaches suppression without discrimination. The checkpoint best on every non-speech arm here is also the one that recovers the fewest genuinely spoken phrases and deletes 58% of repeated speech. The teacher has to be fixed, with both halves of the signal. Our benchmark scores both at once: 11,852 clips over eight arms, 6,267 synthetic positives, 8,296 clips of real audio that made a production model fail, and FLEURS in 58 languages. We collect 40,891 hallucination phrases in 100 languages, choose a text-to-speech system by measurement, and synthesise those phrases as positives, so the model meets the same text both as something to suppress and as something to transcribe. Across 33 matched pairs of fine-tunes differing only by those positives, adding them lowers word emission on real voice-free audio in 31 pairs and raises phrase recovery in 32; among the 30 pairs that hold English accuracy it is 30 out of 30. The best checkpoint takes hallucination on silence from 61.9% to 2.4% and words over real voice-free audio from 99.9% to 47.8%, while raising phrase recovery from 69.8% to 82.7%. The benchmark, lexicon and synthetic corpus are released at https://huggingface.co/datasets/Scicom-intl/Whisper-Hallucination

Abstract: Full-duplex speech language models continuously accumulate acoustic key-value (KV) states, making long-running interactions memory-intensive. During listening, the model can finish processing an audio unit before the next arrives; we term the remaining interval listening-time slack. We propose acoustic-to-text KV compression, which introduces a transcription side channel to convert incoming speech into compact textual memory within this interval. When the cache exceeds a target budget during inference, older acoustic states are evicted while transcripts and recent acoustic context remain. We train the side channel with LoRA using cross-entropy on transcription segments. To preserve listening and speaking behavior, we apply knowledge distillation to the original model's token-level output distributions at native prediction positions. On ten-minute LongSpeech sessions, our MiniCPM-o 4.5 implementation reduces peak streaming KV-cache size by 64.6% compared with the same model without eviction. The proposed method also improves transcription, temporal question answering, and summarization over the baseline. Full-Duplex-Bench evaluations further show comparable pause-handling, turn-taking, and interruption performance.

Abstract: We describe our submission to the Unsupervised Speech in the Wild (UPS) Challenge at Interspeech 2026, a bidirectional Mamba-2 (BiMamba2) encoder trained with masked discrete-unit prediction following the HuBERT-style paradigm. The 47.88M-parameter model is trained on 250 hours of speech across 67 languages from the MLCommons Unsupervised People's Speech dataset, with no labeled data. The objective combines masked k-means pseudo-label prediction with language identification supervision and VICReg regularization. On official evaluation, the system achieves an Adjusted Rand Index of 0.735, exceeding four baselines on speaker clustering. Language identification macro-F1 (0.073) and character error rate (0.870) remain below supervised baselines. We analyze a local-official discrepancy in metric scale and checkpoint ranking, highlighting limitations of in-distribution diagnostics for predicting Dynabench probe outcomes.

Abstract: Contextual biasing improves rare-word recognition in speech large language models (SpeechLLMs), but efficiently exploiting large bias lists remains challenging. We propose PTC-Bias, a two-stage framework based on phoneme-level temporal competition. At the prefill stage, PTC Retrieval performs frame-synchronous phoneme decoding and temporal competition among candidate pronunciations, producing a compact bias-word shortlist and corresponding speech intervals. After SpeechLLM decoding, PTC Correction conducts a second local competition between the retrieved candidates and mismatched transcript spans within these intervals. Selective correction reduces near-homophone and word-segmentation errors while preserving correct transcriptions. Both stages share the same phoneme posteriors and require no additional SpeechLLM forward pass. Experiments on LibriSpeech show consistent gains across two SpeechLLMs and bias lists of up to 2000 words. With Prompt-SLAM-ASR-7B and 2000 bias words, PTC-Bias reduces B-WER by 23.4%/23.9% relative to CTC-Filter on test-clean/test-other, while keeping U-WER nearly unchanged.

Abstract: Speech technology penalizes some voices: recognition errs nearly twice as often for Black speakers, and accuracy declines for second-language accents and older speakers. We introduce TRIAD, an audit grid crossing 120 texts, 24 rendered demographic voice profiles (gender, age band, accent), and ten expressive styles via controllable text-to-speech, isolating perceived demographic attributes from content and affect. For ten open-weights encoders we define axis-fidelity functionals, principal-angle leakage between axis subspaces, and group-conditional gaps; a proposition proves that average probe disparity grows with the same aggregate voice-semantic leakage $Λ$ we measure, and a corollary shows that peak leakage forces worst-case disparity inside an active region. The measured mean-square probe disparity tracks $Λ$ (Pearson r = 0.93), and a black-box protocol exposes the same signature in two closed-source models. ORCA, an adapter combining axis-specific contrastive heads, an orthogonality penalty, and group-balanced sampling, cuts leakage 72% and roughly halves the gaps.

Abstract: Semi-supervised federated learning (SSFL) trains models on clients' unlabeled data using a teacher to generate pseudo-labels, with a small labeled seed dataset on the server. Automatic Speech Recognition (ASR) is particularly fragile here: pseudo-label errors compound across the output sequence and across training rounds into divergence, leaving a large gap to fully-supervised FL. We show that closing this gap turns on two coupled design axes -- the teacher (which model generates the pseudo-labels) and the anchor (the server-side updates on labeled data that stabilize training). On the teacher axis, a per-client online teacher (each client's own evolving model) diverges on its own, but once stabilized it matches or beats the broadcast global teacher (one server model, fixed within a round) -- decisively in-domain and competitively under domain shift. As the seed grows stronger and the online teacher's advantage narrows, a transitioning teacher (global $\rightarrow$ online at round $r$) matches or beats both. On the anchor axis, the server must keep training on labeled data between rounds -- otherwise the online teacher drifts -- and this interleaving, more than the seed model, governs convergence. The two axes are inseparable: aggressive teacher choices pay off only once the anchor stabilizes training, which is highly sensitive to data augmentation and batch size -- the settings that govern how much input and gradient noise the server injects. How much stabilization is needed is domain-dependent, governed by the dispersion of the seed data and its overlap with client data. These findings yield guidelines for SSFL in ASR training, improving over the strongest prior method on 9 of 11 pairs, by $20.8\%$ on average in-domain and $10.0\%$ cross-domain, narrowing the gap to fully-supervised FL.

Abstract: Autoregressive generation of interleaved text and acoustic tokens is a common approach to spoken-response generation in speech large language models. Although this design enables streaming generation with explicit textual guidance, generated acoustic tokens become part of the context for subsequent text predictions. Given identical speech inputs, we observe markedly lower answer accuracy for the internal text generated in speech-to-text-and-speech (S2TS) mode than for speech-to-text (S2T) responses. We term this discrepancy the \emph{output-mode gap} (OMG). To reduce OMG, we propose \emph{Joint-Output On-Policy Distillation} (JO-OPD), which distills the model's stronger S2T policy into joint generation using student-generated S2TS trajectories. At each text position, the S2T teacher provides soft targets from a text-only projection of the student's preceding outputs, while the student predicts from the corresponding full interleaved history. A preservation objective further regularizes native non-text predictions. Experiments on Step-Audio-2-mini and Baichuan-Audio-Instruct reveal OMG across two interleaved generation architectures. On Step-Audio-2-mini, JO-OPD reduces OMG from 42.87 to 16.26 percentage points on Spoken-MQA and from 29.72 to 13.04 points on speech-rendered GSM8K, with little change in S2T accuracy and substantially larger reductions than matched SFT baselines. ASR-based evaluation further shows a 7.49-point improvement in spoken-answer accuracy on Spoken-MQA.

Abstract: Continuous-latent Autoregressive Diffusion Transformer (AR-DiT) models have demonstrated immense potential in zero-shot speech generation. However, they still suffer from limited decoding stability when synthesizing long utterances or complex linguistic structures. This instability primarily stems from a restricted historical receptive field and an acoustic inertia dependency within the diffusion decoder, which causes the model to ignore semantic conditions. To address these challenges, we propose DiTAR+, a dual-optimization framework. First, we introduce Dilated Context Sampling to expand the macro-level historical receptive field without violating physical temporal continuity, thereby preventing cumulative error propagation. Second, we propose Hierarchical Acoustic Masking to prevent shallow layers from attending to acoustic pre-context, explicitly decoupling semantic alignment from acoustic detail reconstruction. Extensive experiments show that our framework effectively mitigates pronunciation errors and semantic hallucinations, enhances generation robustness on challenging sentences, and maintains exceptionally high speaker similarity throughout the entirety of long-form utterances. On the linguistically challenging ZH-Hard set, DiTAR+ reduces the word error rate from 12.478% to 9.893%, and on extended utterances of 25 to 35 seconds it improves speaker similarity from 0.741 to 0.759 while simultaneously lowering the word error rate from 2.778% to 2.173%, outperforming both discrete-token and pure flow-matching baselines.

Abstract: Speech-to-speech translation (S2ST) has advanced significantly with speech LLMs, offering the potential for joint optimization and preserving non-linguistic information. However, these models struggle with predicting high-bitrate speech tokens in LLMs, and face the challenge of relying on S2ST training data with ideally aligned speaker identity and prosody. We propose using low-bitrate tokens based on single-layer vector quantization, trained to reconstruct self-supervised learning (SSL) features. We also employ a separate token-to-waveform decoder named Autowave-X, which is also conditioned on the source speech to improve non-linguistic transfer, thereby relaxing the training data constraints. With the integration of these techniques, we propose an S2ST model named Kraken, which augments a pre-trained LLM with speech feature inputs and the low-bitrate token outputs, followed by Autowave-X vocoder. We built the model upon Qwen3-8B and trained it using 150k hours of multilingual and multitask speech data. We demonstrated that our model exhibited better translation quality than SeamlessM4T-Large v2 and Qwen2.5-Omni, along with improved speaker and prosody transfer capabilities.

Abstract: Large Language Model alignment typically relies on learned proxy reward models, which significantly increase the memory footprint during training and are notoriously prone to instability and reward hacking. While offline methods like Direct Preference Optimization (DPO) bypass the reward model, they lose the ability to perform online exploration. If no optimization constraints are applied, this can lead to format collapse in bounded, continuous spaces. To resolve this, we propose Direct Preference Density Alignment: An alternative framework that removes the need for a learned proxy reward model while strictly preserving the benefits of online reinforcement learning. We leverage large-scale user data (approximately 90,000 samples) to construct non-parametric preference density maps, establishing an empirical reward surface. In addition to removing the reward model, Direct Preference Density Alignment enables the combination of the online structural grounding of Group Relative Policy Optimization (GRPO) with the targeted offline refinement of DPO. We show that this GRPO+DPO combination achieves the highest performance, and in a blind audio equalization listening test, enables a 1.5B-parameter model to achieve perceptual parity with a carefully prompt-engineered GPT-4o mini baseline, using only a fraction of the inference compute.

Abstract: Automatic speech recognition (ASR) systems and audio language models (audio LMs) now report low error rates on monolingual benchmarks, but their behavior on code switched speech in low resource, diacritic rich languages remains poorly characterized. We present a switch aware evaluation of eleven modern systems (six ASR models and five audio LMs) on English Yoruba code-switched speech, using a deterministic 2000 utterance evaluation set and a shared scoring pipeline. Beyond word error rate (WER), we report switch localized diagnostics: a switch entry token error rate (SETER), windowed switch point error rates, language specific error rates, and a diacritic insensitive WER. Our central finding is that aggregate WER hides code switching behavior. The best system by WER (an ASR model) is statistically indistinguishable from a leading audio LM on WER, yet the audio LM is significantly better on every switch localized metric. Across faithful systems, Yoruba token recognition collapses (error 0.97 for almost all systems) while English tokens are recognized far better, and errors concentrate sharply at switches into Yoruba. Several generative audio LMs fail as exact transcribers, producing translation, verbosity, and prompt leakage that are strongly prompt dependent. We release manifests, metric implementations, and evaluation scripts to support reproducible, switch aware benchmarking for African code switched speech.

Abstract: We propose a preference optimization framework with Large Audio-Language Model (LALM) feedback for controllable non-verbal vocalization (NVV) generation in continuous autoregressive speech models. To construct preference data without human preference annotation, we build a bilingual prompt corpus by combining NVV-injected real transcripts with LLM-generated semantically aligned prompts, perform stochastic model rollouts, and use a LALM to rank candidate utterances and form same-prompt chosen--rejected pairs. We then adopt a two-stage optimization strategy: Rejection Sampling Fine-Tuning (RSFT) first adapts the model to LALM-selected high-scoring samples, followed by Anchored Flow-DPO, which formulates pairwise preference optimization using utterance-level flow-matching loss and retains the chosen-sample flow-matching objective as an SFT anchor. This design enables DPO-style preference learning without explicit sequence likelihoods while preserving direct supervision on preferred realizations. On the official 1,600-utterance NVVSpeech Challenge Track~2 test set, our method achieves a Final Track2Score of \textbf{75.80} (79.39 ZH / 72.21 EN), outperforming the VoxCPM2 baseline by \textbf{+1.84}. The improvements are mainly driven by higher NVV Accuracy and NVV Perceptual Effect, while Overall Quality remains stable.

Abstract: Robust speech-to-text translation systems should perform reliably across diverse acoustic conditions, yet practical pipelines lack controllable tools for systematic environment exploration. Large speech models remain sensitive to unseen acoustic conditions, as training data rarely cover the full range of real environments.We present NOPEHYPE, a structured training workflow that combines a controllable environment simulator, coverage-optimal environment reduction on Power Spectral Density (PSD) templates, and a small, interpretable hyperparameter search over simulator knobs. We show that simulator-generated noise achieves performance comparable to balanced realnoise training across Whisper and SeamlessM4T models, provide principled environment prototype sets, and identify practical default simulator configurations from a structured 27-run hyperparameter sweep.

Abstract: Large Audio Language Models (LALMs) have shown strong performance on audio reasoning benchmarks, but accuracy alone cannot distinguish true reasoning from superficial pattern matching, often overestimating reasoning ability since high scores may result from guessing rather than genuine audio understanding. Evaluating the reasoning process itself is essential for improving LALMs' reasoning ability, yet remains challenging. Existing methods either rely on costly human annotation or opaque LLM-as-judge approaches, making them impractical, biased, and lacking transparency. Moreover, audio reasoning introduces unique challenges absent in text-based settings, perceptual hallucination and cross-modal alignment between audio understanding and textual inference, hence text-based evaluation frameworks cannot be directly applied. Therefore, we propose ARIA-Rubrics (Audio Reasoning Integrity Assessment), a lightweight, annotation-free gold reasoning chains, automatic and transparent framework comprising six complementary metrics that evaluate audio reasoning quality across perceptual grounding, reasoning coherence, and answer consistency. We use Chain-of-Thought prompting as an externalization mechanism to make the reasoning process observable. Experiments on 9 models across 2 benchmarks identify three reasoning modes of current LALMs with actionable directions for future development, with ARIA-Rubrics achieving high correlation with human judgments. The code is available at the Github Repository.

Abstract: Large audio-language models (LALMs) can hallucinate audio objects, answering "yes" to absent sound events, thus undermining reliability in audio question answering. We propose Token-Adaptive Decoding (TAD), a training-free strategy for hallucination mitigation that grounds the initial yes/no decision by contrasting logits under real audio with a matched silent reference. TAD introduces a token-adaptive, confidence-guided gate that is decision-critical at the first decoding step and class-conditional on affirmative tokens, using the audio-silent margin to avoid overcorrection when evidence is weak or already sufficient. Experiments on AudioCaps-Hallucination show that, relative to Audio-Aware Decoding (AAD), a contrastive baseline with fixed contrast strength, TAD improves F1 for Qwen2 by 0.059 to 0.117 across Popular, Adversarial, and Random splits, and for Gemma by 0.025 to 0.064, while on Clotho-AQA it raises F1 from 0.810 to 0.816 on Qwen2 and remains comparable to AAD on Gemma.