Wav2Vec

Short-duration speaker verification (SDSV) is crucial for personalized keyword spotting, where test utterances are typically shorter than three seconds. Limited speech duration results in unstable speaker representations and increased sensitivity to noise and phoneme variations, thereby degrading performance. To investigate this issue, we construct VoxPhrase, a large-scale SDSV corpus automatically segmented from the VoxCeleb dataset. Our analysis shows that text-dependent (TD) enrollment is constrained by duration and yields unstable speaker representations. In contrast, although text-independent (TI) enrollment introduces content mismatch, its representations become more stable as the enrollment duration increases. Accordingly, we propose a hybrid-enrollment neural re-scoring framework that combines TD and TI enrollment and performs frame-level comparison via parallel cross-attention. Experiments on VoxPhrase demonstrate consistent improvements across multiple speaker models.

For the last couple of years, the field of speech synthesis has improved dramatically thanks to deep learning. There are more and more deep learning-based TTS systems developed to make it possible to produce voices with high intelligibility and naturalness. Meanwhile, controlling the expressiveness is yet a big deal, generating speech in different styles or manners has received a lot of attention from community recently. This paper aims to give our solutions to deal with the task emotional speech synthesis (ESS) at VLSP 2022 which allows to generate humanlike natural-sounding voice from a given input text with desired emotional expression. By integrating speaker embedding, prosody bottleneck into FastSpeech 2, our systems can promisingly generate emotional speech of a single speaker (Sub-task 1), transfer speaking styles from another speaker to the target speaker with neutral non-expressive data while retaining the target speaker's identity (Sub-task 2).

Recent advances in speech generation have significantly improved the naturalness of synthetic speech, making spoofing detection increasingly challenging. A key limitation of current anti-spoofing systems is their limited robustness to unseen synthesis methods. In this work, we transform a self-supervised speech representation model into a Mixture-of-Experts (MoE) architecture to improve generalization. Feed-forward blocks in selected encoder layers are replaced by multiple expert networks controlled by a layer-wise gating mechanism, allowing experts to capture complementary acoustic patterns while preserving the representations learned during self-supervised pretraining. We further analyze the architectural choices affecting the performance of this MoE conversion and investigate the activation behavior of the experts. The proposed approach is evaluated on 14 spoofing datasets and reduces the macro EER from 5.46% to 4.81%, corresponding to 11.9% relative improvement over the baseline.

The choice of speech representation is critical in speech-driven 3D facial animation. Representations differ in what they encode: SSL features emphasize segmental and semantic cues, neural codecs yield latents optimized for acoustic reconstruction, and ASR-style objectives produce label-based spaces. We evaluate four speech representation families for 3D facial synthesis, comparing their facial reconstruction quality across two facial decoders using objective metrics and a perceptual evaluation. We additionally conduct probing analyses that relate tokenized representations to phonetic units and to articulatory deformations. We found that encoding phonetic classes is beneficial for accurate facial animation prediction on both semantic and label-based representations with comparable facial animation quality. From the latter, we introduce an Audio Visual Text-to-Speech (AVTTS) pipeline that leverages, as a shared space, discrete representations to decode speech and 3D facial motion.

Modern pretrained self-supervised automatic speech recognition models are trained on large-scale audio data to encode speech into contextualized representations. However, their training data are heavily skewed toward high-resource languages with little data from low-resource languages, raising concerns about the potential underrepresentation of typologically uncommon speech sounds such as click consonants primarily found in Khoisan languages. This leads to our central research question: Can these models recognize click consonants as accurately as other speech sounds? To address this question, we fine-tune and compare pretrained self-supervised speech models (Wav2Vec2 and HuBERT) on data from two click-rich Khoisan languages (G|ui and West !Xoon). Our results reveal that the fine-tuned models consistently recognize clicks more accurately than non-clicks, suggesting that self-supervision enables generalization across human speech sounds including rare phonemes.

We investigate what self-supervised speech recognition models (S3Ms) learn about speaker groups (SGs). We examine several states of S3Ms: pretrained, finetuned on speaker identification (SID), finetuned on automatic speech recognition (ASR), and ASR-finetuned using a fairness enhancing algorithm. We find that S3Ms encode information about several speaker group categories (SGCs), including their gender, age, dialect, ethnicity, and whether they are a native speaker. We find that finetuning for SID amplifies certain SGCs, namely those whose variance is more phonetic in nature, though it does not amplify other SGCs, namely those whose variance is more semantic in nature. On the other hand, finetuning for ASR discards phonetically variant speaker group information (SGI) but retains semantically variant SGI. We find that ASR algorithms designed for fairness improvement change to what extent SGI is encoded in S3Ms; however, this is primarily true for for phonetically variant SGCs, and less true for semantically variant SGCs. We discuss how SGI is encoded by each layer, and identify subdimensions of embeddings responsible for encoding different SGCs. Finally, we discuss how our findings could be beneficial in designing fairer ASR algorithms.

We introduce SSL-GMMVC, an interpretable voice conversion method in self-supervised speech space. The method models paired source-target features with a Gaussian mixture model and performs conversion as a posterior-weighted sum of affine transforms. This yields locally linear transformations that adapt to heterogeneous feature-space structure while remaining analytically tractable. Through objective and subjective evaluations, we show that SSL-GMMVC improves speaker similarity with comparable intelligibility and naturalness, and that even a constrained covariance variant surpasses a deep learning baseline as the number of mixture components increases. Further analyses link component selection to phonetic structure and reveal interpretable scaling and rotation in the learned transforms. These findings highlight SSL-GMMVC as an effective, analyzable framework for voice conversion.

Speech Emotion Recognition (SER) aims to identify a speaker's emotional state from audio signals. While recent advances in deep learning have significantly improved SER performance in Indo-European languages, Arabic SER remains underexplored and challenging due to dialectal diversity, limited annotated datasets, and the difficulty of modeling both local spectral cues and long-range temporal dependencies. To address these limitations, this study investigates whether hybrid architectures that jointly model spatial and contextual information can improve emotion recognition in Arabic speech. We propose and evaluate a comparative framework involving three architectures: a CNN-LSTM model, a CNN-Transformer model, and a fine-tuned wav2vec 2.0 model. The first two models leverage MFCC and spectrogram-based representations, while wav2vec 2.0 operates directly on raw audio through self-supervised representations. Experiments conducted on the EYASE and BAVED datasets demonstrate that the proposed CNN-Transformer architecture significantly outperforms the other models, achieving an accuracy of 98.1 percent. This result highlights the effectiveness of combining convolutional feature extraction with Transformer-based global context modeling. The main contribution of this work lies in providing a systematic comparison of hybrid and self-supervised approaches for Arabic SER, and in demonstrating that CNN-Transformer architectures offer a robust solution for capturing both spectral and long-range dependencies in low-resource and dialectally diverse settings.

Speech sound disorders affect approximately 44% of Korean pediatric communication disorder cases, yet automated assessment tools for Korean toddler speech remain underdeveloped. This paper presents an end-to-end pipeline for automated pronunciation evaluation of Korean toddler speech, combining neural speaker diarization with self-supervised speech representation learning. We introduce a novel IRB-approved corpus of 53 recordings from Korean-speaking children aged 2-5 years. A subset of 53 subjects was annotated by three independent reviewers, yielding 1,190 consonant and 748 vowel word-level binary correctness labels. We evaluate three diarization models, finding that NeMo SortFormer achieves 88.69% speaker count accuracy and 33.04% diarization error rate (DER) owing to its arrival-time-sorted transformer architecture, which handles the acoustic confound between young female caregivers exhibiting aegyo and toddler speech. For pronunciation scoring, we compare three self-supervised learning (SSL) backbones across multiple pooling strategies. A cross-model ensemble routing consonant prediction to HuBERT-large and vowel prediction to WavLM-large achieves balanced accuracies of 0.720 and 0.845, with a mean of 0.782.

Parkinson's disease (PD) is a progressive neurodegenerative disorder that frequently causes speech impairments associated with hypokinetic dysarthria. As speech production relies on the precise coordination of complex neuromuscular mechanisms, speech analysis has emerged as a promising non-invasive and cost-effective biomarker for early PD detection. Recent deep learning approaches have shown encouraging results; however, most existing methods rely on a single speech representation, potentially overlooking complementary pathological information encoded across different feature spaces. In this work, we propose a multi-branch deep learning framework for automatic PD detection from speech. Each recording is segmented into 5-second chunks and represented using three complementary modalities: Log-Mel spectrograms, MFCCs, and HuBERT embeddings extracted from raw waveforms. The spectrograms are processed using a pre-trained ResNet-18 encoder, MFCC sequences are modeled through a BiLSTM network, and raw speech is encoded using a pre-trained HuBERT model. To effectively integrate these heterogeneous representations, we introduce a context-guided cross-modal attention mechanism that dynamically weights temporal HuBERT embeddings according to the global acoustic context derived from the spectrogram and MFCC branches. Experiments conducted on the publicly available Spanish PC-GITA corpus under strict speaker-independent 5-fold cross-validation demonstrate the effectiveness of the proposed approach. The proposed architecture achieves an accuracy of 91.51%, an F1-score of 91.24%, and an AUC of 95.97%. Furthermore, ablation studies confirm the contribution of both the proposed context-guided cross-modal attention mechanism and the integration of complementary speech representations. These findings highlight the potential of heterogeneous speech modeling for robust and clinically reliable PD detection.

We present a voice conversion (VC) framework that utilizes K-Nearest Neighbors (KNN) retrieval over WavLM representations to align non-parallel source and target speech, constructing synthetic training pairs for supervised learning. The retrieved segments serve as synthetic inputs, while real target audio provides ground-truth outputs, forming a synthetic-to-real training paradigm that naturally supports multilingual data without requiring parallel corpora or explicit alignment. To ensure consistent target-speaker identity, we incorporate a speaker loss derived from a pretrained speaker verification model. Experiments across multiple languages demonstrate that the proposed approach achieves high naturalness and strong speaker similarity, outperforming competitive VC baselines, despite being trained exclusively on English data. Samples can be accessed at: https://palindromic-vc.github.io.

Sophisticated generative speech technology can undermined the reliability of voice biometrics. While spoofing detection systems excel when assessed under in-domain conditions, generalisation to out-of-domain settings is often poor. In this paper, we show that such issues could be caused by speaker bias, where models learn individual voice traits rather than markers of manipulation or generation. We propose a teacher-student framework for speaker-invariant spoofing detection that disentangles identity without requiring speaker labels. We leverage a pre-trained speaker recognition teacher to guide a student model via a gradient reversal layer. To control the balance between suppressing cues related to voice identity with the preservation of those related to spoofing detection, we integrate a Variational Information Bottleneck. Evaluations across nine datasets show our model achieves a 25.7% relative reduction to the EER compared to the MHFA baseline.

Self-supervised learning (SSL) has achieved remarkable representation learning performance, but many existing methods rely on large batch sizes, memory banks, momentum encoders, or global synchronization mechanisms that substantially increase computational cost and training complexity. In this work, we propose Semantic Mutual Information (SMI), a lightweight self-supervised objective derived from a mutual-information-inspired dependency formulation under Gaussian assumptions. Unlike conventional correlation matching objectives that operate on high-dimensional feature correlation matrices, SMI performs optimization on a sample-level dependency matrix through a nonlinear transformation of pairwise correlations. This formulation induces distinct optimization dynamics that emphasize strongly dependent semantic pairs while maintaining representation diversity. Experimental results on ImageNet using a ResNet-50 backbone demonstrate that SMI achieves competitive linear evaluation performance relative to state-of-the-art SSL approaches while substantially reducing computational complexity. Across multiple low-resource benchmarks, SMI consistently improves transfer performance over Barlow Twins, particularly on fine-grained datasets. Furthermore, analyses of optimization dynamics and representation geometry suggest improved alignment--redundancy balance, greater feature diversity, and more spatially localized semantic representations. These results indicate that nonlinear dependency optimization provides an effective and computationally efficient alternative to conventional correlation-based self-supervised learning objectives.

Discrete speech units obtained via k-means clustering of self supervised embeddings entangle phonetic, speaker, and language information, causing speaker mixing and cross-lingual interference in multilingual multi-speaker speech generation. Despite growing use in Audio LLMs and speech to speech systems, unit vocoders remain underexplored. We analyze a BigVGAN based unit vocoder, across four Indian languages. We study the interaction between cluster size and conditioning strategies using WER, speaker similarity, and unit level metrics. Results show that cluster size governs intelligibility by improving phonetic discriminability, while explicit speaker conditioning is indispensable for preventing identity collapse. Language supervision yields further gains mainly at lower cluster sizes where units remain ambiguous. Our analysis shows similar phonemes across languages collapse to the same cluster IDs at smaller inventories, with larger clusters progressively separating them.

Audio-Visual Speech Recognition (AVSR) enhances speech recognition robustness by leveraging visual cues, while real-world scenarios remain challenging due to viewpoint variation, audio distortion, and visual occlusion, which degrade modality quality and increase audio-visual asynchrony. In this paper, we propose a novel Modality-aware Multi-view Self-supervised representation framework for robust Audio-Visual Speech Recognition (M2S-AVSR). First, we introduce a multi-view representation learning encoder to learn view-invariant visual speech representations. Next, we employ a modality-aware module that explicitly models modality quality and cross-modal synchrony to perform fine-grained modality-aware fusion, enabling fine-grained visual information injection during decoding. In addition, we release AISHELL8-RealScene, a public multi-scenario, multi-view conversational audio-visual dataset recorded in real-world environments, and establish a speech recognition benchmark on it. Experiments on English and Mandarin benchmarks demonstrate the effectiveness of the proposed method under challenging conditions. On LRS3, M2S-AVSR achieves up to 29.4% relative improvement under viewpoint perturbation and visual degradation settings. Our method also achieves new state-of-the-art performance on the MISP2021-AVSR test set. On AISHELL8-RealScene, it achieves the best result in outdoor scenes. The proposed method and dataset provide useful support for future research on robust speech and multimodal tasks under realistic conditions.

Objective: laryngectomees depend on an electromechanical device to generate electrolaryngeal (EL) speech. Compared with normal speech, EL speech suffers from severe distortion, limited phonetic variation, unnatural prosody, and temporal shifts, degrading naturalness and intelligibility. Although sequence-to-sequence (seq2seq) voice conversion (VC) based EL-speech-to-normal-speech conversion (EL2SP) is promising, substantial mismatches between EL and normal speech inevitably cause cumulative mapping errors that limit performance. To address this, we describe a novel representation learning framework integrating speech and text representations to improve mapping and reconstruction quality within a seq2seq VC model. Methods: our methodology comprises two main stages: 1) representation integration and learning, and 2) reconstruction training. A network capable of incorporating auxiliary text information is first constructed with pretrained modules to learn speech--text-based integrated representations. Then, an autoencoder-style reconstruction strategy finalizes EL2SP model to inherit these representations without increasing model complexity. We introduce three fusion strategies including middle-, input-, and hybrid-level fusion strategies that progressively enhance learning. Moreover, besides standard seq2seq VC objectives, an additional reconstruction loss on the integrated representation is introduced to refine representation transfer. Results: experiments under different EL2SP datasets consistently demonstrate that our methods, combined with data augmentations, outperform baselines relying solely on speech representations. Furthermore, progressive improvements with system design depth validate the effectiveness of our methods. Significance: the proposed methods provide an extensible and practical methodology for EL speech enhancement and assistive communication technologies.

In real-world scenarios of linguistic steganalysis, tested texts usually come from unseen domains with different vocabularies, topics, writing styles, and steganographic generation patterns, which can significantly degrade the detection performance. Although existing cross-domain steganalysis methods can effectively alleviate this problem through distribution alignment, domain-invariant feature learning, etc., the detection performance is not satisfactory. In this paper, we propose a post-training representation editing method for cross-domain linguistic steganalysis. Specifically, the detector is first trained on source-domain data, and then the feature extractor and classifier are kept frozen, and the intermediate representations are deterministically edited before classification. For domain adaptation, we construct a domain-offset vector from marginal source and target representations. For domain generalization, we derive a source-domain cover-to-stego direction to guide sample-specific editing. Experimental results show that compared with the advanced methods, the proposed method can achieve high cross-domain detection performance, especially in terms of F1-score, while requiring no architecture modification or parameter updates after source-domain training.

The preservation of under-resourced languages requires digital tools and resources shaped by and for their speakers. We present the first dedicated ASR resources for Puno Quechua (ISO 639-3: qxp): (1) the largest speech corpus for any single Quechua variety, consisting in 66 hours of recordings for scripted and spontaneous speech (including 36 hours of manually transcribed and validated data), collected via a participatory design campaign; (2) the first systematic ASR benchmark for Puno Quechua, evaluating state-of-the-art models and fine-tuning Whisper-base, wav2vec2-base, and XLS-R-300M, with and without continued pre-training (CPT); (3) an open release of all datasets and fine-tuned models.

This study examines the relationship between speech representations and the hierarchical structure of cognitive assessment in mild cognitive impairment. Utilizing 5,754 German neuropsychological assessment recordings, we evaluate six cognitive tasks across three score levels: task, domain, and global levels. We compare hand-crafted acoustic features with self-supervised learning (SSL) embeddings. Results show that although SSL representations generally outperform hand-crafted features at lower levels, this trend reverses for MCI classification. Furthermore, task-specific constraints influence performance: tasks with greater response freedom exhibit performance dilution as hierarchical levels increase, suggesting specialist'' representations, whereas the performance of highly structured tasks increases toward higher levels, suggesting generalist'' representations. These findings show links between task constraints and assessment hierarchy in automated clinical speech analysis.

Phonological features provide a language-general and linguistically grounded representation of speech. We present PhonoQ-2.0, a multilingual frame-level phonological feature recognizer built on self-supervised speech models. The system directly predicts a structured 22-dimensional feature vector per frame encoding manner, vowel quality, place, and voicing, instead of deriving features from phoneme outputs. To ensure phonologically coherent predictions, we introduce a manner-conditioned gating mechanism that activates valid feature groups. Evaluated across multiple languages and corpora, PhonoQ-2.0 achieves an average macro-F1 of 91.3% in-domain and 88.9% out-of-domain. Compared to a strong CTC phoneme baseline, it delivers consistent gains of +8.8 F1 in-domain and +8.6 out-of-domain on average. In unseen-language evaluation, PhonoQ-2.0 improves macro-F1 from 66.9% to 73.6% (+6.7 on average), with gains of up to +10.8 points.

Alzheimer's disease (AD) is a progressive neurodegenerative disorder and the leading cause of dementia, affecting memory, reasoning, communication, and daily functioning. Early diagnosis is particularly important, as timely intervention may help slow cognitive decline and improve patient care. Recent studies have demonstrated that spontaneous speech contains valuable linguistic and acoustic biomarkers associated with dementia. However, existing approaches often rely on independently trained modality-specific models, feature concatenation strategies, ensemble methods, or attention-based fusion mechanisms that do not explicitly maximize the dependency between speech and transcript representations. In this work, we propose a multimodal deep learning framework for automatic dementia detection that jointly exploits speech and transcript information in an end-to-end trainable manner. Specifically, speech recordings are divided into 10-second segments and passed through a pre-trained HuBERT model to extract contextualized acoustic representations. To better capture informative temporal speech characteristics, attentive statistics pooling is employed to aggregate frame-level acoustic embeddings. For the textual modality, transcripts are encoded using a pre-trained BERT model, where the [CLS] token representation is used as the linguistic embedding. The acoustic and textual representations are subsequently combined using an attention-based Audio-Text Fusion (AT-Fusion) mechanism. In addition, we introduce a MINE objective to maximize the mutual information between modalities and improve multimodal representation alignment. The fused multimodal representation is finally used for dementia classification. Experiments conducted on the publicly available ADReSS Challenge and PROCESS-2 dataset demonstrate the effectiveness and robustness of the proposed approach for speech-based dementia assessment.

Speech and audio systems operate in inherently non-stationary environments, yet continual learning (CL) research in this domain, especially in the foundation model era, remains fragmented that fail to account for the coupled, geometry-sensitive nature of acoustic representations. Modern speech foundation models operate over highly entangled, continuous representations that jointly encode linguistic, speaker, and paralinguistic factors within a shared latent space. CL is therefore fundamentally about preserving and evolving shared representation structure rather than retaining isolated task knowledge. In this work, we revisit CL for speech from a representation-centered perspective, and introduce a new taxonomy that organizes CL according to how underlying representation geometry evolves under non-stationary acoustic conditions. We further identify key mismatches between current CL assumptions and speech foundation model behavior, and finally outline a set of open challenges and future research directions.

Non-intrusive intelligibility prediction estimates how well hearing-impaired listeners understand hearing-aid-processed speech without a clean reference. We study this task in the 3rd Clarity Prediction Challenge using two frozen speech encoders, Canary and WavLM. The central question is not only whether complementary pretrained representations should be combined, but where their interaction should occur. We compare single-backbone baselines, uniform score averaging, pool-late fusion, cross-attention, frame-aligned fusion, and reverse alignment under a shared left/right-preserving binaural framework. Among the compared systems, the best model temporally prepares WavLM with a learnable strided convolution and fuses it with Canary on the coarser Canary timeline before pooling, reaching Eval RMSE 24.96$\pm$0.06 and Eval Corr 0.796$\pm$0.001. Severity, enhancement-system, layer-window, and temporal-shift analyses indicate that coarse local temporal correspondence before pooling is a useful inductive bias for this task.

Globalization and multiculturalism continue to produce increasingly diverse speech varieties. Yet current spoken dialogue systems frequently fail on under-represented dialects and accents, often misidentifying the input language and causing cascading failures in downstream dialogue tasks. Addressing this dialectal variance under low-resource constraints remains an open challenge, as standard fine-tuning is computationally expensive and prone to overfitting on high-dimensional speech data. We propose Convex Language Detection (CLD), a novel framework that integrates theoretically grounded convex optimization techniques into the spoken dialogue systems pipeline. Our method is efficiently implemented via multi-GPU Alternating Direction Method of Multipliers (ADMM) in JAX, thus providing global optimality guarantees and fast training in polynomial time. Theoretically, we prove that our convex objective induces certified margin stability and provide guarantees against feature perturbations. Empirically, we demonstrate sample efficiency and robustness to input dialectical variation, achieving 97-98% accuracy in challenging low-resource regimes. Our open-source package is available at https://pypi.org/project/jaxcld/

Speech-driven gestures and facial animations are fundamental to expressive digital avatars in games, virtual production, and interactive media. However, existing methods are either limited to a single modality for audio motion alignment, failing to fully utilize the potential of massive human motion data, or are constrained by the representation ability and throughput of multimodal models, which makes it difficult to achieve high-quality motion generation or real-time performance. We present UMo, a unified sparse motion modeling architecture for real-time co-speech avatars, which processes text, audio, and motion tokens within a unified formulation. Leveraging a spatially sparse Mixture-of-Experts framework and a temporally sparse, keyframe-centric design, UMo efficiently performs real-time dense reconstruction, enabling temporally coherent and high-fidelity animation generation for both facial expressions and gestures. Furthermore, we implement a multi-stage training strategy with targeted audio augmentation to enhance acoustic diversity and semantic consistency. Consequently, UMo preserves fine-grained speech-motion alignment even under strict latency constraints. Extensive quantitative and qualitative evaluations show that UMo achieves better output quality under low latency and real-time performance constraints, offering a practical solution for high-fidelity real-time co-speech avatars.

Deep learning-based AMC methods have achieved remarkable performance, but their practical deployment remains constrained by the high cost of labeled data. Although self-supervised learning (SSL) reduces the reliance on labels, existing SSL-based AMC methods often rely on task-agnostic pretext objectives misaligned with modulation classification, leading to representations entangled with nuisance factors such as symbol, channel, and noise. In this paper, we identify intra-instance modulation consistency as a task-aware structural prior, whereby different temporal segments of the same signal may differ in waveform while preserving the same modulation type, thus providing a principled cue for task-aligned self-supervision. Based on this prior, we propose Mod-CL, a Modulation consistency-based Contrastive Learning framework that constructs positive pairs from different temporal segments of the same signal instance, to encourage the model to learn shared modulation information while suppressing nuisance variations. We further develop a contrastive objective tailored to Mod-CL, which jointly exploits temporal segmentation and data augmentation to pull together views sharing the same modulation semantics while avoiding supervisory conflicts within each signal instance. Extensive experiments on RadioML datasets show that Mod-CL consistently outperforms strong baselines, especially in low-label regimes, achieving substantial improvements in linear probing accuracy.

High-quality, real-time talking head synthesis remains a fundamental challenge in computer vision. Existing reconstruction- and rendering-based methods typically rely on identity-specific models, limiting cross-identity generalization. To address this issue, we propose SDTalk, a one-shot 3D Gaussian Splatting (3DGS)-based framework that generalizes to unseen identities without personalized training or fine-tuning. Our framework comprises two modules with a two-stage training strategy. In the first stage, we incorporate structured facial priors into the reconstruction module and separately predict 3DGS parameters for visible and occluded regions, enabling complete head reconstruction from a single image. In the second stage, we introduce a dual-branch motion field to model coarse and fine facial dynamics, improving detail fidelity and lip synchronization. Experiments demonstrate that SDTalk surpasses existing methods in both visual quality and inference efficiency.

Audio deepfake detection systems are increasingly deployed in high-stakes security applications, yet their fairness across demographic groups remains critically underexamined. Prior work measures gender disparity but does not investigate where it comes from or how to fix it systematically. We present the first diagnosis-first framework that identifies bias source before applying targeted mitigation, evaluated on two models, AASIST and Wav2Vec2+ResNet18, on ASVSpoof5. Our diagnosis shows that bias does not stem from imbalanced training data but from acoustic representation differences, gender leakage in learned features, and structural evaluation asymmetry. We test mitigation strategies across in-processing, post-processing and combined families, including novel methods introduced in this work. Adjusting the decision threshold separately per gender reduces unfairness by 54% to 75% at no cost to detection accuracy, and our new epoch-level fairness regularisation method outperforms existing per-batch approaches. Adversarial debiasing succeeds only when gender leakage is localised, and fails when it is diffuse, an outcome correctly predicted by our diagnosis before training. No single method fully closes the fairness gap, confirming that bias sources must be identified before fixes are applied and that fairer benchmark design is equally important

Integrating speech understanding and generation is a pivotal step toward building unified speech models. However, the different representations required for these two tasks currently pose significant compatibility challenges. Typically, semantics-oriented features are learned from self-supervised learning (SSL), and acoustic-oriented features from reconstruction. Such fragmented representations hinder the realization of truly unified speech systems. We present WavCube, a compact continuous latent derived from an SSL speech encoder that simultaneously supports speech understanding, reconstruction, and generation. WavCube employs a two-stage training scheme. Stage 1 trains a semantic bottleneck to filter off-manifold redundancy that makes raw SSL features intractable for diffusion. Stage 2 injects fine-grained acoustic details via end-to-end reconstruction, while a semantic anchoring loss ensures the representation remains grounded within its original semantic manifold. Comprehensive experiments show that WavCube closely approaches WavLM performance on SUPERB despite an 8x dimensional compression, attains reconstruction quality on par with existing acoustic representations, delivers state-of-the-art zero-shot TTS performance with markedly faster training convergence, and excels in speech enhancement, separation, and voice conversion tasks on the SUPERB-SG benchmark. Systematic ablations reveal that WavCube's two-stage recipe resolves two intrinsic flaws of SSL features for generative modeling, paving the way for future unified speech systems. Codes and checkpoints are available at https://github.com/yanghaha0908/WavCube.

This paper describes a submission to the Environment-Aware Speech and Sound Deepfake Detection Challenge (ESDD2) 2026, which addresses component-level deepfake detection using the CompSpoofV2 dataset, where speech and environmental sounds may be independently manipulated. To address this challenge, a dual-branch deepfake detection framework is proposed to jointly model speech and environmental contextual representations from input audio. Two pretrained models, XLS-R for speech and BEATs for environmental sound, are used to extract complementary contextual representations. A Matching Head is introduced to model representation differences through statistical normalization and representation interaction, enabling estimation of the original class. In parallel, multi-head cross-attention enables effective information exchange between speech and environmental components. The refined representations are processed with residual connections and layer normalization, and passed to an AASIST classifier to predict speech-based and environment-based spoofing probabilities. The model outputs original, speech, and environment predictions. On the test set, the proposed system achieves an F1-score of 70.20% and an environmental EER of 16.54%, outperforming the baseline system.

Self-supervised speech models (S3Ms) achieve strong downstream performance, yet their learned representations remain poorly understood under natural and adversarial perturbations. Prior studies rely on representation similarity or global dimensionality, offering limited visibility into local geometric changes. We ask: how do perturbations deform local geometry, and do these shifts track downstream automatic speech recognition (ASR) degradation? To address this, we present GRIDS, a framework using Local Intrinsic Dimensionality (LID) across layer-wise representations in WavLM and wav2vec 2.0. We find that LID increases for all low signal-to noise ratio (SNR) perturbations and diverges at high SNR: benign noise converges toward the clean profile, while adversarial inputs retain early-layer LID elevation. We show LID elevation co-occurs with increased WER, and that layer-wise LID features enable anomaly detection (AUROC 0.78-1.00), opening the door to transcript-free monitoring in S3Ms.

Subtle hand differences make sign language recognition challenging, yet many existing methods rely on encoders pretrained on generic action datasets that poorly capture such fine-grained cues. We propose a self-supervised pretraining method for sign language recognition that uses segmentation-based masking to adapt to the presence and motion of key body parts, rather than treating hand poses as static visual tokens. The resulting mask-and-reconstruct objective improves fine-grained sign representation learning. On WLASL, NMFs-CSL, and Slovo, our encoder achieves state-of-the-art performance, improving per-instance and per-class Top-1 accuracy while using fewer input frames and modalities than comparable encoders.

Understanding how neural models represent human-interpretable concepts is challenging. Prior work has explored linear concept subspaces from diverse perspectives, such as probing and concept erasure. We introduce a unified framework to study these subspaces along two axes: \textit{containment}, which tests if a concept is fully represented in a subspace but not outside it, and \textit{disentanglement}, which tests for isolation from other concepts. In experiments on both text and speech models, we first highlight that concept subspaces may not be uniquely determined, and discuss the implications for concept subspace analysis. Then, we compare properties of concept subspaces estimated using five estimators, proposed in different communities. We find that (1) the choice of estimator impacts the containment and disentanglement properties; (2) the state-of-the-art concept erasure method, LEACE, performs well on both testing axes, but still struggles to generalize to unseen data; and (3) in HuBERT speech representations, phone information is both contained and disentangled from speaker information, while speaker information is hard to contain in a compact subspace, despite being disentangled from phones.

Recent advances in speech synthesis have shifted from phoneme representations to direct grapheme modeling. While phonemes address the one-to-many mapping between text and acoustics, they rely on grapheme-to-phoneme (G2P) systems that fail to capture speaker-specific acoustic variation. Prior work demonstrates that grapheme-based models outperform phoneme-based systems at scale, but not in low-resource settings. In this paper, we propose SPARCLE, a speaker-aware grapheme representation model that enriches characters with their precise acoustic realizations. SPARCLE is trained with a contrastive objective to align graphemes with corresponding Wav2Vec2 acoustic representations while conditioned on speaker identity. The resulting model serves as a replacement to G2P systems for downstream text-to-speech (TTS) tasks. We demonstrate that SPARCLE improves generation quality, reducing word error rates by half in extreme low-resource settings compared to standard grapheme-based models.

A speaker encoder used in multilingual voice cloning should treat the same speaker identically regardless of which script the audio was uttered in. Off-the-shelf encoders do not, and the failure is accent-conditional. On a 1043-pair Western-accented voice corpus across English, Hindi, Telugu, and Tamil, WavLM-base-plus-sv loses 0.082 absolute cosine similarity when the same voice changes script and ECAPA-TDNN loses 0.105. On a 1369-pair Indian-accented voice corpus, the gap shrinks to 0.006 (WavLM-SV) and 0.044 (ECAPA-TDNN). The leak is largest where it matters most for cross-script TTS: when a system projects a non-Indic-trained voice into Indic scripts. We present LASE (Language-Adversarial Speaker Encoder), a small projection head over frozen WavLM-base-plus trained with two losses: a supervised contrastive loss over voice identity, and a gradient-reversal cross-entropy against a 4-language classifier that pushes the embedding to be language-uninformative while remaining speaker-informative. Trained on 1118 quality-gated cross-script pairs synthesised from 8 commercial multilingual voices, LASE's residual gap is consistent with zero on both corpora (Delta = 0.013 Western, Delta = 0.026 Indian; both bootstrap 95% CIs include zero) and amplifies the cross-script-vs-floor margin 2.4-2.7x over both baselines. An ECAPA+GRL ablation shows the GRL objective improves either backbone but the WavLM choice contributes too. In synthetic multi-speaker diarisation, LASE matches ECAPA-TDNN on cross-script speaker recall (0.788 vs 0.789) with ~100x less training data. We release the r1 checkpoint, both corpora, and the bootstrap recipe.

Supervised contrastive learning (SupCon) is widely used to shape representations, but has seen limited targeted study for audio deepfake detection. Existing work typically combines contrastive terms with broader pipelines; however, the focus on SupCon itself is missing. In this work, we run a controlled study on wav2vec2 XLS-R (300M) that varies (i) similarity in SupCon (cosine vs angular similarity derived from the hyperspherical angle) and (ii) negative scaling using a warm-started global cross-batch queue. Stage 1 fine-tunes the encoder and projection head with SupCon; Stage 2 freezes them and trains a linear classifier with BCE. Trained on ASVspoof 2019 LA and evaluated on ASV19 eval plus ITW and ASVspoof 2021 DF/LA, Cosine SupCon with a delayed queue achieves the best ITW EER (8.29%) and pooled EER (4.44), while angular similarity performs strongly without queued negatives (ITW 8.70), indicating reduced reliance on large negative sets.

Human-imitated speech poses a greater challenge than AI-generated speech for both human listeners and automatic detection systems. Unlike AI-generated speech, which often contains artifacts, over-smoothed spectra, or robotic cues, imitated speech is produced naturally by humans, thereby preserving a higher degree of naturalness that makes imitation-based speech forgery significantly more challenging to detect using conventional acoustic or cepstral features. To overcome this challenge, this study proposes an auditory perception-based Spectro-Temporal Modulation (STM) representation framework for human-imitated speech detection. The STM representations are derived from two cochlear filterbank models: the Gammatone Filterbank (GTFB), which simulates frequency selectivity and can be regarded as a first approximation of cochlear filtering, and the Gammachirp Filterbank (GCFB), which further models both frequency selectivity and level-dependent asymmetry. These STM representations jointly capture temporal and spectral fluctuations in speech signals, corresponding to changes over time in the spectrogram and variations along the frequency axis related to human auditory perception. We also introduce a Segmental-STM representation to analyze short-term modulation patterns across overlapping time windows, enabling high-resolution modeling of temporal speech variations. Experimental results show that STM representations are effective for human-imitated speech detection, achieving accuracy levels close to those of human listeners. In addition, Segmental-STM representations are more effective, surpassing human perceptual performance. The findings demonstrate that perceptually inspired spectro-temporal modeling is promising for detecting imitation-based speech attacks and improving voice authentication robustness.

Modern automatic speech recognition (ASR) systems have been observed to function better for certain speaker groups (SGs) than others, despite recent gains in overall performance. One potential impediment to progress towards fairer ASR is a more nuanced understanding of the types of modeling errors that speech encoder models make, and in particular the difference between the structure of embeddings for high-performance and low-performance SGs. This paper proposes a framework typifying two types of error that can occur in modeling phonemes in ASR systems: random error/high variance in phoneme embedding, vs systematic error/embedding bias. We find that training phoneme classification probes only on a single, typically disadvantaged SG, sometimes improves performance for that SG, which is evidence for the existence of SG-level bias in phoneme embeddings. On the other hand, we find that speakers and SGs with higher levels of phoneme variance are the same as those with worse phoneme prediction accuracy. We conclude that both types of error are present in phoneme embeddings and both are candidate causes for SG-level unfairness in ASR, though random error is likely a greater hindrance to fairness than systematic error. Furthermore, we find that finetuning encoder models using a fairness-enhancing algorithm (domain enhancing and adversarial training) changes neither the benefits of in-domain phoneme classification probe training, nor measured levels of random embedding error.

Using self-supervised learning (SSL) models has significantly improved performance for downstream speech tasks, surpassing the capabilities of traditional hand-crafted features. This study investigates the amalgamation of SSL models, with the aim to leverage both their individual strengths and refine extracted features to achieve improved speech recognition models for naturalistic scenarios. Our research investigates the massive naturalistic Fearless Steps (FS) APOLLO resource, with particular focus on the FS Challenge (FSC) Phase-4 corpus, providing the inaugural analysis of this dataset. Additionally, we incorporate the CHiME-6 dataset to evaluate performance across diverse naturalistic speech scenarios. While exploring previously proposed Feature Refinement Loss and fusion methods, we found these methods to be less effective on the FSC Phase-4 corpus. To address this, we introduce a novel deep cross-attention (DCA) fusion method, designed to elevate performance, especially for the FSC Phase-4 corpus. Our objective is to foster creation of superior FS APOLLO community resources, catering to the diverse needs of researchers across various disciplines. The proposed solution achieves an absolute +1.1% improvement in WER, providing effective meta-data creation for the massive FS APOLLO community resource.

Wav2vec 2.0 (W2V2) has shown strong performance in pathological speech analysis by effectively capturing the characteristics of atypical speech. Despite its success, it remains unclear which components of its learned representations are most informative for specific downstream tasks. In this study, we address this question by investigating the regression of dysarthric speech descriptors using annotations from the Speech Accessibility Project dataset. We focus on five descriptors, each addressing a different aspect of speech or voice production: intelligibility, imprecise consonants, inappropriate silences, harsh voice and monoloudness. Speech representations are derived from a W2V2-based feature extractor, and we systematically compare layer-wise and time-wise aggregation strategies using attentive statistics pooling. Our results show that intelligibility is best captured through layer-wise representations, whereas imprecise consonants, harsh voice and monoloudness benefit from time-wise modeling. For inappropriate silences, no clear advantage could be observed for either approach.

Speaker diarization (SD) is the task of answering "who spoke when" in a multi-speaker audio stream. Classically, an SD system clusters segments of speech belonging to an individual speaker's identity. Recent years have seen substantial progress in SD through end-to-end neural diarization (EEND) approaches. DiariZen, a hybrid SD pipeline built upon a structurally pruned WavLM-Large encoder, a Conformer backend with powerset classification, and VBx clustering, represents the leading open-source state of the art at the time of writing across multiple benchmarks. Despite its strong performance, the DiariZen architecture spans several repositories and frameworks, making it difficult for researchers and practitioners to understand, reproduce, or extend the system as a whole. This tutorial paper provides a self-contained, block-by-block explanation of the complete DiariZen pipeline, decomposing it into seven stages: (1) audio loading and sliding window segmentation, (2) WavLM feature extraction with learned layer weighting, (3) Conformer backend and powerset classification, (4) segmentation aggregation via overlap-add, (5) speaker embedding extraction with overlap exclusion, (6) VBx clustering with PLDA scoring, and (7) reconstruction and RTTM output. For each block, we provide the conceptual motivation, source code references, intermediate tensor shapes, and annotated visualizations of the actual outputs on a 30s excerpt from the AMI Meeting Corpus. The implementation is available at https://github.com/nikhilraghav29/diarizen-tutorial, which includes standalone executable scripts for each block and a Jupyter notebook that runs the complete pipeline end-to-end.

Speech-preserving facial expression manipulation (SPFEM) aims to modify facial emotions while meticulously maintaining the mouth animation associated with spoken content. Current works depend on inaccessible paired training samples for the person, where two aligned frames exhibit the same speech content yet differ in emotional expression, limiting the SPFEM applications in real-world scenarios. In this work, we discover that speakers who convey the same content with different emotions exhibit highly correlated local facial animations in both spatial and temporal spaces, providing valuable supervision for SPFEM. To capitalize on this insight, we propose a novel spatial-temporal coherent correlation learning (STCCL) algorithm, which models the aforementioned correlations as explicit metrics and integrates the metrics to supervise manipulating facial expression and meanwhile better preserving the facial animation of spoken content. To this end, it first learns a spatial coherent correlation metric, ensuring that the visual correlations of adjacent local regions within an image linked to a specific emotion closely resemble those of corresponding regions in an image linked to a different emotion. Simultaneously, it develops a temporal coherent correlation metric, ensuring that the visual correlations of specific regions across adjacent image frames associated with one emotion are similar to those in the corresponding regions of frames associated with another emotion. Recognizing that visual correlations are not uniform across all regions, we have also crafted a correlation-aware adaptive strategy that prioritizes regions that present greater challenges. During SPFEM model training, we construct the spatial-temporal coherent correlation metric between corresponding local regions of the input and output image frames as an additional loss to supervise the generation process.

Speech encoder models are known to model members of some speaker groups (SGs) better than others. However, there has been little work in establishing why this occurs on a technological level. To our knowledge, we present the first layerwise fairness analysis of pretrained self-supervised speech encoder models (S3Ms), probing each embedding layer for speaker identification (SID) automatic speech recognition (ASR). We find S3Ms produce embeddings biased against certain SGs for both tasks, starting at the very first latent layers. Furthermore, we find opposite patterns of layerwise bias for SID vs ASR for all models in our study: SID bias is minimized in layers that minimize overall SID error; on the other hand, ASR bias is maximized in layers that minimize overall ASR error. The inverse bias/error relationship for ASR is unaffected when probing S3Ms that are finetuned for ASR, suggesting SG-level bias is established during pretraining and is difficult to remove.

We present a phoneme-level analysis of automatic speech recognition (ASR) for two low-resourced and phonologically complex East Caucasian languages, Archi and Rutul, based on curated and standardized speech-transcript resources totaling approximately 50 minutes and 1 hour 20 minutes of audio, respectively. Existing recordings and transcriptions are consolidated and processed into a form suitable for ASR training and evaluation. We evaluate several state-of-the-art audio and audio-language models, including wav2vec2, Whisper, and Qwen2-Audio. For wav2vec2, we introduce a language-specific phoneme vocabulary with heuristic output-layer initialization, which yields consistent improvements and achieves performance comparable to or exceeding Whisper in these extremely low-resource settings. Beyond standard word and character error rates, we conduct a detailed phoneme-level error analysis. We find that phoneme recognition accuracy strongly correlates with training frequency, exhibiting a characteristic sigmoid-shaped learning curve. For Archi, this relationship partially breaks for Whisper, pointing to model-specific generalization effects beyond what is predicted by training frequency. Overall, our results indicate that many errors attributed to phonological complexity are better explained by data scarcity. These findings demonstrate the value of phoneme-level evaluation for understanding ASR behavior in low-resource, typologically complex languages.

Existing lip-sync deepfake detectors rely on pixel artifacts or audio-visual correspondence, and both fail under generator or language shift because the features they learn are tied to the training distribution. We take a different approach. Authentic lip motion is constrained by tissue mechanics and neuromuscular bandwidth; current generators typically do not impose these constraints, producing trajectories with elevated variance in velocity, acceleration, and jerk that real speech does not exhibit. We exploit this signal, which we term temporal lip jitter, by computing kinematic statistics from 64 perioral landmarks over short sliding windows and feeding them into a lightweight three-branch network. The model uses only landmark coordinates: no pixels, no audio, and no voiceprint data. We train only on English data and test in a zero-shot setting on five unseen generators and seven languages.

We introduce Echoes, a new dataset for music deepfake detection designed for training and benchmarking detectors under realistic and provider-diverse conditions. Echoes comprises 4,468 tracks (131 hours of audio) spanning multiple genres (pop, rock, electronic), and includes content generated by ten popular AI music generation systems. To prevent shortcut learning and promote robust generalization, the dataset is deliberately constructed to be challenging, enforcing semantic-level alignment between spoofed audio and bona fide references. This alignment is achieved by conditioning generated audio samples directly on bona-fide waveforms or song descriptors. We evaluate Echoes in a cross-dataset setting against three existing AI-generated music datasets using state-of-the-art Wav2Vec2 XLS-R 2B representations. Results show that (i) Echoes is the hardest in-domain dataset; (ii) detectors trained on existing datasets transfer poorly to Echoes; (iii) training on Echoes yields the strongest generalization performance. These findings suggest that provider diversity and semantic alignment help learn more transferable detection cues.

Sign language production requires more than hand motion generation. Non-manual features, including mouthings, eyebrow raises, gaze, and head movements, are grammatically obligatory and cannot be recovered from manual articulators alone. Existing 3D production systems face two barriers to integrating them: the body-model fittings they train on retain a facial space too low-dimensional to encode these articulations, and, if richer representations are adopted, standard discrete tokenization suffers from codebook collapse, leaving most of the expression space largely unreachable. We propose SMPL-FX, which couples FLAME's rich expression space with the SMPL-X body parameterisation, and tokenize the resulting representation with modality-specific Finite Scalar Quantization VAEs for body, hands, and face, raising face codebook utilization from 78.8% to 99.0%. M3T is an autoregressive transformer trained on this multi-modal motion vocabulary, with an auxiliary sign-to-text translation objective that encourages semantically grounded embeddings. Across three standard datasets, M3T achieves state-of-the-art sign language production quality, and on NMFs-CSL, where signs are distinguishable only by non-manual features, our model reaches 58.3% accuracy against 49.0% for the strongest comparable baseline, without large-scale sign-language pre-training. Project page: https://cogvis-cvssp.github.io/papers/m3t/

The limited availability of dysarthric speech data makes cross-lingual detection an important but challenging problem. A key difficulty is that speech representations often encode language-dependent structure that can confound dysarthria detection. We propose a representation-level language shift (LS) that aligns source-language self-supervised speech representations with the target-language distribution using centroid-based vector adaptation estimated from healthy-control speech. We evaluate the approach on oral DDK recordings from Parkinson's disease speech datasets in Czech, German, and Spanish under both cross-lingual and multilingual settings. LS substantially improves sensitivity and F1 in cross-lingual settings, while yielding smaller but consistent gains in multilingual settings. Representation analysis further shows that LS reduces language identity in the embedding space, supporting the interpretation that LS removes language-dependent structure.

Tonal low-resource languages are widely spoken but remain underserved by modern speech technologies. A central challenge is learning speech representations that are robust to nuisance variation, such as speaker gender, while preserving lexical tone, which carries word meaning. We propose SITA, a lightweight adaptation recipe for pretrained wav2vec-style self-supervised speech encoders. Rather than designing a new backbone or objective, SITA combines existing objectives in a staged optimization framework to reduce tone collapse while preserving ASR capability. Stage 1 improves speaker invariance without erasing tonal contrasts by combining a cross-gender contrastive loss with a tone-repulsive loss that separates same-word, different-tone realizations. Stage 2 restores recognition-oriented linguistic information through CTC fine-tuning and knowledge distillation on upper encoder layers. We evaluate SITA primarily on Hmong, a tonal language with limited digital resources and a small speaker pool. Against multilingual, speaker-adversarial, label-aware, and semi-supervised baselines, SITA achieves the best trade-off between cross-gender lexical retrieval and tone separation, while maintaining ASR accuracy close to an ASR-adapted XLS-R teacher. Results on Mandarin show consistent gains, suggesting that SITA is a general plug-in recipe for tonal speech representation learning.

This paper introduces MauBERT, a multilingual extension of HuBERT that leverages articulatory features for robust cross-lingual phonetic representation learning. We continue HuBERT pre-training with supervision based on a phonetic-to-articulatory feature mapping in 55 languages. Our models learn from multilingual data to predict articulatory features or phones, resulting in language-independent representations that capture multilingual phonetic properties. Through comprehensive ABX discriminability testing, we show MauBERT models produce more context-invariant representations than state-of-the-art multilingual self-supervised learning models. Additionally, the models effectively adapt to unseen languages and casual speech with minimal self-supervised fine-tuning (10 hours of speech). This establishes an effective approach for instilling linguistic inductive biases in self-supervised speech models.

Child-centered daylong recordings are essential for studying early language development, but existing speech models trained on clean adult data perform poorly due to acoustic and linguistic differences. We introduce BabyHuBERT, a self-supervised speech model trained on 13,000 hours of multilingual child-centered recordings from 40+ languages. Evaluated on voice type classification, the task of identifying who produces speech and when in child-centered recordings (key child, other children, male, and female adults), BabyHuBERT-VTC achieves F1-scores from 55.0% to 76.1% across six corpora, consistently outperforming W2V2-LL4300 and HuBERT (pretrained on English daylongs and clean adult speech, respectively). Notable gains include 14.0 and 18.3 absolute F1 points over HuBERT on Vanuatu and Solomon Islands, demonstrating effectiveness on underrepresented languages. We share code and models to support researchers working with child-centered recordings across diverse linguistic contexts.

A recent line of research on automated speaking assessment (ASA) has benefited from self-supervised learning (SSL) representations, which capture rich acoustic and linguistic patterns in non-native speech without underlying assumptions of feature curation. However, speech-based SSL models capture acoustic-related traits but overlook linguistic content, while text-based SSL models rely on ASR output and fail to encode prosodic nuances. Moreover, most prior arts treat proficiency levels as nominal classes, ignoring their ordinal structure and non-uniform intervals between proficiency labels. To address these limitations, we propose an effective ASA approach combining SSL with handcrafted indicator features via a novel modeling paradigm. We further introduce a multi-margin ordinal loss that jointly models both the score ordinality and non-uniform intervals of proficiency labels. Extensive experiments on the TEEMI corpus show that our method consistently outperforms strong baselines and generalizes well to unseen prompts.

A recent line of research on spoken language assessment (SLA) employs neural models such as BERT and wav2vec 2.0 (W2V) to evaluate speaking proficiency across linguistic and acoustic modalities. Although both models effectively capture features relevant to oral competence, each exhibits modality-specific limitations. BERT-based methods rely on ASR transcripts, which often fail to capture prosodic and phonetic cues for SLA. In contrast, W2V-based methods excel at modeling acoustic features but lack semantic interpretability. To overcome these limitations, we propose a system that integrates W2V with Phi-4 multimodal large language model (MLLM) through a score fusion strategy. The proposed system achieves a root mean square error (RMSE) of 0.375 on the official test set of the Speak & Improve Challenge 2025, securing second place in the competition. For comparison, the RMSEs of the top-ranked, third-ranked, and official baseline systems are 0.364, 0.384, and 0.444, respectively.