This work presents the first multilingual spoken hallucination benchmark comprising 12,013 news samples across English, Russian, and Kazakh with controlled hallucinations of three types and three severity levels, and assesses fine-tuned multilingual encoders and, in zero-shot in-context settings, multimodal decoder models on transcript-based versus direct audio processing.
Abstract
While text-based hallucination detection has been extensively studied, spoken hallucination detection remains largely unexplored, particularly for low-resource languages. We present the first multilingual spoken hallucination benchmark comprising 12,013 news samples across English, Russian, and Kazakh with controlled hallucinations of three types and three severity levels. Samples comprise original articles and aligned hallucinated counterparts in text and audio. We complement the synthetic corpus with 290 fact-checked fake news items collected natively in Russian (225) and Kazakh (65), translated into the other language and rendered through the same TTS-ASR pipeline. We assess fine-tuned multilingual encoders and, in zero-shot in-context settings, multimodal decoder models on transcript-based versus direct audio processing. Transcript-based detection generally outperforms direct audio processing, with binary-task degradation for strong encoders tracking per-language ASR error. On real-world fakes, synthetic-trained detectors transfer strongly (macro-F1 0.82-0.88 on original text), while Russian provenance analysis reveals both veracity-related and model-dependent machine-style signals, quantifying a key confound in synthetic hallucination benchmarks.
Large language models generate fluent text that can contain unfaithful claims -- a phenomenon known as hallucination. We present a multi-signal detection pipeline combining fine-tuned DeBERTa-v3 classification, Monte Carlo (MC) Dropout uncertainty quantification, and temperature-scaled calibration for response-level hallucination detection. Evaluated on the HaluEval benchmark, our pipeline achieves F1=0.915 and AUROC=0.977 on general-domain tasks, with per-task F1 scores of 0.97 (QA), 0.96 (Summarization), and 0.82 (Dialogue). MC Dropout inference further improves accuracy to 93.2%. A context ablation study confirms the model performs genuine entailment reasoning rather than exploiting surface patterns, with summarization F1 dropping 24% when knowledge context is removed. Learning curve analysis reveals that 25% of training data captures 77% of full-data performance. Beyond detection, we apply Direct Preference Optimization (DPO) to a Qwen2.5-0.5B generator, reducing its hallucination rate from 85.5% to 37.7% (55.9% relative reduction) as measured by our detector. Cross-domain evaluation on the SciFact biomedical benchmark shows that general-domain training transfers poorly (F1=0.52), motivating domain-specific fine-tuning. PubMedBERT fine-tuned on SciFact achieves F1=0.63 and AUROC=0.81, demonstrating that domain-matched pre-training is the strongest adaptation strategy. Code and models are available at https://github.com/varunteja99/hallucination-detection-nlp
Whisper is a widely used foundation model for automatic speech recognition (ASR), but its generative decoder can produce fluent hallucinated transcripts for inputs containing little or no speech. We propose a training-free, inference-time method to reduce these hallucinations using low-rank projection of decoder activations. A compact hallucination-associated subspace is estimated from non-speech calibration data, and decoder hidden states are projected away from this subspace during inference. We evaluate two variants: always-on, which applies projection to all inputs, and gated, which applies it only when Whisper predicts that an input is likely non-speech. Across non-speech benchmarks, always-on projection reduces average hallucination rate (HR) from 31.31% to 2.44%, a 92.21% relative reduction, while gated projection reduces HR to 3.74%, an 88.05% relative reduction, with lower false rejection of genuine speech. On LibriSpeech, gated projection increases absolute word error rate (WER) by 0.33-4.39 percentage points and yields false-rejection rates (FRR) of 0.41--9.97% across model and split settings. These results show that low-rank activation projection can substantially suppress Whisper hallucinations without retraining, while providing a controllable trade-off between hallucination suppression and speech recognition performance.
Introduction Aphasia affects expressive and receptive communication and may influence the affective tone expressed during clinical speech tasks. This study presents an exploratory weakly supervised NLP analysis of positive/negative affective-tone proxies in AphasiaBank transcripts with paired audio. Methods We extracted sentence embeddings from DistilBERT (e∈R768) and recording-level acoustic summaries (MFCC13, ZCR, RMS, spectral centroid, and spectral bandwidth; a∈R17). Text and acoustic features were concatenated (x=[e;a]∈R785) and classified using Random Forest models. Sentiment labels were generated using an SST-2-derived weak-supervision pipeline and should be interpreted as pseudo-labels rather than clinical ground truth. To evaluate modality contribution and potential leakage, we compared text-only, audio-only, and fused text-audio models under utterance-level and recording-disjoint splits. A small five-rater evaluation was used to examine human judgment alignment. Results Under the recording-disjoint split, both the text-only and fused text-audio models achieved 97.9% accuracy and 0.791 macro-F1, while the audio-only model achieved 55.4% accuracy and 0.388 macro-F1. These results indicate that classification performance was primarily driven by textual embeddings, while the recording-level acoustic summaries did not improve performance over text-only features. Across the pseudo-labeled corpus, aphasic utterances were more often labeled negative than control utterances. Age-stratified summaries showed subgroup variation in pseudo-label distributions, but these patterns were treated descriptively because labels were model-derived. Human-rater agreement was low for aphasic utterances, indicating that affective-tone interpretation in fragmented clinical speech is ambiguous. Discussion These findings should be interpreted as exploratory evidence about weakly supervised affective-tone proxies, not as validated clinical sentiment recognition. The results highlight both the promise of clinical NLP for aphasia discourse analysis and the need for independent human-labeled validation, utterance-aligned acoustic features, and careful control of domain, task, age, and topic bias.
Shamiha Binta Manir, A. Kothari, William L. Gross et al.· Frontiers in Digital Health· 0 citations
Large audio-language models (LALMs) can hallucinate audio objects, answering"yes"to absent sound events, thus undermining reliability in audio question answering. We propose Token-Adaptive Decoding (TAD), a training-free strategy for hallucination mitigation that grounds the initial yes/no decision by contrasting logits under real audio with a matched silent reference. TAD introduces a token-adaptive, confidence-guided gate that is decision-critical at the first decoding step and class-conditional on affirmative tokens, using the audio-silent margin to avoid overcorrection when evidence is weak or already sufficient. Experiments on AudioCaps-Hallucination show that, relative to Audio-Aware Decoding (AAD), a contrastive baseline with fixed contrast strength, TAD improves F1 for Qwen2 by 0.059 to 0.117 across Popular, Adversarial, and Random splits, and for Gemma by 0.025 to 0.064, while on Clotho-AQA it raises F1 from 0.810 to 0.816 on Qwen2 and remains comparable to AAD on Gemma.
He-Yu Chang, Nianwen Si, Hao Zhang et al.· 0 citations
Advancement of zero-shot text-to-speech synthesis is currently hindered for low-resource languages by the scarcity of large-scale, high-fidelity speech datasets. Traditional alignment-based methods require rare verbatim transcripts, while standard in-the-wild pipelines often rely on single-model automatic speech recognition and silence-based segmentation, leading to transcription errors and truncated prosody. To address these challenges for the Persian language, this paper introduces PersianVox, a fully automated pipeline designed to generate high-quality speech corpora from unlabeled web data. Our approach integrates a novel prosody-aware segmentation strategy that utilizes acoustic turn-detection to preserve linguistic completeness and optimize utterance duration for long-context modeling. Furthermore, we employ a dual-model agreement mechanism, leveraging two distinct model architectures to filter unreliable transcriptions without ground truth. This pipeline yields a 2,400-hour multi-speaker dataset, the largest open-source speech resource available for Persian to date. Additionally, we provide the first comparative benchmark of speech quality assessment methods for Persian, releasing a human-annotated subset to facilitate future research.
Saeedreza Zouashkiani, Soheil Khalesi, Saman Soleimani Roudi et al.· 0 citations
An audio-guided distillation framework for multilingual VSR that exploits synchronized audio-visual speech during training while preserving visual-only inference is introduced, and results further confirm the contribution of LLM decoding, teacher initialization, and audio-guided distillation to visual-only recognition performance.
B. Mocanu, Ruxandra Tapu· Multimedia tools and applica...· 0 citations
A new method, called CW-Net, translates the reasoning process of an autonomous vehicle’s AI system into understandable concepts that explain its behavior.
MIT News · Artificial Intelligence· news.mit.eduAug 31, 2026
With millions of users across the world, Julia has been used to conduct cutting-edge research and to design new drugs, jet engines, heat pumps, and more.
