arXiv subjects

Mengzhe Geng

Publications and source records attributed to Mengzhe Geng.

At least 19 recordsLinked to original sources

arXiv2026

Differentiable and Severity-invariant Discrete Tokens for Dysarthric Speech Recognition

This paper proposes novel differentiable and severity-invariant (DSI) discrete token approaches that are not only tightly integrated with downstream dysarthric speech recognition tasks, but also minimise discrete token diversity across speech impairment severity groups. Experiments conducted on the UASpeech and TORGO corpora suggest that Conformer models trained using the DSI tokens outperform the comparable baseline HuBERT discrete/continuous features by statistically significant WER reductions of 2.22\%/0.78\% absolute (9.14\%/3.41\% relative) and 1.78\%/1.06\% absolute (18.43\%/11.86\% relative) on the two tasks, respectively. After system combination, the lowest WERs of 18.90\% and 6.38\% were obtained on UASpeech and TORGO. Phoneme-specific T-SNE visualizations show that severity-invariant regularization reduces severity-dependent variation by producing greater overlap and less distinct boundaries among severity-group distributions.

eess.AS↗

arXiv2026

SURE-Voice: A Front-End Baseline for Speech-Evidence Filtering in Speech LLMs

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

eess.AS↗

arXiv2026

Auditing Generative Audio Calls for Known-Task Audio-LLM Evaluatio

Speech and audio LLMs are often evaluated by asking whether a waveform prompt beats an automatic speech recognition (ASR) transcript. For known closed-set tasks, that comparison conflates two factors: access to acoustic evidence and the need to call a generative audio model. We evaluate this distinction as a controlled call-decision problem. For each example, a policy chooses among keeping a transcript label, using encoder evidence from Contrastive Language-Audio Pretraining (CLAP), Audio Spectrogram Transformer (AST), or WavLM, and calling Qwen2-Audio, Qwen2.5-Omni, or MOSS-Audio; the decisive ablation removes all generative actions while keeping the selector and development protocol fixed. On VocalSound, the transcript-only representation reaches 0.296 accuracy, so it is insufficient for this label set. Yet supervised CLAP and WavLM controls reach 0.850 and 0.854 with no generative audio calls. A selector with generative actions reaches 0.925 accuracy using 12.5% calls, compared with 0.921 for the matched no-call selector (paired difference 0.004; 95% CI [-0.025, 0.033], including zero). Agreement and stacking features improve weaker selectors but do not beat the strongest no-call control. For known-task endpoint-value statements, the relevant quantity is the marginal value of the generative call after transcript and encoder evidence have already been used.

cs.SD↗

arXiv2026

From Scores to Evidence: Auditable Decisions Can Improve Speech Deepfake Detection

Speech deepfakes can mimic a speaker's voice convincingly enough to deceive listeners and automated systems. This has driven strong progress in speech deepfake detection, but most detectors still end with one score per utterance. That score is useful for ranking systems, yet it says little about why a borderline item should be trusted, deferred, or reviewed. Two utterances can fall in the same score band for different reasons, for example because passive and retrieval evidence disagree or because the keyed probe is unavailable. We ask whether the final decision can remain scalar without discarding that provenance. We answer this question with an auditable decision record that carries four aligned cues into a late calibration step: a passive detector score, a conditional keyed-probe score on a marked derivative, retrieval support, and a speaker-profile margin, together with explicit disagreement coordinates. On the 4,080-example ASVspoof 5 Track 1 matched subset, the fixed retrieval-augmented rule improves on retrieval-only evidence, reducing equal error rate (EER) from 15.84% to 11.91%, and late calibration over the full record reaches 8.43% EER. At a 33.75% review budget, the exposed cue union covers 82.85% of the calibrated model's errors. The best passive WavLM run still reaches 6.71% EER, so we do not present the decision record as a stronger standalone detector. Its contribution is to preserve the evidence behind each surfaced utterance while still producing one operating score for thresholding and review.

cs.SD↗

arXiv2026

VoxReason: Listener-Free Evaluation of Source-Grounded Speech Planning Before Synthesis

Expressive speech systems have to decide how an utterance is delivered before any waveform is rendered. In dialogue agents, narration, and role-conditioned TTS, that planning step sets affect, pitch, energy, rate, pause, emphasis, and stance, yet standard audio metrics rarely show whether those choices were actually licensed by the source record. This leaves a practical evaluation gap: a system may sound plausible while relying on a memorized script instead of the cue that governs delivery. VoxReason casts this pre-synthesis step as a listener-free task for source-grounded speech planning. Systems output a source-cited speaking plan, and a deterministic verifier checks citation legality, slot agreement, unsupported state, schema validity, and one-cue counterfactual locality. On 1,440 checked source-label cases, shortcut controls show why slot accuracy alone is unsafe: a key-lookup oracle reaches 1.000 plan-slot accuracy on seen keys, while an emotion prior still reaches 0.958 slot accuracy on source-key-disjoint cases without citing intensity or identity. In a separate 100-case learned source-key-disjoint comparison, a 7B locality SFT+CF repair improves plan-slot accuracy/locality from 0.684/0.141 to 0.919/1.000, and removing source records lowers citation-required grounded score by 0.488. The resulting benchmark isolates whether planned delivery is warranted by the record before waveform evaluation or listener studies are used.

cs.SD↗

arXiv2026

AudioNoisePrints: Model-free audio watermarking using spatial correlation in flow matching TTS

We present AudioNoisePrints, a training-free watermarking pipeline for flow matching and diffusion TTS models, which requires minimal extra computation during inference and does not require retraining the TTS model or reducing the generation quality. We exploited the fact that there are strong correlations between the initial Gaussian noises and the generated outputs in diffusion and flow matching models, such that a simple cosine correlation between the initial noise and the generated output can be used to perform watermaking. Moreover, we train a lightweight detector on top for more aggressive augmentations. Our method outperforms AudioSeal, a strong baseline for audio watermarking under strong augmentations. We experimented on F5TTS and other TTS and vocoder models, and concluded that they all exhibit similar spatial correlation properties, suggesting our watermarking scheme can be used for more flow-matching TTS models and even vocoders in the future.

cs.SD↗

arXiv2026

Decoding while Adapting: Zero-Shot Online Speaker Adaptation via Audio-Textual Prompts for Elderly Speech Recognition

This paper proposes a novel cross-utterance audio-textual prompts based speaker adaptation approach for elderly speech recognition. It enables zero-shot, real-time adaptation to unseen speakers. Speech and text embeddings are extracted from the current and a few preceding utterances, before being fused in a cross-modal manner to produce compact speaker prompts that are more consistent than i/x-vectors and ECAPA-TDNN features. Experiments on the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets suggest that the proposed online adaptation outperforms the speaker-independent (SI) model by statistically significant word error rate (WER) or character error rate (CER) reductions of 0.61% and 1.22% absolute (2.99% and 4.48% relative). Real-time factor (RTF) speed-up ratios of up to 9.83 times are obtained over offline batch-mode adaptation.

eess.AS↗

arXiv2026

Confidence Score Guided Incremental and Speaker Adaptive Pseudo-Labeling for Semi-Supervised Elderly Speech Recognition

This paper proposes a novel confidence score guided incremental and speaker adaptive pseudo-labeling approach for semi-supervised elderly speech recognition. It facilitates higher-quality pseudo-label selection and progressive refinement, while also mitigating speaker heterogeneity. A confidence estimation module is designed to rank the reliability of untranscribed data, enabling a curriculum learning trajectory that progressively folds in unlabeled data subsets from high to low confidence. Speaker-specific characteristics are captured through speaker adaptive training with learnable prompts. Experiments on the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets suggest that the proposed method outperforms the semi-supervised baseline using no confidence scores guided incremental or speaker adaptive pseudo-labeling by statistically significant word error rate (WER) or character error rate (CER) reductions of 1.45% and 2.27% absolute (6.21% and 6.98% relative).

eess.AS↗

arXiv2026

Explainable and Trustworthy Speech Emotion Recognition Using Confidence Score and Reinforcement Learning Rectified Speech Emotion Descriptors

Explainable and trustworthy speech emotion recognition (SER) remains a challenging task to date, largely due to the scarcity of SER data with reliable speech emotion descriptor (SED) labels, such as prosodic features and speaker traits. This paper presents a confidence score and reinforcement learning (RL) based on-the-fly SED rectification approach for post-training SER systems on automatically annotated SED labels. Experiments on IEMOCAP and MELD suggest that explainable SER systems incorporating the proposed confidence score and RL-based SED rectification approach consistently outperform baselines without data selection or SED rectification. The best performing system, which integrates both components, surpasses the baseline without data selection and SED rectification, achieving SER gains of 2.9% and 3.3% absolute (3.7% and 5.4% relative) on IEMOCAP and MELD benchmarks, respectively.

cs.SD↗

arXiv2026

Towards Data-free and Training-free Compression for Speech Foundation Models Using Parameter Clustering

This paper presents a novel data-free and training-free compression approach for speech foundation models using channelwise clustering via k-means. More fine-grained, mixed sparsity pruning by layer-level varying number of parameter clusters is also explored. Experiments conducted on the LibriSpeech dataset suggest that when operating with pruning sparsity of 50% on HuBERT-large, consistent WER reductions of 27.73%/18.61% absolute (34.37%/21.91% relative) over the magnitude-based pruning were obtained on the test-clean and test-other subsets before fine-tuning and 0.19%/0.79% absolute (3.36%/4.62% relative) after fine-tuning with only 3 epochs. Similar WER reductions of 2.86%/5.02% absolute (59.21%/55.29% relative) were observed against magnitudebased pruning on Whisper-large-v3 at 10% sparsity, all with no significant WER increase relative to the uncompressed baseline.

cs.SD↗

arXiv2026

Towards Personalized Federated Learning for Dysarthric Speech Recognition

Speech recognition is challenging for dysarthric speakers. While federated learning (FL)-based ASR can be an effective tool for protecting privacy, it suffers from heterogeneity issues caused by speaker variability. Forcing all speakers to share the same model components can be suboptimal under such heterogeneity, making personalization a promising direction; however, related research on dysarthric speech remains limited. To this end, this paper explores two aggregation strategies to achieve personalization, including the parameter-based averaging strategy and the embedding-based averaging strategy. Experiments on UASpeech and TORGO show that the proposed methods outperform the baseline regularized FedAvg by statistically significant WER reductions of up to 0.99% absolute (3.15% relative) on UASpeech and 0.56% absolute (4.73% relative) on TORGO, respectively.

cs.SD↗

arXiv2026

Multi-Channel Speech Enhancement for Cocktail Party Speech Emotion Recognition

This paper highlights the critical importance of multi-channel speech enhancement (MCSE) for speech emotion recognition (ER) in cocktail party scenarios. A multi-channel speech dereverberation and separation front-end integrating DNN-WPE and mask-based MVDR is used to extract the target speaker's speech from the mixture speech, before being fed into the downstream ER back-end using HuBERT- and ViT-based speech and visual features. Experiments on mixture speech constructed using the IEMOCAP and MSP-FACE datasets suggest the MCSE output consistently outperforms domain fine-tuned single-channel speech representations produced by: a) Conformer-based metric GANs; and b) WavLM SSL features with optional SE-ER dual task fine-tuning. Statistically significant increases in weighted, unweighted accuracy and F1 measures by up to 9.5%, 8.5% and 9.1% absolute (17.1%, 14.7% and 16.0% relative) are obtained over the above single-channel baselines. The generalization of IEMOCAP trained MCSE front-ends are also shown when being zero-shot applied to out-of-domain MSP-FACE data.

cs.SD↗

arXiv2025

Towards Effective and Efficient Non-autoregressive decoders for Conformer and LLM-based ASR using Block-based Attention Mask

Automatic speech recognition (ASR) systems often rely on autoregressive (AR) Transformer decoder architectures, which limit efficient inference parallelization due to their sequential nature. To this end, non-autoregressive (NAR) approaches aim primarily to achieve significant decoding speedup while the maintaining recognition accuracy that is comparable to AR baselines. This paper proposes a novel NAR block-based attention mask decoder (AMD) that effectively improves decoding efficiency while maintaining ASR accuracy, and also offering flexibility in balancing the performance-efficiency trade-off on both Conformer and large language model (LLM)-based ASR systems. The proposed AMD performs parallel inference within contiguous blocks of output labels while maintaining monotonic left-to-right prediction between blocks. A one-pass beam search algorithm is designed to dynamically fuse Connectionist Temporal Classification (CTC), AR decoder, and AMD probabilities. Experiments are conducted on normal speech LS960 and DBank elderly speech across: a) The Conformer encoder-decoder ASR system with filterbank input features; b) its integration with WavLM features; and c) further advancement by integrating an LLM-based decoder. On the LS960 task, the proposed AMD empowered tripartite decoder achieves decoding speedup ratios of up to 1.44x, 1.55x, and 2.31x under the three model configurations over the CTC + AR baselines, without statistically significant WER increases. When operating with real-time factors (RTFs) comparable to the baselines, the tripartite decoder produces statistically significant WER reductions of 0.19%, 0.62% and 0.13% absolute (4.3%, 16.3%, and 3.8% relative). Similar improvements are also obtained on the DBank task.

eess.AS↗

arXiv2025

Exploring Cross-Utterance Speech Contexts for Conformer-Transducer Speech Recognition Systems

This paper investigates four types of cross-utterance speech contexts modeling approaches for streaming and non-streaming Conformer-Transformer (C-T) ASR systems: i) input audio feature concatenation; ii) cross-utterance Encoder embedding concatenation; iii) cross-utterance Encoder embedding pooling projection; or iv) a novel chunk-based approach applied to C-T models for the first time. An efficient batch-training scheme is proposed for contextual C-Ts that uses spliced speech utterances within each minibatch to minimize the synchronization overhead while preserving the sequential order of cross-utterance speech contexts. Experiments are conducted on four benchmark speech datasets across three languages: the English GigaSpeech and Mandarin Wenetspeech corpora used in contextual C-T models pre-training; and the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets used in domain fine-tuning. The best performing contextual C-T systems consistently outperform their respective baselines using no cross-utterance speech contexts in pre-training and fine-tuning stages with statistically significant average word error rate (WER) or character error rate (CER) reductions up to 0.9%, 1.1%, 0.51%, and 0.98% absolute (6.0%, 5.4%, 2.0%, and 3.4% relative) on the four tasks respectively. Their performance competitiveness against Wav2vec2.0-Conformer, XLSR-128, and Whisper models highlights the potential benefit of incorporating cross-utterance speech contexts into current speech foundation models.

eess.AS↗

arXiv2025

Supporting SENCOTEN Language Documentation Efforts with Automatic Speech Recognition

The SENCOTEN language, spoken on the Saanich peninsula of southern Vancouver Island, is in the midst of vigorous language revitalization efforts to turn the tide of language loss as a result of colonial language policies. To support these on-the-ground efforts, the community is turning to digital technology. Automatic Speech Recognition (ASR) technology holds great promise for accelerating language documentation and the creation of educational resources. However, developing ASR systems for SENCOTEN is challenging due to limited data and significant vocabulary variation from its polysynthetic structure and stress-driven metathesis. To address these challenges, we propose an ASR-driven documentation pipeline that leverages augmented speech data from a text-to-speech (TTS) system and cross-lingual transfer learning with Speech Foundation Models (SFMs). An n-gram language model is also incorporated via shallow fusion or n-best restoring to maximize the use of available data. Experiments on the SENCOTEN dataset show a word error rate (WER) of 19.34% and a character error rate (CER) of 5.09% on the test set with a 57.02% out-of-vocabulary (OOV) rate. After filtering minor cedilla-related errors, WER improves to 14.32% (26.48% on unseen words) and CER to 3.45%, demonstrating the potential of our ASR-driven pipeline to support SENCOTEN language documentation.

cs.SD↗

arXiv2025

Regularized Federated Learning for Privacy-Preserving Dysarthric and Elderly Speech Recognition

Accurate recognition of dysarthric and elderly speech remains challenging to date. While privacy concerns have driven a shift from centralized approaches to federated learning (FL) to ensure data confidentiality, this further exacerbates the challenges of data scarcity, imbalanced data distribution and speaker heterogeneity. To this end, this paper conducts a systematic investigation of regularized FL techniques for privacy-preserving dysarthric and elderly speech recognition, addressing different levels of the FL process by 1) parameter-based, 2) embedding-based and 3) novel loss-based regularization. Experiments on the benchmark UASpeech dysarthric and DementiaBank Pitt elderly speech corpora suggest that regularized FL systems consistently outperform the baseline FedAvg system by statistically significant WER reductions of up to 0.55\% absolute (2.13\% relative). Further increasing communication frequency to one exchange per batch approaches centralized training performance.

eess.AS↗

arXiv2025

MOPSA: Mixture of Prompt-Experts Based Speaker Adaptation for Elderly Speech Recognition

This paper proposes a novel Mixture of Prompt-Experts based Speaker Adaptation approach (MOPSA) for elderly speech recognition. It allows zero-shot, real-time adaptation to unseen speakers, and leverages domain knowledge tailored to elderly speakers. Top-K most distinctive speaker prompt clusters derived using K-means serve as experts. A router network is trained to dynamically combine clustered prompt-experts. Acoustic and language level variability among elderly speakers are modelled using separate encoder and decoder prompts for Whisper. Experiments on the English DementiaBank Pitt and Cantonese JCCOCC MoCA elderly speech datasets suggest that online MOPSA adaptation outperforms the speaker-independent (SI) model by statistically significant word error rate (WER) or character error rate (CER) reductions of 0.86% and 1.47% absolute (4.21% and 5.40% relative). Real-time factor (RTF) speed-up ratios of up to 16.12 times are obtained over offline batch-mode adaptation.

eess.AS↗

arXiv2025

Towards LLM-Empowered Fine-Grained Speech Descriptors for Explainable Emotion Recognition

This paper presents a novel end-to-end LLM-empowered explainable speech emotion recognition (SER) approach. Fine-grained speech emotion descriptor (SED) features, e.g., pitch, tone and emphasis, are disentangled from HuBERT SSL representations via alternating LLM fine-tuning to joint SER-SED prediction and ASR tasks. VAE compressed HuBERT features derived via Information Bottleneck (IB) are used to adjust feature granularity. Experiments on the IEMOCAP and MELD benchmarks demonstrate that our approach consistently outperforms comparable LLaMA-based SER baselines, including those using either (a) alternating multi-task fine-tuning alone or (b) feature disentanglement only. Statistically significant increase of SER unweighted accuracy by up to 4.0% and 3.7% absolute (5.4% and 6.6% relative) are obtained. More importantly, emotion descriptors offer further explainability for SER.

cs.SD↗