Speech Large Language Models

We present a comprehensive benchmark of Federated Learning (FL) for multilingual Automatic Speech Recognition (ASR), evaluating four Speech-LLM architectures on the Multilingual LibriSpeech dataset. We compare FedAvg and FedProx across frozen and unfrozen encoder configurations, demonstrating that optimized learning rates are critical for performance. Specifically, independently tuning the learning rates for the speech encoder, connector, and decoder yields the lowest error rates, with full three-component adaptation (LoRA for encoder and decoder, full training for the connector) producing the best FL results. We observe that FedProx efficacy is architecture-dependent, providing notable advantages in multilingual pre-trained architectures (e.g., EuroLLM over TinyLlama when keeping the encoder fixed); this indicates that LLM backbone capacity plays a key role in mediating resilience to heterogeneous data distributions. These findings offer concrete design guidance for deploying multilingual Speech-LLMs in privacy-sensitive, distributed environments.

Non-verbal vocalizations (NVVs) carry important paralinguistic information but are often omitted by conventional automatic speech recognition (ASR) systems. The ISCSLP NVVSpeech Challenge requires joint transcription of lexical content and 16 NVV categories under limited and highly imbalanced supervision. We present a data-centric NVV-aware ASR pipeline based on cross-dataset label harmonization and a two-stage sampling schedule. We map heterogeneous source labels to the official taxonomy and exclude samples without a reliable mapping. Our schedule first uses square-root category sampling to moderate the long-tailed distribution and then applies uniform-category fine-tuning. On a fixed local validation split, square-root category sampling performs best among the tested single-stage settings. The final two-stage system obtains an official score of 63.86 and ranks fourth in Track 1.

SpeechLLMs have shown strong potential for emotion recognition, yet they read the predicted emotion off a generative decoder not suited for classification: it can emit labels outside the target set and favors frequent classes. We propose a discriminative adaptation that reads the final prompt token's hidden state through a classification head, producing a label in one forward pass without modifying the backbone. Because this readout starts from the hidden state the model would otherwise decode, it gives a controlled comparison of generative and discriminative inference in an otherwise identical speechLLM. We keep the head a single linear layer, trading little accuracy for interpretability: each emotion becomes one direction in the LLM output token space, revealing associated tokens. On IEMOCAP, across two speechLLM architectures, it improves Macro F1 and removes hallucinations, with largest gains on realistic ASR transcripts. Our analysis reveals that these emotion directions encode indirect associations mirroring biases in web-scale text.

High-quality 3D talking face datasets remain largely English- centric, and Korean 3D facial motion data are difficult to combine with standard English benchmarks because of differences in mesh topology, spatial scale, coordinate system, and temporal sampling. We present KoUniTalk, a lightweight articulation-centered Korean-English 3D talk- ing face benchmark that retargets VOCASET and the released Korean speech-based 3D talking face data to a shared mesh topology using de- formation transfer. Rather than proposing a new deformation-transfer algorithm or a full-head identity-preserving avatar dataset, KoUniTalk provides an identity-neutral canonical output space for controlled speech- driven facial articulation training and evaluation across English and Ko- rean. The unified template contains 1,176 vertices and focuses on the mouth and adjacent lower- and mid-face regions, reducing the output dimensionality from 15,069 and 72,147 dimensions to 3,528 dimensions, corresponding to 4.27-fold and 20.45-fold reductions compared with VO- CASET/FLAME and the original Korean mesh, respectively. To exam- ine whether retargeting preserves speech-relevant motion, we evaluate semantic mouth-landmark trajectories, including mouth opening, mouth width, aperture ratio, and mouth-opening dynamics. Since the official test set of the Korean dataset is not publicly released, we additionally define a subject-disjoint Korean benchmark split. The processed matched benchmark contains 22 speakers, 4,978 sequences, and 642,781 frames, enabling Korean-English cross-dataset evaluation of speech-driven 3D fa- cial animation models in a single compact articulation-template space. Source-reported inventory counts are listed separately from these pro- cessed counts

This paper proposes a phonemically comprehensive, ASCII-only romanization scheme for Thai and Lao, treating the two closely related languages as a unified cross-lingual design problem. The scheme represents segmental contrasts, vowel length, and lexical tone while maintaining one-symbol-one-phoneme transparency and systematic correspondence between Thai and Lao. The scheme prioritizes synchronic phonetic correspondence, including correspondence with Pinyin and Jyutping where applicable, while preserving historical-phonological correspondence where it does not conflict with phonetic transparency. Tone uses a compact single-digit default notation, supplemented by optional tone-value and historical tone-category representations. The resulting scheme provides a readable, keyboard-friendly, and machine-processable phonemic representation for language learning and cross-lingual speech processing.

Spontaneous speech offers a scalable, noninvasive signal for respiratory health assessment, yet interpretable models that generalize across languages remain challenging because disease-related acoustic changes are confounded by language-specific phonetic variation. We present CL-DAF, a Cross-Lingual Disease-Alignment Framework that identifies acoustic dimensions whose disease effects remain consistent across languages. Using 201 English and 75 newly collected Bangla speakers, we construct a common 272-dimensional acoustic representation and quantify disease alignment using signed rank-biserial effects and the Language Invariance Score. We first show that spontaneous Bangla speech separates COPD from controls (AUC 0.85); however, 133 features reverse their disease direction across languages and the full representation transfers poorly (AUC 0.49 from Bangla to English). CL-DAF isolates 26 disease-aligned features that raise AUCs to 0.825 and 0.722 from English to Bangla and Bangla to English, respectively. These findings provide a foundation for multilingual clinical speech models emphasizing pathology over language-dependent variation.

Contextual ASR can benefit from semantic cues or from target words explicitly provided in the context. We introduce HearInContext, a Mandarin-English benchmark that pairs shared synthetic speech with assistant replies supporting different interpretations. The benchmark comprises 3,764 semantic test cases built around homophones. Implicit contexts exclude candidate words; explicit contexts name the target. No-context and unrelated-context controls measure the benefit of relevant history and sensitivity to irrelevant history. Context-capable models benefit from implicit cues but achieve higher target recall with explicit hints. Fine-tuning Qwen3-ASR-1.7B improves implicit-context target recall by 11.4 percentage points in both Mandarin and English, while absolute CER/WER changes on AISHELL-1 and LibriSpeech remain below 0.1 percentage points. Gains extend to explicit conditions excluded from fine-tuning and to Mandarin hotword recognition on real recordings. Code and data are available at https://github.com/OPPO-Mente-Lab/HearInContext

Automatic speech recognition (ASR) systems exhibit unequal error rates across speaker groups, motivating interventions on their internal representations. We ask whether speaker-linked attributes that are linearly readable from pretrained ASR encoders yield useful directions for reducing group word-error-rate (WER) gaps. Across Whisper-medium, HuBERT-large, and Wav2Vec2-large on Common Voice and the Speech Accent Archive, we probe every encoder layer for metadata-derived sex/gender, age, and native/accent labels; construct centroid and probe-derived directions; inject them at selected layers; and compare downstream probe trajectories with matched WER changes. Sex labels are highly decodable (best macro-F1 0.924--0.941), native/accent labels are also above chance (0.544--0.696), and age is weaker (0.354--0.397). Of 22 post-selected reruns, nine have 95% paired-bootstrap intervals entirely below zero, yet every absolute source-group WER reduction is below 0.7 percentage points. Conversely, a local target-class probe rate can rise from 8.09% to 99.87% while WER worsens. Linear readability is therefore neither evidence of causal use nor a reliable mitigation method. Our results motivate evaluating speech-bias interventions jointly at representation, propagation, and task levels.

While Large Language Models excel in natural language processing, efficiently extending their capabilities to spoken input remains a significant challenge. Existing methods for building SpeechLLMs often rely on computationally expensive full-model fine-tuning, or employ parameter-efficient projectors that suffer from inefficient token sequence lengths and costly full-model supervision. In this paper, we introduce Aligned Continuous Integrate-and-Fire, a highly efficient framework for zero-shot speech processing. Our method dynamically compresses continuous acoustic frames into the exact discrete token length of the target text utilizing explicit Dynamic Time Warping alignments. This allows our initial training stage to establish a robust acoustic-to-semantic bridge using lightweight distance metrics, entirely bypassing the computationally expensive LLM forward pass. For subsequent fine-tuning, we propose a memory-efficient knowledge distillation objective that targets a single LLM layer, performing competitively with full-model cross-entropy training at a fraction of the computational cost. Through extensive evaluations on Automatic Speech Recognition and Speech Translation, we demonstrate that our method achieves superior performance compared to prior parameter-efficient baselines.

In Taiwanese Hokkien automatic speech recognition (ASR), prior studies often treat tone sandhi as a major challenge under the assumption that models fail to process implicit phonological variations. However, our experiments on Taiwanese Hokkien reveal that speech foundation models actually handle tone sandhi variations effectively, and the real performance bottleneck stems from a localized confusion between these variations and retained citation tones. To address this, we propose T-SANDHI to explicitly decouple surface acoustics from underlying lexical intent on top of a frozen Whisper backbone. Using a lexicon-guided multi-task learning structure driven by text-derived pseudo labels, our lightweight hybrid injection module integrates independent citation and sandhi phonetic streams via dynamic gating. Extensive evaluation on the TAT-MOE corpus and two blind test sets demonstrates that this explicit disentanglement effectively resolves tonal mapping confusion, outperforming baselines with strict parameter efficiency.

Speech Large Language Models (Speech-LLMs), typically built from a pre-trained speech encoder, a modality projector, and an LLM fine-tuned with Low-Rank Adapters (LoRA), have shown strong Automatic Speech Recognition (ASR) performance on general-domain speech. However, adapting them to domain-shifted speech, such as child or dialectal speech, remains challenging under limited target-domain data. Given the dominant role of the LLM in Speech-LLMs, with cross-entropy loss applied only at the LLM output, the speech encoder may receive insufficient adaptation to new acoustic conditions. In this paper, we propose Encoder Awakening via Adapters (EAVA), a simple yet effective domain-adaptive fine-tuning method for Speech-LLM-based ASR. First, lightweight adapters are inserted into each encoder layer and trained exclusively, enabling target-domain acoustic knowledge to be incorporated into the encoder while preserving its pre-trained knowledge. Second, the full model is jointly fine-tuned on the target domain with LoRA applied to the LLM. Experiments on three domain-shifted ASR datasets, covering child and dialectal speech, show that EAVA consistently outperforms vanilla fine-tuning and other baselines, achieving new state-of-the-art performance.

Audio deepfake detectors need to transfer to languages absent from training, as multilingual speech synthesis outpaces labeled anti-spoofing resources. While detectors increasingly rely on self-supervised speech models (S3Ms), these backbones encode language-dependent structure that confounds spoof cues. We address this confound through language orthogonalization, a target-free ridge map that removes S3M variation projected onto continuous language-identification (LID) embeddings. Across six languages, six S3M backbones, and all Leave-N-Out settings, it consistently reduces EER across unseen languages. Cross-lingual EER correlates with LID-space distance, where orthogonalization yields larger gains for more distant transfers.

Extending large-scale multilingual automatic speech recognition (ASR) models to low-resource languages remains challenging. Model performance is skewed toward high-resource languages and degrades sharply for languages with limited labeled data and pre-training exposure. To address this, we investigate parameter-efficient approaches for transferring knowledge from resource-rich source languages to low-resource target languages on Whisper. Alongside warm initialization and attention-based fusion, we propose Sequential Adapter Stacking, which places a trainable target-language adapter on top of a frozen source-language adapter. Under controlled experiments, these approaches are evaluated on three target languages unsupported by Whisper -- Asturian, Assamese, and Xhosa -- using source languages with varying degrees of relatedness. Sequential Adapter Stacking with the closest related source consistently and significantly outperforms full fine-tuning across the three targets, with 5--8% relative WER reductions. These gains largely persist with only one hour of target training data.

Automatic speech recognition (ASR) systems, trained on paired speech-text data, have been improved by leveraging language models (LMs) trained on text-only data. LM fusion methods such as shallow fusion and density ratio are well-established methods that incorporate external LMs during ASR decoding. However, they incur additional computational costs due to LM inference, which is particularly problematic for recent larger LMs. In this study, we propose incorporating external LMs via model merging. This method integrates the LMs directly into the parameters of an LLM-based ASR model, requiring no additional computational cost at inference. We formulate domain extension and transfer via arithmetic operations on LoRA parameters. Experimental evaluations were conducted for the domain adaptation of LLM-based ASR trained on CSJ and LibriSpeech. We show that our LM merging consistently improved the ASR performance in the target domains, without degrading inference speed or memory footprint.

Zero-shot text-to-speech (TTS) can clone a speaker's voice from a short audio prompt, yet most TTS systems still require the audio prompt transcript during inference. This dependency prevents cross-lingual voice cloning when the audio prompt transcript is unavailable, particularly for unseen languages. Cross-Lingual F5-TTS removes this dependency and enables transcript-free cross-lingual voice cloning, but it prepares its training data with forced alignment. Forced alignment is sensitive to boundary errors, and its cost grows as more languages are covered. Its speaking rate predictor is also unreliable at estimating duration when the audio prompt begins or ends with silence. In this paper, we present Cross-Lingual F5-TTS 2, a simplified framework for transcript-free cross-lingual voice cloning without forced alignment. Instead of using forced alignment to segment real utterances, we build same-speaker prompt and target pairs using a pretrained F5-TTS model and fine-tune the same model on these constructed pairs. This simplifies data preparation and preserves the acoustic modeling capability of the pretrained model, enabling adaptation with only a short fine-tuning stage. We further make the syllable-level speaking rate predictor robust to leading and trailing silence through silence-aware augmentation. Experiments show that Cross-Lingual F5-TTS 2 reaches higher speaker similarity than F5-TTS and Cross-Lingual F5-TTS while maintaining intelligibility. All related resources are publicly available.

False wake-up activations remain a persistent challenge in conversational AI. Speech phonetically similar to a device's wake word can produce a syntactically valid and semantically coherent ASR transcript that the assistant incorrectly executes. Most existing systems make a single intent decision in isolation, without a mechanism to learn from recurring errors over time or adapt to individual users through personalized learning. We introduce the Feedback-Driven Adaptive Self-Correcting Inference Layer (ASCIL), a complementary post-ASR correction framework that re-evaluates wake-up intent before response generation by fusing acoustic embeddings, linguistic cues, device context, and patterns from past misclassifications. ASCIL interprets implicit signals, including hesitation, disengagement, and silence, and explicit signals, including cancellation and repetition, as automatically inferred, noisy behavioral indicators of potential misclassification. These signals drive online pattern updates without manual annotation, whereas the intentional/unintentional reference labels used for offline evaluation are human-annotated. It generalizes from prior errors, applies corrective adjustments at inference time, and continuously updates in parallel with natural-language execution. Evaluated on a proprietary dataset of 3,667 interactions with human-annotated intentional/unintentional reference labels spanning 14 acoustic and contextual conditions, ASCIL achieves 54.27% relative error reduction on a session-disjoint subset constructed from baseline failures, and up to 24.39% relative error reduction at threshold 0.90 on the issue-tagged evaluation slice. These gains are achieved while improving intentional acceptance rates, with a median added latency below 60 ms in the reported benchmark.

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.

Checking Quran recitation from an ASR transcript requires distinguishing unresolved mistakes from repetitions, repairs, opening formulas and accepted spelling differences. We report a completed human annotation of 100 production recording cases: 348 scored units and 162 localized events across ten combined labels. An executable evaluator scores labels and word positions together. A plain diff reaches label-aware F1 0.525 and localization F1 0.826; adapted production cleaner/alignment components reach 0.518 and 0.786, with exact-span F1 0.505 for both. Correcting the adapter's word coordinates recovers all five annotated repetition events, showing why annotation interfaces must be checked before interpreting baseline failures. In a preliminary pilot, eight single 20-minute runs across three coding agents and eight models span label-aware F1 0.143 to 0.892: seven land far above every baseline, and one collapses below the naive diff from a missing normalization step. Across the six, 970 of 972 gold-event instances draw an overlapping prediction, so what remains is not detection but convention: span extent, and the labels whose boundary is stipulated by adjudication rather than visible in the text. Seven of 162 events defeat all six same-day runs, five of them one orthographic rule, and the strongest run still misses the same ones. No run annotated before building, so the pilot measures the algorithm half of the task only.

Speech Quality Assessment (SQA) is essential for modern speech technologies, and recent non-intrusive SQA predictors increasingly rely on Speech Foundation Models (SFMs). However, because SFMs expose representations from many layers, it remains unclear which depths are most informative for MOS prediction and how multi-layer information should be combined reliably across backbones and datasets. We benchmark ten SFMs on four MOS datasets under three regimes: full fine-tuning, last-layer probing with a frozen encoder, and naive cross-layer weighted aggregation. We find that the best layer is strongly backbone- and dataset-dependent, and that naive weighted fusion can be unstable across settings. We further evaluate a layer-calibrated aggregation variant that applies per-layer adapters before pooling, which improves the robustness of multi-layer fusion and narrows the gap to full fine-tuning while keeping the backbone frozen.

In Yorùbá, pitch alone separates \d{o}k\d{o} (husband, Mid), \d{o}k\d{ò} (vehicle, Low), and \d{o}k\d{ó} (hoe, High) -- the diacritics ARE the tone marks. Yet character error rate (CER), the standard automated metric for text-to-speech (TTS), is in practice computed from ASR output that drops those marks: a synthesizer can ace CER and still say vehicle for husband. We introduce DunDun -- named for the dùndún, the Yorùbá talking drum that speaks through pitch alone -- an automated, reference-free lexical-tone metric that needs no tone-labelled corpus. The gold High/Mid/Low sequence is read from the input text's diacritics (in TTS that text exists by construction, so no reference recording is needed); the prediction comes from the audio's pitch track. We validate three ways. Flattening pitch with PSOLA resynthesis collapses DunDun while CER does not move. Inverting High and Low in the answer key of 300 native recordings drives the two-class readout to 0.14, symmetrically below its 0.35 chance level -- a consistency check on the scoring path, not independent evidence. And three native listeners, over 67 blind A/B trials, pick the tone-correct clip 89.6% of the time (95% CI 80.0-94.8; p < 1e-4); whether DunDun tracks those judgements trial by trial is not resolved at this sample size. Applied to a massively multilingual zero-shot TTS model, DunDun shows what CER cannot: Yorùbá tone sits near the native anchor before any Yorùbá fine-tuning (0.567 +/- 0.02 over five decode seeds vs. 0.596; chance 0.33), despite the 21.4% CER the model's own paper reports; and a few hours of clean audio halve CER (5.6% to 2.7% by 5h, 1.7% by 15h) while tone saturates within the hour. On non-tonal Swahili, CER already captures the gains: the metric a language needs is language-dependent. We release the metric and the complete validation protocol.

Reducing audio-encoder depth lowers the inference cost of speech large language models, but removing complete blocks perturbs the embeddings consumed by the decoder and can cause deletion and premature end-of-sequence errors. We introduce X-AuT, a progressive framework that selects layer combinations through short behavioral probes and restores the pruned model through representation alignment, cross-scale distillation, scheduled student-policy supervision, and LoRA finetuning. The language-model backbone remains frozen, while attention LoRA adapters and the tied output embedding adapt during distillation. Training uses the highest-agreement tier from a transcript-consistency pipeline, followed by source reweighting during finetuning. On ten public Chinese--English benchmarks, compressing Qwen3-ASR-0.6B from 18 to 16 audio-encoder layers reduces macro-average error from 5.61% to 5.27%. The 14-layer model reaches 5.75% with 20.7% fewer audio-tower parameters. Under the matched recipe, the 1.7B teacher yields 5.55% mean error, compared with 8.45% for self-distillation, and progressive 18$\rightarrow$14 pruning outperforms direct pruning (5.75% vs. 6.73%). These single-run results establish two practical operating points and show that the accuracy effects vary across benchmarks. Project website: https://xpeng-ai.github.io/x-aut

Speech large language models (SpeechLLMs) offer reduced latency and retain paralinguistic nuances that are typically lost in cascaded automatic speech recognition (ASR) and text-based LM architectures. However, they continue to lag behind text-only LLMs on complex reasoning tasks, while real-time spoken interaction imposes strict latency constraints. Although prior works employ Chain-of-Thought (CoT) and concurrent reasoning to enhance reasoning capabilities without inducing prohibitive delays, an inherent accuracy-latency trade-off persists. In this paper, we investigate whether a streaming SpeechLLM can dynamically revise its reasoning traces on the fly. We introduce RetroThinker, a multi-stage post-training framework that equips the Moshi model to self-verify and forward-correct CoT steps during inference. RetroThinker combines supervised fine-tuning (SFT) on curated retrospective thinking data with length-based direct preference optimization (DPO) to optimize retrospective during early reasoning (i.e., reasoning concurrently while the user speaks). Evaluated on the GSM8K benchmark, RetroThinker significantly improves the accuracy-latency trade-off over non-retrospective baselines, achieving an 11% absolute accuracy gain at a comparable latency.

Kurdish is spoken by millions of people, but little technology can read it aloud. A recent study released three Kurdish voices, 35 hours of recorded speech, and a paper describing the work, all free to download. This review checks how well those public files match the paper. The research is careful about its limits, but the files contain several problems: a settings file lists equipment that was never used, test recordings are left unlabeled among training data, and a coding fault mishandles long numbers. The download page also claims a stronger result than the paper reports and recommends one voice for general use. That recommendation matters because Kurdish has major regional and written variation, while these voices were built from three people reading prepared texts. The process therefore removes much everyday and regional speech. English and German benefit from long traditions of dictionaries and linguistic description that help identify wrong pronunciations; Kurdish has far less such support, so software choices can go unchecked. The voices sound fluent, but they represent the reading styles of their speakers rather than Kurdish as a whole. Most of these issues can be fixed using information the team already has, without changing the reported results. Better records would mainly make the work easier for others, especially community linguists, to check and reuse. The license is the main exception: whether audiobook owners allow corrected versions to be shared will affect whether future Kurdish voices can build on this work or must start again.

This paper describes our system for Task2 of the second Multilingual Conversational Speech Language Model (MLC-SLM) Challenge. We adapt Qwen3-Omni-30B-A3B-Instruct with a segment-evidence-aware data and post-training pipeline. A language model converts timestamped ASR into coherent event spans, which are expanded by a boundary margin and cropped from the original recording. We then synthesize complementary semantic MCQs with Qwen3.6-27B and acoustic MCQs with Gemini3.1 Flash-Lite, followed by structural, grounding, answer-consistency, and target-model trainability checks, yielding 359,825 verified MCQs across 21 language and accent variants. A text-only probe partitions the data into weak, text-answerable items used for supervised fine-tuning and strong, audio-dependent items used for reinforcement learning with Group Sequence Policy Optimization (GSPO), stabilized by debiased advantages, sequence-level importance correction, and dynamic filtering. Our system obtains 90.92% accuracy on the final official evaluation set.

As Speech Large Language Models (speech-LLMs) become increasingly multilingual, Arabic remains significantly underrepresented, highlighting the need for dedicated infrastructure to train and evaluate Arabic speech-LLMs. To address this gap, we introduce Nuha-Speech, a comprehensive initiative to develop general-purpose Arabic speech-LLMs spanning dataset construction, model training, and systematic evaluation. Specifically, we constructed a large-scale Arabic Speech Question-Answering (SQA) corpus comprising over 1.5 million training samples to allow instruction tuning over a broad range of core speech tasks. Then, the corpus was used for supervised fine-tuning based on Qwen-Omni model variants at different scales. Finally, we designed an evaluation framework featuring diverse tasks and tailored metrics. Through this work, we aim to establish foundational infrastructures for Arabic Speech-LLMs under constraints imposed by limited Arabic speech resources.

Low-resource multilingual text-to-speech (TTS) systems have expanded language coverage, but their robustness under complex text inputs remains insufficiently diagnosed. Existing evaluations mainly focus on naturalness, speaker similarity, and content consistency using regular test sentences, while providing limited insight into how multilingual TTS systems fail when handling challenging inputs such as numbers, dates, named entities, long sentences, code-switched expressions, and punctuation-related structures. This paper proposes a complex-text robustness diagnosis framework for low-resource multilingual TTS. We evaluate robustness from three dimensions: content consistency, language consistency, and generation stability. A multilingual robustness testing scheme is designed for Thai, Vietnamese, Swahili, and Indonesian, covering ordinary sentences and multiple types of complex text inputs. We further introduce automatic diagnostic metrics, including character error rate, language identification accuracy, and duration abnormal rate. To support input-level risk analysis before speech generation, we propose a lightweight Text Risk Score (TRS), which estimates synthesis risk from interpretable text features without manual annotation or model training. Experiments on three representative multilingual TTS systems, including OmniVoice, VoxCPM2, and MMS-TTS, show that complex text inputs expose systematic failure patterns that are not fully reflected by ordinary short-sentence evaluation. Different systems exhibit distinct vulnerabilities in number normalization, named entity handling, long-text generation, and code-switched input processing. Furthermore, TRS shows a positive correlation with content errors and duration abnormalities, demonstrating its usefulness as a low-cost pre-synthesis indicator for complex-text risk diagnosis in low-resource multilingual TTS.

Per-layer differential privacy (DP) clipping improves gradient fidelity in federated learning by allocating per-matrix clipping budgets proportional to parameter count. We show that this recipe breaks for speech large language models (speech-LLMs), when the acoustic encoder and the language decoder differ by an order of magnitude in update norm. Single-pool per-layer methods suffer \emph{cross-component budget collapse}, dragging word error rate (WER) far from flat global clipping or collapsing training entirely. When the norm imbalance is milder, adaptive single-pool methods partially recover, confirming that collapse severity scales with the inter-component norm ratio. We empirically diagnose the root cause across six per-layer methods and three speech-LLM architectures. We then propose \emph{$\alpha$-split}, a two-pool allocation that normalises encoder and LLM parameters into independent pools, and show that joint $\ell_2$ sensitivity and the original $(\varepsilon,\delta)$-DP guarantee are unchanged. At architecture-calibrated $\alpha$, our method recovers WER utility compared to flat DP, while granting the encoder $4.47{\times}$ tighter per-component noise protection against speaker voice-based gradient-inversion attacks at only $+2.6\%$ LLM noise overhead.

This paper presents an end-to-end evaluation framework for image-triggered command injection against computer-use agents (CUAs). The goal is to test whether a local visual patch can induce verifiable environmental consequences along the full chain of screenshot input, VLM generation, action parsing, and environment execution. We train and deploy patches on author-controlled GitHub Pages pages and a locally deployed CSDN clone, and evaluate them in real environments across five open-source or publicly available GUI-agent or vision-language-model (VLM) backends. Our experiment aggregates 600 instance-level online cases, with T-ASR, TAPR, and E2E-ASR reaching 84.5%, 47.0%, and 20.3%, respectively. Trajectory analysis further shows that in some successful cases the agent first executes a malicious terminal command and then continues the original benign task. These results indicate that optimized local visual signals can affect not only VLM outputs but also propagate through the execution pipeline of open CUAs and create real environmental risk.

Self-supervised speech foundation models are now used in a wide array of downstream applications, including traditional speech recognition and as the basis for tokens in speech-aware language models. Attempts to understand their usefulness have largely focused on probing their representations' ability to discriminate phonemes and words. However, discriminative ability for words need not imply specialized representation of words per se. Good discrimination of words may be explained by good encoding of word form (phonemes) rather than form-independent word representations encoding identity or syntactic/semantic properties. By partialling out phoneme information using residualization, we show that, in later layers, HuBERT and wav2vec 2.0 do in general learn representations which encode words with reasonable fidelity independently of local phonetic content. We show that this simple approach to disentanglement can enhance higher-order linguistic information in word discovery tasks.

Orukeet replaces half of an adapted Parakeet encoder's temporal filters with 12,288 fitted Gabor kernels, freezes these replacements, and trains the remaining parameters on multilingual and multi-accent data. Final adaptation and checkpoint selection use LibriSpeech test-other. Across 20,146 FLEURS recordings in 25 languages, pooled word error rate (WER) falls from Parakeet's 11.01% to Orukeet's 9.85%, a 10.6% relative reduction. Orukeet has lower WER on 23 of the 25 languages. Orukeet outperforms Parakeet on 61 out of 74 tested splits, including LibriSpeech test-clean (1.46% vs. 1.53% WER), test-other (2.86% vs. 3.14%), and FLEURS English (3.82% vs. 4.28%). All comparisons decode the same audio with matched NeMo settings. The fitted kernels are stored as ordinary convolution weights, retaining Parakeet's architecture and inference operators.

Text-aligned speech tokenization methods have emerged to better align speech tokens with LLM token spaces, enabling more effective utilization of pretrained LLMs. However, they rely on offline automatic speech recognition (ASR), leading to two key limitations: (i) the need for complete utterances before tokenization, precluding real-time streaming, and (ii) vocabulary mismatch between ASR and LLMs, which reduces acoustic granularity from the subword to the word level. We introduce StreamAlign, a text-aligned speech tokenization framework that enables streaming tokenization for real-time speech-text joint modeling. StreamAlign performs online speech-text alignment by combining character-level RNN-Transducer alignment with word-level ASR guidance, mitigating ASR-LLM vocabulary mismatch while preserving recognition accuracy. A proactive word boundary classifier anticipates word completion at chunk boundaries, reducing tokenization latency from 560 ms to 270 ms. On LibriSpeech, StreamAlign achieves the lowest WER and highest UTMOS among evaluated tokenizers. Furthermore, StreamAlign-SLM, a spoken language model trained on StreamAlign units, outperforms other end-to-end spoken language models in speech continuation while achieving the strongest overall consistency on SALMon and spoken StoryCloze.

The rapid advancement of audio and multimodal large language models has unlocked transformative speech understanding capabilities, yet evaluation frameworks remain predominantly English-centric, leaving Southeast Asian (SEA) languages critically underrepresented. We introduce SEA-SpeechBench, to the best of our knowledge, the first large-scale multitask benchmark that evaluates speech understanding in 11 SEA languages through 97,194 samples across 99 evaluation sets and 597 hours of curated audio data. Our benchmark comprises 9 diverse tasks across 3 categories: speech processing (automatic speech recognition, speech translation, spoken question answering), paralinguistic analysis (emotion, gender, age, speaker recognition), and temporal understanding, a novel dimension featuring timestamped content queries and temporal localization within extended audio sequences up to 3 minutes. We implement multilingual prompting in both native SEA languages and English to reflect user interactions with audio-language models. Evaluation of leading open-source and proprietary systems reveals marked performance gaps. Across all models, performance remains underwhelming on temporal understanding, emotion recognition, and speech translation. Prompting in low-resource languages such as Burmese and Tamil lags behind English by up to 41 percentage points. Our findings expose critical model limitations and underscore the need for inclusive model development. The SEA-SpeechBench benchmark is available at https://zwenyu.github.io/SEA-SpeechBench/.

We introduce BuzzASR, a collection of language-specialized fine-tuned Whisper models adapted for automatic speech recognition (ASR) in 102 languages. Large end-to-end Transformer-based ASR models such as Whisper have revolutionized ASR, but most prominent models are highly multilingual. As a result, these models often perform poorly on languages less well-represented in their training set. While it has long been known that effective language adaptation can be achieved through simple fine-tuning on monolingual data, this strategy has only been applied to a small number of languages. We massively scale up this simple approach to 102 languages covered in the FLEURS dataset, while also implementing a more complex language adaptation strategy that integrates monolingual tokenizer replacement and data augmentation using text-only fine-tuning. BuzzASR models outperform Whisper-large-v3 on 77 out of 102 languages, reducing character error rates (CER) by a factor of over 2.8 on average. Our models achieve state-of-the-art CER among open-source systems on 27 of 102 languages on the combined FLEURS and Common Voice test set. Our tokenizer replacement strategy yields an average 3.3x improvement in compression rate (characters per token) over Whisper's multilingual BPE, with gains of up to 21.7x. We release all models, code, and detailed results: https://lemn-lab.github.io/buzz-asr

Whisper is a widely used foundation model for automatic speech recognition (ASR), but its generative decoder can produce fluent hallucinated transcripts for inputs containing little or no speech. We propose a training-free, inference-time method to reduce these hallucinations using low-rank projection of decoder activations. A compact hallucination-associated subspace is estimated from non-speech calibration data, and decoder hidden states are projected away from this subspace during inference. We evaluate two variants: always-on, which applies projection to all inputs, and gated, which applies it only when Whisper predicts that an input is likely non-speech. Across non-speech benchmarks, always-on projection reduces average hallucination rate (HR) from 31.31% to 2.44%, a 92.21% relative reduction, while gated projection reduces HR to 3.74%, an 88.05% relative reduction, with lower false rejection of genuine speech. On LibriSpeech, gated projection increases absolute word error rate (WER) by 0.33-4.39 percentage points and yields false-rejection rates (FRR) of 0.41--9.97% across model and split settings. These results show that low-rank activation projection can substantially suppress Whisper hallucinations without retraining, while providing a controllable trade-off between hallucination suppression and speech recognition performance.

The ambition of the 2025 PNPL competition (Landau et al., 2025) was to launch a multi-year curriculum for non-invasive speech decoding. Designed to progress from foundational tasks toward the linguistic complexity required for a practical brain-computer interface (BCI), it set the stage with speech detection and phoneme classification tasks. Winning submissions reached F1-macro scores of 95.6% and 73.6% on the respective tasks (Elvers et al., 2026), highly significant advances. This success was built on the LibriBrain dataset (Özdogan et al., 2025), the largest within-subject MEG dataset recorded at the time with ${\sim}50$ hours of data for one subject. However, while within-subject scale drives strong decoding performance, a practical BCI must generalise to new users from minutes of data, not hours. The 2026 PNPL competition responds to this challenge with LibriBrain100 (Mantegna et al., 2026), an extended LibriBrain dataset with 32 additional subjects (${\sim}40$ minutes each) plus even more within-subject data (${\sim}80$ hours). Advancing the curriculum of tasks to focus on word classification, two complementary tracks are presented in this competition: the Deep track targets within-subject word classification at scale, aiming at the best possible performance; the Broad track targets cross-subject generalisation, progressively reducing the amount of subject-specific fine-tuning data from ${\sim}40$ to ${\sim}20$ to ${\sim}10$ minutes, a duration that falls within a clinically feasible range and brings us a step closer to a non-invasive BCI capable of restoring communication to people living with profound paralysis.

Multi-talker automatic speech recognition (MT-ASR) remains challenging in the presence of overlapping speech. Hard segmentation introduces irreversible errors, whereas serialized output training (SOT) avoids explicit segmentation but does not condition a pretrained encoder on speaker activity. We propose Soft Posterior Speaker Injection (SPSI). A Soft Posterior Head predicts per-frame speaker posteriors $\hat{\mathbf{P}}$ and injects them into Whisper through Multi-layer Feature-wise Linear Modulation (MFLM) and Speaker Memory Prompts (SMP). The benefit of SPSI is largest where overlap is heaviest and under domain transfer. On controlled two-speaker LibriSpeech overlap, SPSI reduces concatenated minimum-permutation word error rate (cpWER) from $61.5\%$ to $60.0\%$ in the high-overlap bin, and from $51.9\%$ to $51.0\%$ on the full set, relative to SOT. By contrast, Speaker CE, SD-CTC, SA-DiCoW, and Pipeline (oracle/est.\ VAD) do not outperform SOT. Freeze-posterior overlap-heavy adaptation reduces held-out LibriCSS cpWER from $42.3\%$ to $36.8\%$ on sessions $8$--$9$, a $5.5$-point gain over SOT. The source code is available at https://github.com/HackerHyper/SPSI.

Current speech synthesis struggles with code-switching, which mixes a foreign language phrase into a primary language utterance, causing the phrase to be spoken with the primary language's accent rather than its native one. We propose Phrase-Localized Language-Contrastive Guidance (LCG), a training-free inference framework that restores a native accent to code-switched phrases in cross-lingual text-to-speech. LCG replaces the single language guidance applied across the whole utterance with a separate guidance for each region, so each part is guided by its own language. To choose where to apply this localized guidance, we propose a self-attention probing technique that finds the phrase boundaries without external alignments. Together, these components generate speech in which each region carries the accent of its own language, requiring no fine-tuning or auxiliary models. Across diverse language pairs, LCG robustly increases the nativeness of the code-switched phrase while suppressing accent leakage, and preserving overall speaker identity and naturalness.

Scene text recognition is reported as 89--97% accurate on the six standard benchmarks, and the problem is widely treated as saturated. We present an alternative reading. When the same test images are stratified jointly by ground-truth word rarity and character n-gram novelty against a reference corpus, accuracy at the rare-word x rare-trigram corner of the resulting 5x5 grid drops 10--18 pt below the q3/q3 centre across nine English specialised recognisers, and the same direction (corner below centre) holds on all 13 of 13 (language, model) pairs we test across four writing systems (Latin, Han, Han+kana, Arabic). The drop is not a capacity bottleneck. A 6x vision-backbone scale-up (CLIP4STR-Base 158M -> CLIP4STR-Huge 1.0B, OpenCLIP ViT-H/14 LAION-2B) leads every benchmark in aggregate accuracy yet leaves the stress corner unchanged (86.9 -> 86.5, within paired-bootstrap noise). Four converging probes--layer-wise probing, confidence-when-wrong, attention re-balancing, and a cross-script commit-vs-abstain error split--localise the failure to the autoregressive decoder's lexical prior. We then ask how much of the gap existing techniques recover. Of 16 non-architectural mitigations, the largest mean q5/q5 gain is +1.3 pt and none clears the paired-bootstrap noise floor; the only intervention that does is the architectural shift from autoregressive to CTC decoding (SVTRv2, +2.5 pt, p=0.02, n=474). A confidence-routed AR-CTC ensemble adds a directionally consistent +0.6 pt that stays within noise, and its dominant learned coefficient is each model's own minimum-softmax confidence--independently echoing the mechanism above. No configuration we test improves both the compositional corner and aggregate accuracy. The rare-input long tail thus points to architectural change rather than added capacity.

Retrieval-Augmented Generation (RAG) grounds large language models in external corpora, but implicit trust in retrieved documents creates a critical attack surface: PoisonedRAG shows that a handful of crafted passages can dominate dense retrieval and steer generation toward attacker-chosen answers. We present the Tri-Layer Sieve, a middleware defense that sanitizes retrieved evidence through cross-embedding-space clustering with an independent judge model, structural filtering of trigger-payload artifacts, and LLM consistency verification. The design exploits a key weakness of retrieval-stage poisoning: a single document must satisfy one embedding geometry, one internal Trigger-Payload structure, and one generation objective - rarely all three simultaneously, a fragility that persists even against an adaptive attacker who paraphrases around it. On Natural Questions, HotpotQA, and MS-MARCO with Contriever retrieval (k=50), the Sieve reduces black-box Attack Success Rate from 67.0/87.0/64.0% to 3.0/14.0/4.0%, mitigates white-box HotFlip attacks from ~74% to 27.8% on NQ with Layer 3 enabled, and drives poisoned-document MRR to 0.000, while restoring clean accuracy from 13-33% under attack to 58-76%. Under an architecture-aware adversary who paraphrases triggers to evade the structural filter, enabling the consistency layer halves adaptive ASR (32.0% to 15.0% on NQ) while raising clean accuracy by 18 points, at an added latency of ~16-19 s/query under live retrieval.

Speech-based Alzheimer's disease (AD) detection increasingly relies on speech-enhanced and curated versions of the Pitt Corpus, where speech enhancement, sample selection, and demographic balancing are often treated as beneficial preprocessing steps. However, whether these transformations improve real-world AD detection or instead affect model generalization and prediction behavior remains unclear. In this work, we revisit the role of speech preprocessing and dataset curation across widely used benchmarks for speech-based AD detection. We evaluate the speech quality of different datasets, the cross-dataset generalization of multiple deep learning models under matched and mismatched enhancement settings, and the behavior of several recent large audio-language models (LALMs). Experimental results show that across multiple supervised speech models, speech-enhanced datasets often improve in-domain performance while reducing robustness in cross-domain evaluation. Matched enhancement between training and test data alleviates, but does not eliminate, this degradation. LALMs show a similar sensitivity: enhanced datasets induce stronger class imbalance and prediction shifts than unprocessed data. These results suggest that speech preprocessing and dataset curation can substantially influence downstream AD detection behavior, indicating that ``cleaner'' speech datasets are not necessarily more reliable for real-world AD detection.

While automatic speech recognition (ASR) models have achieved remarkable improvements in recent years, performance disparities persist across different speaker populations. One such disparity is for speakers whose first languages (L1) are from families distant from English. This paper investigates the relationship between first language background and English ASR performance. Through empirical analysis, we observe that the correlation between speakers' L1 distance and ASR error rates yields a systematic effect on English Speech, with its strength varying across datasets and models. This association is statistically significant in a follow-up analysis accounting for dataset-level variation in Tweedie mixed-effects models ($p<0.001$ across evaluated models). In addition, analysis of the latent space reveals a L1-based spatial segregation across deeper acoustic layers in the majority of evaluated architectures

Obtaining data from neuromorphic sensors and processing it with Spiking Neural Networks is a promising solution to lower the energy cost of artificial intelligence. The current rarity of natively neuromorphic datasets promotes the development of software tools to translate input sensory data into spikes. However, highly bio-mimetic simulators can be challenging to implement on digital hardware. In this work, we evaluate the neuromorphic encoding and subsequent classification of audio into spikes using a non-learnable, high-level, programmable encoder targeting hardware implementation on FPGA. We quantify the pipeline's efficiency with hardware-agnostic metrics based on the quantitative spiking activity. Our study focuses on the simultaneous optimisation of encoder and classifier: the first provides efficient and informative data so that the latter achieves a better performance with an overall lower energy cost at learning and inference. This work introduces the first end-to-end neuromorphic spike-encoding and evaluation of the TIMIT dataset. Our simple feedforward network reaches a classification accuracy of 99.77% on a spike-encoded Heidelberg Digits, overcoming the neuromorphic state of the art on this benchmark dataset.

Speech enhancement (SE) is commonly applied as a preprocessing step in spoken AI pipelines under the assumption that better audio quality improves downstream task performance. Whether SE-induced distortions propagate to downstream LLM task performance remains an open question. We introduce Output Divergence Rate (ODR), which measures how often SE changes an LLM's intent classification relative to clean speech, and benchmark five conditions on 2,974 SLURP clips using Whisper large-v3 and wav2vec2-large cascades. Every condition produces ODR significantly above zero ($p < 0.001$, binomial test). MetricGAN{+} more than doubles ODR versus unenhanced noisy speech (0.318 vs. 0.135) despite improving PESQ, and unmitigated echo reaches an ODR of 0.836 through speaker substitution, a failure WER cannot capture. Audio quality metrics range from near-zero to moderate correlation with ODR (SQUIM-MOS $ρ=-0.068$, PESQ $ρ=-0.467$). The MetricGAN{+} and echo results replicate across ASR architectures, indicating that standard audio quality metrics are insufficient for LLM pipeline quality.

Speech enhancement is often used as a front-end for robust ASR, yet recurrent temporal and cross-band modules introduce sequential dependencies that reduce parallel efficiency. In this paper, we present a sequence-parallel band-split enhancement front-end built on a Parallel Time-Band Mixer (PTBM) block that eliminates within-block recurrent unrolling. PTBM integrates intra-band temporal mixing and per-frame cross-band attention within a unified parallel architecture, enabling efficient contextual modeling across both time and frequency dimensions. The system retains the mask-plus-residual reconstruction interface and introduces learned Observation-Adding (LOA) to suppress ASR-sensitive artifacts without development-set tuning. Experiments on DNS Challenge and CHiME-4 with frozen Whisper back-ends show that the proposed front-end consistently reduces word error rate relative to recurrent band-split baselines while requiring only 0.96 M parameters and 0.58 GMAC/s for the front-end network.

Speech language models (speech LLMs) can generate plausible outputs from audio that contains no usable speech evidence. We study this failure as a pre-generation support-estimation problem and present SURE-Voice, a training-free front end that decides whether an audio prompt contains intelligible speech evidence before calling a speech LLM. We build SURE-Challenge with a 640-example SURE-Core split and a 1,920-example SURE-Extended split derived from 120 LibriSpeech source utterances. Using one fixed operating point, an energy screen plus Whisper token confidence raises unsupported accuracy on the held-out Extended test from 0.000--0.133 to 0.919 for six non-degenerate speech LLM backbones, while supported accuracy remains 0.919--0.970 and downstream calls fall from 480 to 287. A 500-clip ESC-50 sanity set shows the same pattern on real environmental audio, with vocal non-speech as a residual failure mode. An overlap diagnostic shows that source attribution remains separate from speech-evidence filtering. The evidence supports a controlled benchmark baseline and a deployment-oriented analysis; it does not establish universal robustness to semantic answerability, gain variation or natural conversations.

Research on automatic speaking assessment (ASA) has increasingly adopted multimodal speech large language models to assess learners' speaking performance. However, existing studies provide limited analysis of how acoustic and content information contribute to predictions and how stable the resulting performance is. We propose CASA, a simpler architecture combining Whisper-medium and Qwen3.5-2B that achieves state-of-the-art performance while providing a more interpretable separation between speech delivery and content. On the Speak & Improve Corpus 2025, CASA achieves a root mean square error (RMSE) of 0.358, improving on the previous best RMSE while using approximately half the estimated inference parameters. The general-purpose architecture is designed for adaptation to other ASA corpora without structural changes and relies on three handcrafted fluency features. Through ablations and repeated runs, we analyze the individual and complementary contributions of acoustic and content information, examine performance variability, and demonstrate the potential of large language model reasoning for training-free content validation.

Reliable relative pose estimation between autonomous underwater vehicles (AUVs) is critical for cooperative ocean exploration, sampling, and multi-robot coordination. However, achieving robust vision-based relative localization in underwater environments remains challenging due to severe optical degradation, including turbidity, illumination variations, reflections, and intermittent feature occlusions. This paper presents AMR-Pose, an active LED marker-based relative pose estimation framework for cooperative AUVs. A compact marker module consisting of one red central LED and three blue peripheral LEDs is developed and integrated onto the leader AUV to provide distinctive visual features under complex underwater conditions. Building upon the detected marker observations, a probabilistic switching Perspective-n-Point estimator (PSwPnP) is developed by combining Lie-group pose propagation on $SE(3)$, probabilistic marker association, and visibility-adaptive measurement fusion for robust six-degree-of-freedom relative pose estimation. The proposed framework dynamically adapts the estimation process according to marker visibility, maintaining geometric consistency and temporal stability during partial observations and visibility transitions. Extensive water-tank experiments with motion-capture ground truth validate that AMR-Pose achieves accurate, smooth, and robust relative pose estimation under challenging underwater conditions. Closed-loop leader-follower experiments further demonstrate its feasibility for real-time relative pose feedback in cooperative underwater robotics.

Recent advances in large language models (LLMs) and multimodal LLMs (MLLMs) have created new opportunities for wearable speech interfaces, with smart glasses providing an egocentric platform for continuous audio sensing and assistance. However, speech recognition and understanding in this setting remain challenging because of dynamic acoustic conditions, speaker overlap, and the spatial ambiguity introduced by wearer-centered recording geometry. To support systematic evaluation in this setting, we introduce the IEEE SLT 2026 SmartGlasses Challenge for egocentric multi-speaker speech processing. The challenge consists of two tracks, Dyadic Dialogue Understanding and Multi-party Meeting Understanding, and jointly evaluates Time-Stamped Speaker-Attributed Automatic Speech Recognition (TSA-ASR) and Spoken Language Understanding (SLU). It is built on a 106-hour four-channel egocentric speech dataset containing 714 sessions collected in real-world scenarios. This paper describes challenge tasks, dataset construction, submissions, and summarizes the main findings from the shared evaluation. The results show that heavy speaker overlap remains a major factor affecting TSA-ASR performance, while paralinguistic acoustic understanding continues to be difficult for current audio-language models in complex SLU settings. Further details can be found on the official challenge website.

Recent advances in zero-shot text-to-speech (TTS) have substantially improved speech quality and voice cloning fidelity. However, many zero-shot TTS systems still depend on audio prompt transcripts at inference time. This dependency limits cross-lingual voice cloning, since in-the-wild reference audio is often untranscribed. In this technical report, we present Confucius4-TTS, a multilingual zero-shot TTS system that supports 14 languages and performs both intra-lingual and cross-lingual reference cloning without requiring transcripts of audio prompts. Confucius4-TTS follows a two-stage architecture, consisting of text-to-semantic (T2S) and semantic-to-acoustic (S2A) modules. The LLM-based T2S module uses a learnable speaker encoder to extract timbre features from self-supervised speech representations, and the conditional flow-matching S2A module converts the predicted semantic tokens into mel-spectrograms. The same model also supports continuation cloning when a reference transcript is available. Confucius4-TTS is trained on large-scale multilingual speech data. It achieves high intelligibility and speaker similarity on public benchmarks. On the CV3-Eval cross-lingual benchmark, Confucius4-TTS obtains an average WER of 3.73% across six directions. On our internal cross-lingual set, it achieves the best average overall rank in human evaluation among recent open-source and commercial systems. We release code, model checkpoints, and demos at https://github.com/netease-youdao/Confucius4-TTS.

Audio transcription is a critical bottleneck in language documentation. While multilingual Automatic Speech Recognition (ASR) models like Whisper offer solutions, field linguists often lack the expertise to utilise them. We present Easper, an open-source, no-code workflow enabling linguists to iteratively fine-tune ASR models via cloud resources directly from ELAN annotations. Deploying ASR also raises a cold start problem: deciding which recordings to transcribe first to bootstrap an accurate model. Using Easper, we evaluate transcription prioritisation strategies on three Vanuatu languages (Bislama, Nafsan, Nguna). We fine-tune models by recording session, comparing Character Error Rate trajectories when prioritising acoustic cleanliness versus linguistic richness. We demonstrate that prioritising lexically rich narratives and increasing acoustic-phonetic repetition, even in noisy environments, leads to faster improvements in transcription quality.

Recent advances in model design and self-supervised audio representations have improved speech and audio understanding, yet infant-centered naturalistic recordings remain challenging due to limited labeled data, low signal-to-noise ratio, and cross-family domain shifts. We present a family-conditioned, multi-tier audio tagger that combines a LoRA-finetuned Whisper encoder with a lightweight, target-speaker-aware Transformer for long-context inference and framewise prediction across tiers. To improve temporal coherence, we incorporate a simple sequence-level smoothing loss, and to enhance robustness across households, we introduce a factorized speaker-token design with a shared tier token and a learned family-specific offset, reducing family bias and promoting generalizable representations. Together, these choices enable efficient and effective infant-centered audio tagging of daylong audio recordings in home environments.

Low-resource automatic speech recognition (ASR) commonly relies on cross-lingual transfer, where models are adapted from higher-resource donor languages. However, selecting donors remains challenging for spontaneous speech from under-resourced language communities, due to linguistic variation, evolving orthographic conventions, and uneven resource availability. We present DonorRank, a learning-to-rank framework for predicting effective donor languages for zero-shot ASR. We evaluate DonorRank on two multilingual speech corpora of Indic and African language families. It accurately predicts donor language rankings and improves donor selection over common heuristics based on genetic similarity or high-resource languages. Beyond improving transfer, we show how DonorRank is a general framework for analyzing donor language selection itself. Our analyses show that the composition of the donor set determines which linguistic cues are useful in predicting successful transfer. We also identify transfer patterns that provide practical guidance for multilingual ASR in low-resource settings.

A common approach to adding audio to a vision-language model is to train or adapt a large omni-modal system. We show that a lightweight alternative can be highly effective for music audio-visual question answering (AVQA). Qwen-MusicAVQA-7B connects a frozen Whisper encoder to Qwen2-VL-7B-Instruct through learned linear projections. The same frozen encoder processes both the video's music track and a TTS-spoken question through separate projectors, while the language model fuses visual frames, music, and question audio through pretrained self-attention, with no task-specific fusion network. On MUSIC-AVQA, our system reaches 96.0% +/- 3.9% accuracy across three independent training seeds on the 7,402-question available-video test subset. Our central finding is that downstream accuracy tracks how much fine-grained local temporal information the audio representation preserves. In a matched 32-token comparison, a stride-pooled Whisper frame sequence outperforms a globally pooled PANNs representation expanded to the same budget by 26 percentage points, even though PANNs sees at least as much audio and uses a far larger projector. The effect is not simply sequence versus vector: within Whisper alone, reducing temporal resolution at a fixed token budget costs a comparable amount. Under matched data and inputs, fine-tuned Qwen2.5-Omni-7B reaches 80.9%, against 95.9% for our 30 s variant; because the systems differ in backbone and adaptation, this is a system-level comparison. Accuracy remains high on sampled head and tail splits of the rephrased MUSIC-AVQA-R benchmark (96.5% and 95.6%). Because both encoders stay frozen and the music features are cached, the entire adaptation is cheap to train: the complete two-stage AVQA run takes approximately 5 hours on a single A100 80GB, and every run reported here fits on that one GPU.

Although Whisper models benefit from large-scale multilingual pre-training, their performance on Burmese medical speech remains limited. This work presents a Burmese medical speech recognition framework built on a high-quality 28-hour corpus recorded and validated by native speakers. We fine-tune Whisper models using full fine-tuning (FFT) and parameter-efficient fine-tuning (PEFT) with LoRA. To evaluate robustness, we apply waveform- and spectrogram-level data augmentation under controlled noise and simulated room acoustics. While augmentation reduces performance on clean speech, it significantly improves robustness in noisy and reverberant environments across FFT and PEFT settings. Our best-performing system, fully fine-tuned myMediWhisper-Medium without augmentation, achieves a state-of-the-art Word Error Rate (WER) of 23.44%, outperforming much larger general-domain fine-tuned models. Dataset and other resources can be found at the Huggingface repository: https://huggingface.co/datasets/LULab/mediTalk-mm-rdy.

Accurate and responsive turn-taking is essential for spoken dialogue systems, which must distinguish in real time between user interruptions, backchannels that should be ignored, and the completion of an utterance. Prior modular approaches typically optimize turn state prediction at the utterance or fixed-chunk level, creating a mismatch with the continuous turn state estimate, and often depend on an auxiliary ASR model, which limits responsiveness and increases overall system complexity. Therefore, we present X2-Turn, a frame-synchronous turn state prediction method via delayed-stream modeling. Specifically, building on the pretrained Voxtral Realtime model, we introduce a frame-synchronous turn state head that operates in parallel with the ASR head on shared streaming representations, jointly predicting ASR tokens and fine-grained turn states at the frame level. Experiments on bilingual EasyTurn and Full-Duplex-Bench demonstrate that the proposed method achieves an effective trade-off between turn state accuracy and decision latency.

The signal ambiguity of whispered speech drives ASR systems toward two opposing failure modes: failing to capture whispered speech or hallucinatory transcription of noise. This paper introduces the Whisper-Aware LLM, a framework that teaches an Audio-LLM to perceive and react to this uncertainty. Our model develops an intrinsic self-awareness by learning to quantify the physical deficiencies of acoustic signals through targeted self-supervised tasks. This learned uncertainty is then operationalized via a novel Confidence-Fused Decoding mechanism, which provides both high-level instructions and frame-level attention modulation to the LLM decoder. Our experiments confirm the effectiveness of this approach. The model sets a new state-of-the-art on whispered speech with a 17% relative CER reduction on AISHELL6-Whisper. At the same time, it directly addresses the reliability trade-off, with hallucination rates dropping from over 25% to 4.5%.

ASR-roundtrip evaluation is widely used as a scalable proxy for text-to-speech (TTS) intelligibility, but it can produce false negatives for reading errors perceived by listeners. We study Chinese news TTS spans whose correct reading depends on context or domain conventions, such as sports scores, aircraft models, technical units, and membership names. In these cases, Raw TTS can choose a plausible but wrong reading while ASR transcribes the audio as the intended or surface-correct text. A targeted audit over 110 high-risk MiMo TTS cases, reported with a complete denominator, confirms 46 masked false negatives, 9 exposed TTS errors, and 55 cases with no Raw TTS error. A span-isolation diagnostic re-exposes 18/46 previously masked errors. A Raw-only CosyVoice audit on the same targeted pool confirms 51 masked cases. Across the 97 TTS-specific audio files labeled confirmed masked across the two audits, Qwen3-ASR surface-recovers 40 cases, whereas Paraformer does so in only 2. The results suggest that ASR-roundtrip is useful for screening but insufficient as standalone ground truth for Chinese news reading-risk evaluation.

Many studies have shown that specially crafted inputs can induce large language models (LLMs) to generate excessively long outputs, resulting in significant computational overhead and resource consumption. While most existing denial-of-service (DoS) attacks target text-only LLMs, end-to-end (E2E) speech LLMs are rapidly emerging. Existing text-based DoS attacks primarily rely on prompt engineering, such as adversarial suffixes or semantic inducement, which exploit the discrete nature of text inputs and therefore cannot be directly transferred to continuous speech inputs. Moreover, prior studies on speech model security mainly focus on ASR or TTS systems, leaving the DoS vulnerability of E2E speech LLMs largely unexplored. To address this gap, we propose the perturbation-based DoS attack targeting E2E speech models. Instead of inducing long outputs through prompt manipulation, our method optimizes imperceptible acoustic perturbations to directly influence the model's autoregressive generation process while preserving the original input length. Specifically, we formulate the attack as a composite optimization objective that jointly suppresses EOS generation, encourages prolonged decoding, and largely preserves semantic consistency by integrating weighted EOS loss, top-k logit loss, length loss, and semantic alignment loss. To further improve stealthiness, we employ voice activity detection (VAD) to inject perturbations only into voiced regions. Extensive experiments on three open-source E2E speech LLMs demonstrate that our method achieves stable attack success rate while significantly increasing generation length and GPU resource consumption, revealing security risks in modern ALLMs.

Detecting phonemes from children's speech has historically been difficult due to the scarcity of training data, and unique characteristics of children's speech. During a phoneme detection competition, we found that training a lightweight model to predict the age of the learner, as well as the phoneme sequence, enabled a 94M-parameter model to outperform WavLM Large models (317M) on the target DrivenData distribution, and fall within approximately 0.04 CER of competition ensembles with 90 times the parameters. This has enabled the creation of PhonemeTrainer, an application that can run on most modern cellular phones. This will ultimately enable better Automated Speech Recognition (ASR) and pronunciation helper apps for children's speech, with the privacy and compliance benefits that come with edge processing.

Unconditional diffusion checkpoint merging assumes benign sources, yet a compromised public checkpoint can transfer a dormant backdoor while clean generation appears normal. Mitigation is difficult without knowing the compromised source, trigger, or target, and broad sanitization may degrade image quality. We introduce DiffSafeMerge (DSM), which uses a small unlabeled clean set and fixed, attack-agnostic stress probes to score source blocks, shrink suspicious contributions toward a trusted reference, and select attenuation under a clean denoising-loss budget. We evaluate four attacks, two datasets, and 21 target conditions. Intended merging already has zero worst-target ASR in 10 of 14 source cases; DSM preserves these outcomes and records no target match in the remaining four over three seeds, including three with baseline ASR of 48--100%. Among methods with zero worst-target ASR on both datasets, DSM obtains the lowest case-averaged FID in the matched seed-0 comparison.