Privacy-Preserving End-to-End Full-Duplex Speech Dialogue Models

End-to-end full-duplex speech models feed user audio through an always-on LLM backbone, yet the speaker privacy implications of their hidden representations remain unexamined. Following the VoicePrivacy 2024 protocol with a lazy-informed attacker, we show that the hidden states of SALM-Duplex and Moshi leak substantial speaker identity across all transformer layers. Layer-wise and turn-wise analyses reveal that leakage persists across all layers, with SALM-Duplex showing stronger leakage in early layers while Moshi leaks uniformly, and that Linkability rises sharply within the first few turns. We propose two streaming anonymization setups using Stream-Voice-Anon: a waveform-level front-end (Anon-W2W) and a feature-domain replacement (Anon-W2F). Anon-W2F raises EER by over 3.5x relative to the discrete encoder baseline (11.2% to 41.0%), approaching the 50% random-chance ceiling, while Anon-W2W retains 78-93% of baseline sBERT across setups with sub-second response latency (FRL under 0.8 s).

Targeted Speaker Poisoning Framework in Zero-Shot Text-to-Speech

Zero-shot Text-to-Speech (TTS) voice cloning poses severe privacy risks, demanding the removal of specific speaker identities from trained TTS models. Conventional machine unlearning is insufficient in this context, as zero-shot TTS can dynamically reconstruct voices from just reference prompts. We formalize this task as Speech Generation Speaker Poisoning (SGSP), in which we modify trained models to prevent the generation of specific identities while preserving utility for other speakers. We evaluate inference-time filtering and parameter-modification baselines across 1, 15, and 100 forgotten speakers. Performance is assessed through the trade-off between utility (WER) and privacy, quantified using AUC and Forget Speaker Similarity (FSSIM). We achieve strong privacy for up to 15 speakers but reveal scalability limits at 100 speakers due to increased identity overlap. Our study thus introduces a novel problem and evaluation framework toward further advances in generative voice privacy.

StyleBench: Evaluating Speech Language Models on Conversational Speaking Style Control

Speech language models (SLMs) have significantly extended the interactive capability of text-based Large Language Models (LLMs) by incorporating paralinguistic information. For more realistic interactive experience with customized styles, current SLMs have managed to interpret and control speaking style intensity from user prompts during the dialogue process. However, there remains a lack of systematic benchmarks that quantifies and evaluates the style intensity control ability in conversations. In this paper, we propose StyleBench, a multi-turn dialogue benchmark for comprehensively evaluating the style intensity control ability across four dimensions: emotion, speed, volume, and pitch. Our results reveal the performance gaps between leading SLMs and omni language models (OLMs), suggesting the underlying reasons and promising approaches for future exploration.

EmbedTalk: Triplane-Free Talking Head Synthesis using Embedding-Driven Gaussian Deformation

Real-time talking head synthesis increasingly relies on deformable 3D Gaussian Splatting (3DGS) due to its low latency. Tri-planes are the standard choice for encoding Gaussians prior to deformation, since they provide a continuous domain with explicit spatial relationships. However, tri-plane representations are limited by grid resolution and approximation errors introduced by projecting 3D volumetric fields onto 2D subspaces. Recent work has shown the superiority of learnt embeddings for driving temporal deformations in 4D scene reconstruction. We introduce $\textbf{EmbedTalk}$, which shows how such embeddings can be leveraged for modelling speech deformations in talking head synthesis. Through comprehensive experiments, we show that EmbedTalk outperforms existing 3DGS-based methods in rendering quality, lip synchronisation, and motion consistency, while remaining competitive with state-of-the-art generative models. Moreover, replacing the tri-plane encoding with learnt embeddings enables significantly more compact models that achieve over 60 FPS on a mobile GPU (RTX 2060 6 GB). Our code will be placed in the public domain on acceptance.

VoiceSHIELD-Small: Real-Time Malicious Speech Detection and Transcription

Voice interfaces are quickly becoming a common way for people to interact with AI systems. This also brings new security risks, such as prompt injection, social engineering, and harmful voice commands. Traditional security methods rely on converting speech to text and then filtering that text, which introduces delays and can ignore important audio cues. This paper introduces VoiceSHIELD-Small, a lightweight model that works in real time. It can transcribe speech and detect whether it is safe or harmful, all in one step. Built on OpenAI's Whisper-small encoder, VoiceSHIELD adds a mean-pooling layer and a simple classification head. It takes just 90-120 milliseconds to classify audio on mid-tier GPUs, while transcription happens at the same time. Tested on a balanced set of 947 audio clips, the model achieved 99.16 percent accuracy and an F1 score of 0.9865. At the default setting, it missed 2.33 percent of harmful inputs. Cross-validation showed consistent performance (F1 standard deviation = 0.0026). The paper also covers the model's design, training data, performance trade-offs, and responsible use guidelines. VoiceSHIELD is released under the MIT license to encourage further research and adoption in voice AI security.

Nwāchā Munā: A Devanagari Speech Corpus and Proximal Transfer Benchmark for Nepal Bhasha ASR

Nepal Bhasha (Newari), an endangered language of the Kathmandu Valley, remains digitally marginalized due to the severe scarcity of annotated speech resources. In this work, we introduce Nwāchā Munā, a newly curated 5.39-hour manually transcribed Devanagari speech corpus for Nepal Bhasha, and establish the first benchmark using script-preserving acoustic modeling. We investigate whether proximal cross-lingual transfer from a geographically and linguistically adjacent language (Nepali) can rival large-scale multilingual pretraining in an ultra-low-resource Automatic Speech Recognition (ASR) setting. Fine-tuning a Nepali Conformer model reduces the Character Error Rate (CER) from a 52.54% zero-shot baseline to 17.59% with data augmentation, effectively matching the performance of the multilingual Whisper-Small model despite utilizing significantly fewer parameters. Our findings demonstrate that proximal transfer within South Asian language clusters serves as a computationally efficient alternative to massive multilingual models. We openly release the dataset and benchmarks to digitally enable the Newari community and foster further research in Nepal Bhasha.

Exploring the potential and limitations of Model Merging for Multi-Domain Adaptation in ASR

Model merging is a scalable alternative to multi-task training that combines the capabilities of multiple specialised models into a single model. This is particularly attractive for large speech foundation models, which are typically adapted through domain-specific fine-tuning, resulting in multiple customised checkpoints, for which repeating full fine-tuning when new data becomes available is computationally prohibitive. In this work, we study model merging for multi-domain ASR and benchmark 11 merging algorithms for 10 European Portuguese domains, evaluating in-domain accuracy, robustness under distribution shift, as well as English and multilingual performance. We further propose BoostedTSV-M, a new merging algorithm based on TSV-M that mitigates rank collapse via singular-value boosting and improves numerical stability. Overall, our approach outperforms full fine-tuning on European Portuguese while preserving out-of-distribution generalisation in a single model.

Oral to Web: Digitizing 'Zero Resource'Languages of Bangladesh

We present the Multilingual Cloud Corpus, the first national-scale, parallel, multimodal linguistic dataset of Bangladesh's ethnic and indigenous languages. Despite being home to approximately 40 minority languages spanning four language families, Bangladesh has lacked a systematic, cross-family digital corpus for these predominantly oral, computationally "zero resource" varieties, 14 of which are classified as endangered. Our corpus comprises 85792 structured textual entries, each containing a Bengali stimulus text, an English translation, and an IPA transcription, together with approximately 107 hours of transcribed audio recordings, covering 42 language varieties from the Tibeto-Burman, Indo-European, Austro-Asiatic, and Dravidian families, plus two genetically unclassified languages. The data were collected through systematic fieldwork over 90 days across nine districts of Bangladesh, involving 16 data collectors, 77 speakers, and 43 validators, following a predefined elicitation template of 2224 unique items organized at three levels of linguistic granularity: isolated lexical items (475 words across 22 semantic domains), grammatical constructions (887 sentences across 21 categories including verbal conjugation paradigms), and directed speech (862 prompts across 46 conversational scenarios). Post-field processing included IPA transcription by 10 linguists with independent adjudication by 6 reviewers. The complete dataset is publicly accessible through the Multilingual Cloud platform (multiling.cloud), providing searchable access to annotated audio and textual data for all documented varieties. We describe the corpus design, fieldwork methodology, dataset structure, and per-language coverage, and discuss implications for endangered language documentation, low-resource NLP, and digital preservation in linguistically diverse developing countries.

Boosting ASR Robustness via Test-Time Reinforcement Learning with Audio-Text Semantic Rewards

Recently, Automatic Speech Recognition (ASR) systems (e.g., Whisper) have achieved remarkable accuracy improvements but remain highly sensitive to real-world unseen data (data with large distribution shifts), including noisy environments and diverse accents. To address this issue, test-time adaptation (TTA) has shown great potential in improving the model adaptability at inference time without ground-truth labels, and existing TTA methods often rely on pseudo-labeling or entropy minimization. However, by treating model confidence as a learning signal, these methods may reinforce high-confidence errors, leading to confirmation bias that undermines adaptation. To overcome these limitations, we present ASR-TRA, a novel Test-time Reinforcement Adaptation framework inspired by causal intervention. More precisely, our method introduces a learnable decoder prompt and utilizes temperature-controlled stochastic decoding to generate diverse transcription candidates. These are scored by a reward model that measures audio-text semantic alignment, and the resulting feedback is used to update both model and prompt parameters via reinforcement learning. Comprehensive experiments on LibriSpeech with synthetic noise and L2 Arctic accented English datasets demonstrate that our method achieves higher accuracy while maintaining lower latency than existing TTA baselines. Ablation studies further confirm the effectiveness of combining audio and language-based rewards, highlighting our method's enhanced stability and interpretability. Overall, our approach provides a practical and robust solution for deploying ASR systems in challenging real-world conditions.

When Denoising Hinders: Revisiting Zero-Shot ASR with SAM-Audio and Whisper

Recent advances in automatic speech recognition (ASR) and speech enhancement have led to a widespread assumption that improving perceptual audio quality should directly benefit recognition accuracy. In this work, we rigorously examine whether this assumption holds for modern zero-shot ASR systems. We present a systematic empirical study on the impact of Segment Anything Model Audio by Meta AI, a recent foundation-scale speech enhancement model proposed by Meta, when used as a preprocessing step for zero-shot transcription with Whisper. Experiments are conducted across multiple Whisper model variants and two linguistically distinct noisy speech datasets: a real-world Bengali YouTube corpus and a publicly available English noisy dataset. Contrary to common intuition, our results show that SAM-Audio preprocessing consistently degrades ASR performance, increasing both Word Error Rate (WER) and Character Error Rate (CER) compared to raw noisy speech, despite substantial improvements in signal-level quality. Objective Peak Signal-to-Noise Ratio analysis on the English dataset confirms that SAM-Audio produces acoustically cleaner signals, yet this improvement fails to translate into recognition gains. Therefore, we conducted a detailed utterance-level analysis to understand this counterintuitive result. We found that the recognition degradation is a systematic issue affecting the majority of the audio, not just isolated outliers, and that the errors worsen as the Whisper model size increases. These findings expose a fundamental mismatch: audio that is perceptually cleaner to human listeners is not necessarily robust for machine recognition. This highlights the risk of blindly applying state-of-the-art denoising as a preprocessing step in zero-shot ASR pipelines.