Papers

Frozen self-supervised (SSL) speech encoders are strong, low-cost front ends for audio deepfake detection, and recent comparisons agree that large, multilingual, discriminative encoders generalize best out of domain. These comparisons fail to control for encoder capacity, pretraining objective, and multilingual coverage together, identifying which encoder wins without isolating why. We present a controlled decomposition with a fixed pipeline and trainable capacity. We vary multilingual coverage on four wav2vec2-family encoders, matched to ~315M parameters. We isolate the pretraining objective on two encoders matched on identical data. Coverage does not help monotonically, as out-of-domain error drops sharply at the ~100-language scale (XLS-R) but does not improve further at the 1406-language extreme (MMS). We find that a mid-coverage encoder is strongest on farther out-of-domain sets, matching or surpassing a 577M-parameter model at 315M. Its lead on these far sets, statistically significant under paired bootstrap, and on the official ASVspoof 5 cost metric holds under two backends. Separately, masked-prediction pretraining generalizes better than contrastive on identical data (In-the-Wild EER 26.5% vs. 46.8%). Within this fixed frozen-encoder recipe, we find that broader and larger models are not reliably better.

In cross-lingual zero-shot text-to-speech, the accent of the reference leaks into the target speech. We propose accent analogy guidance (AAG), a training-free sampler term that subtracts an accent direction estimated from the model's own predictions for one synthetic voice rendered in both languages, so the voice cancels and only the accent remains. By a blind LLM accent judge on real dubbing data, reweighting classifier-free guidance between reference and text, and its variants, stay near one identity-accent trade-off curve; we score a method by its speaker similarity above that curve at equal accent ($\Delta$SIM). Across four open TTS models AAG lies above the curve: on OmniVoice $\Delta$SIM is +0.11 to +0.27 on three test sets (accent 3.51 to 4.28 on a 1-5 scale at speaker similarity 0.29, where reweighting keeps 0.02); MaskGCT and CosyVoice 2 also lie above their curves, and on F5-TTS it is more native than any reweighting setting. An LLM-free language-ID measure and a twelve-listener panel agree. A premise test and the reach of a model's own curve indicate in advance whether and roughly how much AAG can gain, predicting the one model where it gains nothing (X-Voice).

The constant-Q transform (CQT) represents audio on a logarithmic frequency axis. Its nonstationary Gabor formulation is exactly invertible, but the unequal numbers of time coefficients in its bands complicate GPU computation. An exact factorisation combines spectral selection, conjugation, windowing, and reordering into a fixed map between one packed Fourier transform and the shorter band inverse transforms. The factors give waveform reconstruction, real adjoints for backpropagation, and bounds on arithmetic depth and block width; overlapping slices permit streaming with bounded memory. Tests on two GPU models show that Flash-CQT reduces analysis-synthesis round-trip time by factors of two to eight relative to a baseline computing the same CQT. The proposed implementation also uses over 30% less peak temporary workspace and reaches a negligible reconstruction error, with a signal-to-noise ratio of about 130 dB, in single-precision floating-point arithmetic. These advances make Flash-CQT a practical, computationally efficient front end for spectral analysis and modern audio machine-learning systems.

We present a personalized Korean visual speech recognition (VSR) system and quantify, on the nine-camera OLKAVS corpus, the gap between the population-level benchmark score and an individual user's error. A video-only Conformer initialized from English-trained weights attains 9.95 - 12.19% character error rate (CER) under the corpus protocol against the published 26.64, and 19.00 - 21.52 on unseen wording. Per speaker, CER spans 1.0 to 52.2%, with seen wording lowering CER by 7.0 - 9.0 points and professional delivery and spontaneous speech raising it by 8.5 - 10.5 and 12.7 points. A low-rank adapter with 4.6% of the parameters, trained on 4 to 29 minutes of the user's frontal video, lowers the CER of twelve high-error speakers by 2.13 to 3.58 points, transfers to every camera without loss, and keeps 85% of the full fine-tuning gain at 12% of its cost to other speakers. Cameras above the mouth plane add about six CER points as a constant offset that training on all views keeps small.

Post-training quantization (PTQ) reduces the cost of on-device text-to-speech (TTS), but published evaluations cover one system or method. We evaluate PTQ across TTS architectures under one protocol with three core models, weight and activation ablations of eight more, and two held-out models quantized blind. Four-bit per-channel weights reduce UTMOS, a predicted mean opinion score, by 2.8 on Supertonic and 0.07 on Kokoro, and per-tensor scaling can cause severe degradation even at 8 bits. The same bit width yields different outcomes, because the sensitive component is model-specific and not reliably predicted from the model class. A staged ablation procedure identifies it, and per-layer GPTQ can restore it to within 0.1 UTMOS. Real int8 and int4 kernels reproduce the simulated ordering at hardware-dependent cost. On a Mac mini, a 4-bit weight kernel runs Supertonic at 0.60x the fp32 latency while int8 is slower, so each configuration requires validation on the target runtime.

New words are invented every day. A human listener can learn a new word by hearing it clearly once and inferring its usage from sentence context. This paper proposes granting ASR a similar ability to learn the contextual representations and spellings of new words from unlabeled test data at test time. A frozen CTC acoustic model provides spellings, a frozen language model provides contextual evidence for out-of-vocabulary (OOV) word detection, and an adaptation module expands the vocabulary by learning the lexical token representations with distributions over CTC-generated candidates. The spelling model of each token is optimized by minimizing a Kullback-Leibler divergence (KLD) objective. We demonstrate that the CTC-weighted language model log likelihood ratio can be interpreted as the KLD between the unknown correct ASR and the unsupervised learned ASR, and that, using a Pinsker bound, the square root of KLD can be interpreted as an upper bound on the total variation distance between the true and estimated spelling of the unknown word. Experiments show relative OOV character-error-rate reductions of up to 14.97% on LibriSpeech and 6.67% on dysarthric Speech Accessibility Project data for recurring OOV words, relative to the corresponding rescoring system.

Domain Generalization (DG) aims to develop models that remain robust to conditions unseen during training. While DG has been systematically studied in computer vision through controlled benchmarks and diverse distribution shifts, its evaluation in spoken language recognition remains less structured. Existing speech datasets provide valuable benchmarks for robustness to real-world acoustic conditions, but are primarily designed around specific scenarios and large scale rather than as general-purpose tools for systematically constructing and evaluating domain shifts. In this work, we introduce a smallscale speech dataset and evaluation protocol for controlled studies of acoustic domain shifts. It enables the evaluation of spoken language recognition models under a variety of acousitc domain shifts. We introduce the speech modality into the DomainBed Domain Generalization framework and evaluate three domain generalization algorithms. We show that in-domain performance is not a reliable predictor of cross-domain robustness and verify that explicit domain invariant algorithms such as MMD or DANN algorithms do not outperform ERM. We further validate the generality of these findings on MMS-LID-126, a state-of-the-art spoken language identification system. We release the code.

Full-duplex dialogue systems, which listen while speaking, must distinguish a completed turn from a pause within a turn and an interruption that requests a turn from a brief acknowledgment or speech addressed to a third party. Yet existing conversational corpora provide limited control over these events and limited labels for their intent. We present a pipeline for synthesizing intent-labeled, two-channel conversational speech from relational event lists. An LLM authors each event's speaker, text, conversational act, and attachment to an earlier event without predicting absolute timestamps. Events are synthesized independently, aligned with their source text, and placed on a shared clock, so turn-taking landmarks are measured from the rendered signal while silence durations are specified or sampled from turn-taking distributions. The pipeline covers 42 phenomena across eight families in English and Mandarin, derives frame-level system actions from authored intent, and promotes diversity using small, diverse sets of prior examples and batch prompts that request alternatives with self-reported probabilities. Ablations show gains in each targeted diversity dimension. On a four-action label space for taking, holding, releasing, and not holding the conversational floor, a semantic voice-activity detector using only current and past audio reaches start-speaking and start-listening F1 scores of 0.819 and 0.802. When generating its own responses, the full-duplex speech model Moshi takes 0.85 of the reference turns after fine-tuning on the generated corpus, compared with 0.44 before fine-tuning. Its frame-level precision for predicting system-floor occupancy rises from 0.46 to 0.88. With reference context at each step, its frame-level floor F1 rises from 0.893 to 0.962. These results show that controlled synthesis can provide learnable and transferable supervision for full-duplex turn management.

Self-supervised learning (SSL) countermeasures (CMs) have shown strong performance in recent years. However, they often show degraded performance while facing unseen spoofing attacks and mismatched conditions. This study examines the Voxtral audio-language model (ALM) framework for spoofing detection, as a step toward combining CM capabilities within the ALM framework. We analyze how Voxtral captures spoofing cues through audio-text processing and propose an instruction-guided approach that uses label-sequence likelihoods to evaluate bonafide and spoofed speech. Experiments on the ASVspoof databases show that without task-specific adaptation, the LLM layers emphasize semantic representations, reducing the separability of spoof-discriminative acoustic cues compared to the Whisper-based audio encoder. Consequently, spoofing-related information becomes less separable after language-model processing. We also applied lightweight adaptation using weight-decomposed low-rank adaptation (DoRA) to the Voxtral model and propose the Spooftral model, achieving an equal error rate (EER) of 4.25% on the ASVspoof5 evaluation set.

Neural methods for audio restoration are typically framed as rigid mappings from degraded inputs to single clean outputs, enforcing decisions about what audio content is removed, and potentially adding unwanted content to the restored signal. Because what constitutes a restored audio signal is subjective, we introduce RESTORE, a framework that formulates audio restoration as a six-source semantic decomposition to allow for real-time interactive user control over the process. By expanding a pretrained HTDemucs backbone, a single forward pass disentangles a degraded mixture into vocals, music, broadband hiss, impulsive transients, and an unmodeled residual, while jointly synthesizing a high-frequency extension. Users may steer the restoration by adjusting stem gains, ensuring generative content remains isolated and auditable. RESTORE improves audio quality on diverse historical recordings compared to baselines,lowering Frechet Audio Distance (FAD) (12.13 VGGish; 0.92 CLAP) and delivering aesthetic steerability (Spearman rho greather than 0.91) at 50x real-time on a single GPU. Code and audio samples are available at https://melissachen15.github.io/restore-audio-demo.

The growing use of spatial audio in applications such as augmented and virtual reality has driven the development of binaural reproduction methods for wearable arrays with a limited number of microphones. Binaural signal matching (BSM) is one such method, producing high-quality binaural signals under a diffuse-field assumption, but degrading at high direct-to-reverberant ratios (DRR) where the direct sound dominates. Previous extensions incorporate Field-of-View (FoV) weighting, either with fixed apertures or based on explicit source localization, but these approaches are limited by coarse spatial coverage or reliance on localization estimation accuracy. This paper introduces FoV-BSM-Net, a signal-dependent FoV-BSM formulation that avoids explicit source estimation by learning the FoV parameters end-to-end from the microphone signals using a Convolutional Recurrent Neural Network. The method is evaluated in simulated rooms across varying reverberation conditions, and compared against BSM and a fixed FoV-BSM baseline. Results show that FoV-BSM-Net consistently improves over BSM, with gains that grow with DRR in both binaural NMSE and interaural cue errors, and are further supported by perceptual evaluation showing a substantial advantage over both baselines across low and high DRR conditions.

The ResNet-18 image classification model is employed to determine the physical parameters of a plate reverb from a recording of the impulse response. The model is adapted to derive parameters using normalized and downsampled multi-resolution spectrograms computed from the provided impulse responses (IRs). To refine the prediction of the output location, the spectral phase response is also included as an additional input channel to the network since multiple output locations can give the same magnitude response for high-order resonant modes. On a 5000 IR validation set, our model achieves an average normalized mean squared error (NMSE) of 0.02920 across all parameters, with the lowest average NMSE occurring for parameters yo (0.00228), Ly (0.00347), and xo (0.00574).

Text-to-audio (TTA) generation aims to synthesize realistic audio that faithfully reflects natural-language descriptions. Most TTA systems adopt a two-stage latent paradigm: an audio tokenizer is optimized for reconstruction and then frozen, after which a generative model is trained in the resulting latent space. However, reconstruction-oriented representations may be suboptimal for generation, motivating joint representation and generative learning. To this end, we introduce Unite-Audio, to our knowledge, is the first to jointly learn continuous audio representations and latent flow matching for TTA. By coupling reconstruction with self-supervised generative prediction, Unite-Audio allows the generative objective to directly shape the latent space rather than treating it as a fixed intermediate representation. We further employ Flow-GRPO post-training to improve text-conditioned generation. Experiments show competitive TTA performance with a compact latent flow model, while ablation studies confirm the benefit of jointly learning the audio representation and generative model. Audio samples are available at https://runwushi.github.io/Unite-Audio.

Early reflections in spatial room impulse responses (SRIRs) measured using spherical microphone arrays (SMAs) play an important role in spatial audio analysis, rendering, and reverberation modeling. Their accurate localization and characterization can facilitate the analysis, processing, and manipulation of measured reverberation fields. This paper proposes a method for detecting and characterizing early reflections based on the Herglotz wavefunction formalism. The measured sound field is represented as a continuous superposition of incident plane waves, from which a localization function is derived in the spherical harmonic domain. An adaptive radial Gaussian fitting procedure is then used to estimate both the directions of arrival (DoAs) and the number of incident reflections within a single analysis frame. The proposed framework is evaluated using simulated and measured SRIRs and compared with conventional steered response power (SRP) and multiple signal classification (MUSIC) localization methods. The results demonstrate improved localization accuracy and more reliable detection of multiple simultaneous reflections, highlighting the potential of the proposed Herglotz-based formulation for the analysis of early reflections in SMA-measured SRIRs.

Outer-hair-cell (OHC) loss is a primary deficit of sensorineural hearing loss (SNHL), impairing cochlear amplification and frequency selectivity and thereby elevating hearing thresholds. Biophysically-inspired DNN-based hearing-aid (HA) algorithms have been proposed to compensate for OHC deficits and have shown clear benefits in objective speech intelligibility and quality metrics (e.g. HASPI, HASQI). However, comprehensive subjective validation of these benefits in human listeners is still missing. In this work, we present a subjective evaluation of a biophysically-inspired HA model targeting OHC deficits. The cochlear module of an auditory model was individualized based on each listener's pure-tone audiogram and integrated into a trainable system, which includes the personalized model and a normal-hearing reference model, and the resulting trained HA was evaluated using a Matrix test comparing intelligibility scores for unprocessed and HA-processed noisy speech. The results revealed a significant benefit of the HA model over the unprocessed condition in the range of +1 to +27%, providing behavioral confirmation of the efficacy of the HA model. This study closes the gap between objective and perceptual evidence for this new generation of DNN-based HA algorithms, paving the way for their integration into next-generation DNN-accelerated chips for hearables and hearing aids.

Generating controllable and human-like emphasis remains an open challenge in text-to-speech, even when explicit emphasis control signals are provided in the text input, limiting the communicative accuracy of synthetic speech in real-world applications. Reinforcement learning has recently shown promise for post-training TTS systems to align with human preference, yet existing methods have not been applied to word-level prosodic control. We present EmphTTS, a non-autoregressive TTS system that applies Group Relative Policy Optimization (GRPO) to the duration predictor with an emphasis localization reward, enabling direct optimization for word-level emphasis. Evaluations show that EmphTTS achieves the best emphasis controllability and performs the best in emphasis objective evaluation. In subjective preference tests, EmphTTS is significantly preferred over synthetic groundtruth and most baselines. Ablation studies show that GRPO improves emphasis realization beyond supervised-finetuning-based duration modeling and simple speaking-rate adjustment, while alleviating the mismatch between the independently trained duration predictor and TTS model.

Recent one-step text-to-audio (TTA) models substantially reduce inference cost, yet their generated distributions can still be improved through post-training. We propose DriftAudio, a distributional post-training method that adapts Drifting to pretrained one-step TTA generators. Applying Drifting condition-wise is challenging under free-form text conditioning, where only one or a few real samples are typically available for a particular condition. DriftAudio instead performs Drifting on the marginal audio distribution while retaining text-conditioned generation. The drifting field is estimated in a frozen audio feature space using resampled real samples, a rolling bank of generated samples, and the current batch of generated samples. The resulting Drifting field provides a detached training target for updating only the generator, keeping the original one-step inference procedure. On AudioCaps, starting from MeanAudio, DriftAudio reduces FAD and FD by 33.9% and 17.6%, respectively, while also improving KL and CLAP. Starting from FdAudio, it further reduces FAD, FD, and KL, with trade-offs in IS and CLAP. These results demonstrate the effectiveness of marginal distributional post-training for one-step TTA generation.

This paper summarizes the Interspeech2026 second Multilingual Conversational Speech Language Model (MLC-SLM) Challenge, which aims to advance the development of effective multilingual conversational speech language models. We describe the two challenge tasks: multilingual conversational speech diarization and recognition, and multilingual conversational speech understanding, together with the released real-world conversational speech dataset, evaluation protocols, and baseline systems. The challenge attracted 91 teams worldwide, with 704 valid leaderboard results and 14 technical reports across the two tasks. Based on the participating systems, we summarize representative approaches and distill practical insights into multilingual conversational speech recognition and understanding to support future research in the community.

Anomalies are rare, and anomalous data are often unavailable during development, making it difficult to determine which anomaly detection models and configurations will generalize to unseen anomalies. Recent approaches address this challenge by generating pseudo-anomalies and using bounds on the achievable area under the ROC curve (AUC) to select the optimal configuration from a finite set of candidates. Instead, we use the AUC bound as a differentiable, anomaly-free objective for directly optimizing continuous parameters of anomaly detection systems. We demonstrate this framework by optimizing ensemble weights and introducing a learnable score-rescaling mechanism that adapts pseudo-anomaly scores, enabling optimization beyond a predefined candidate set. Experiments across multiple datasets and embedding models show that AUC-bound optimization achieves significant performance gains over conventional model selection and prior development-set-based parameter selection. The results further show that direct optimization is less sensitive to the choice of pseudo-anomaly construction.

Speaker diarization must solve two problems: counting how many speakers are present in a given conversation, and assigning speech to each one. This becomes harder in crowded conversations with five or more speakers. We evaluate how end-to-end neural diarization models count speakers, and find systematic under-counting in crowded recordings. The standard diarization error rate (DER) hides this failure, because it is duration-weighted and barely penalizes the dropped, low-activity speakers. We therefore also report the Jaccard error rate (JER) and explicit counting metrics. We propose a gated loss that couples speaker existence with frame activity. This loss can be computed in two ways, duration-weighted or speaker-weighted, and the speaker-weighted variant, used as a regularizer, reduces the under-count. On real-world crowded recordings our method clearly improves the counting metrics, lowering JER by about 6% relative and DER by about 9% relative, while leaving sparse recordings unharmed.

Self-supervised learning (SSL) for speech and audio has largely progressed along separate tracks: speech models emphasise time-domain prediction, whereas audio representation learning has focused on time-frequency patterns. This separation creates a compatibility gap, limiting cross-domain generalization. In this work, we introduce JASPER, Joint Audio and Speech Pre-trained Encoder Representations, a framework that augments speech-pretrained models with time-frequency objectives. Specifically, JASPER performs masked prediction of temporal and spectral targets over long audio segments, enabling spectro-temporal representation learning of speech and audio signals. The proposed method consistently outperforms multiple baselines and existing speech/audio encoders on diverse speech, audio, and music tasks, demonstrating the effectiveness of unified spectro-temporal modeling.

Voice conversion (VC) systems fall into two families: non-parametric embedding-space methods, which need no trained model but degrade on short target utterances, and spectrogram-based neural architectures, which achieve strong quality via multi-module pipelines with tens of millions of parameters. We propose kNN-FM-VC, a single conditional flow-matching network that learns to approximate the kNN-VC mapping between WavLM embedding distributions of source and target speakers, replacing explicit pointwise kNN matching with a neural regressor trained on kNN-generated pairs. The model is conditioned on the target speaker via cross-attention and FiLM, and trained under three Gaussian conditional paths (Schrödinger bridge, straight line, and constant-variance Gaussian tube), enabling few-step sampling. Unlike Phoneme Hallucinator, which uses an upsampling stage followed by kNN matching, our 13M-parameter model performs conversion with a single learned network and supports one-step inference. On LibriSpeech, the one-step Gaussian Bridge achieves lower WER and higher estimated speech quality than FreeVC and Phoneme Hallucinator. Relative to kNN and kDOT, it substantially reduces WER.

Pneumonia is difficult to diagnose in older long-term care residents; multimorbidity and atypical presentations obscure signs, motivating operationally efficient objective testing. We analyzed multi-channel digital stethoscope recordings from 185 Japanese residents (73 pneumonia, 112 symptomatic without), using radiologist-confirmed chest X-rays and clinician diagnoses as supervisory signals that train convolutional neural networks, multimodal fusion, and channel-based variants with time-domain Grad-CAM interpretability. Models were evaluated with repeated patient-level cross-validation showing models with X-ray supervision outperformed clinician supervision (F1 0.729, accuracy 0.783 vs. F1 0.637, accuracy 0.711). Additionally, a three-channel selection protocol maintained performance (F1 0.736; accuracy 0.803), with two mid-thoracic sites ranking highest and Grad-CAM attention overlapping adventitious sounds. These findings indicate automated multi-channel lung-sound analysis can aid long-term care pneumonia diagnosis, with X-ray supervision being more reliable than clinical, and fewer channels preserving performance while lowering acquisition times.

Monitoring cognitive impairment (CI) in motor neuron disease (MND) is essential for timely treatment and care, yet challenging due to co-occurring speech difficulties. The Edinburgh Cognitive and Behavioural ALS Screen (ECAS) provides a robust metric for CI assessment, with the Verbal Fluency Index (VFI) a central element. Building on recent advances in automated speech analysis, this study proposes a system for estimating VFI. It leverages a unique MND dataset and combines ASR (WhisperX) and VAD (Silero) with refined timestamping to predict the VFI and extract several clinically interpretable measures. Our approach outperformed systems based on traditional acoustic features and self-supervised embeddings, evaluated using multiple regression algorithms. Clinically inspired features consistently outperformed the other sets, with the best models achieving strong results (P-words: R2 0.9, NRMSE 0.05; S-words: R2 0.8, NRMSE 0.08), demonstrating the feasibility of automated VFI estimation.

Linear arrays of omnidirectional microphones produce beampatterns that are symmetric about the array axis. For these arrays, a beamformer is considered steerable if its beampattern maintains the same shape in the semicircular plane across all look directions from 0{\deg} to 180{\deg}. For dual-microphone arrays, conventional differential beamformers are generally non-steerable and restricted to first-order, which significantly limits spatial selectivity. To address these limitations, this study presents a neural differential beamformer (NDBF) with a dual-microphone array. The contributions are as follows: (i) NDBF is steerable; (ii) NDBF achieves high-order frequency-invariant beampatterns; and (iii) NDBF enables stereo recording using only two closely spaced omnidirectional microphones. Experimental results demonstrate that NDBF outperforms existing methods while overcoming the limitations of classical differential beamforming.

The inclusion of generative audio in the music production process has led to an increase in hybrid music tracks that blend authentic human performances with AI-generated stems, challenging traditional AI music detectors which operate in a binary setting. In this work, we propose a stem-agnostic framework for identifying synthetic audio sources within hybrid musical mixtures. We introduce the inspectrogram, a novel time-frequency representation that maps localized probabilities of synthetic content across the audio spectrum. By combining the inspectrogram with a Wiener filter estimating target stem energy dominance, a single CNN model evaluates whether the specific stem is generated. Trained on rendered hybrid mixtures and evaluated across various stem classes, our model achieves strong performance on high-frequency sources such as vocals, drums, and guitar, but struggles on the low-frequency, narrow-band bass. We conclude that the quality of separation impacts the detection accuracy and identify source separation as a primary bottleneck and a crucial direction for future research.

Temporal envelopes carry cues critical to speech intelligibility, yet ASR frontends based on log-mel spectrograms do not explicitly model continuous sub-band envelope structure. This limitation is particularly acute for children's speech, where high acoustic variability demands robust feature representations. We propose a modular time-domain frontend that decomposes speech into sub-band envelopes using mel-spaced windowed-sinc filters and the Hilbert transform, with learnable per-channel energy normalization (PCEN) jointly optimized with the Whisper model. On the MyST children's speech corpus, systematic ablations identify full-band windowed-sinc filters, Hilbert envelopes, a 25 Hz smoothing cutoff, and learnable PCEN as the best configuration. Under the same Whisper-small fine-tuning setup, the frontend reduces WER from 13.16% to 11.08%, a 15.8% relative reduction over the log-mel baseline, and outperforms the evaluated Kid-Whisper checkpoint on the same cleaned test split. These results show that temporal-envelope representations and learnable frontend normalization are effective complements to backend adaptation for children's ASR.

Active speaker detection (ASD) requires reliable association between visible faces and acoustic speech, yet existing systems often degrade under challenging domains or incomplete observations. We introduce ROAM-ASD, a robust audiovisual framework that jointly models audio, full-face, and fine-grained mouth representations. A unified joint self-attention mechanism processes all input streams together with modality-agnostic query tokens, enabling direct interaction among available modality inputs. Modality dropout further improves robustness when input streams are unavailable. ROAM-ASD achieves state-of-the-art performance across five ASD benchmarks: 98.8% mAP on WASD, 87.9% on UniTalk, 96.5% on AVA, 99.3% on ASW, and 98.2% on Talkies, improving over previous best systems by 5.1, 4.7, 0.9, 1.0, and 2.1 mAP points, respectively. ROAM-ASD also substantially improves zero-shot cross-dataset generalization and remains robust to missing observations.

In LLM-based speech translation, transcription-based chain-of-thought (CoT) suffers from a mismatch between reference transcripts used in supervised fine-tuning (SFT) and model-generated transcripts at inference. To address this, we propose joint recognition and translation fine-tuning via group relative policy optimization (GRPO). We score both transcripts and translations, with translation conditioned on model-generated transcripts, and compare three token advantage strategies. Using Qwen2.5-Omni-3B across four languages, we evaluate CoT against direct speech translation (Direct ST) under SFT and GRPO, training on CoVoST 2 and testing on CoVoST 2 and FLEURS. CoT GRPO outperforms Direct ST GRPO by 1.77 and 0.83 average BLEU points on CoVoST 2 and FLEURS. Compared to CoT SFT, GRPO boosts BLEU by 0.82 and 0.67 points and reduces word error rate (WER) by 8.8% and 7.2% relatively. These results highlight reinforcement fine-tuning as an effective method to mitigate the training-inference mismatch, jointly improving recognition and translation.

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