PaperlayerView exact revision on arXiv

Source-linked AI summary

Lost in Speech: Trilingual Spoken Hallucination Detection Across Audio and Transcripts

Meruyert Aristombayeva, Jason S. Lucas, Chaewan Chun, Dongwon Lee

arXiv:2608.24707v1cs.CL

TL;DR

Spoken hallucination detection lacks multilingual evidence, especially for low-resource languages. This paper builds and evaluates a trilingual benchmark across text, transcripts, and audio, finding that transcript-based detection generally outperforms direct audio processing while synthetic-trained detectors transfer to real-world fakes.

  • Problem

    Spoken hallucination detection remains largely unexplored, particularly for low-resource languages, while speech pipelines introduce noise and cascading errors.

  • Method

    The paper constructs a 12,013-sample English–Russian–Kazakh benchmark with controlled hallucinations, text/audio/transcript variants, and real-world fake-news evaluation.

  • Results

    Transcript-based detection generally outperforms direct audio processing, with degradation most pronounced for Kazakh; synthetic-trained detectors transfer strongly to real-world fakes.

  • Takeaways & Limitations

    Intermediate textual representations remain more reliable than raw audio for detecting semantic inconsistencies under the evaluated multilingual speech conditions.

  • Takeaways & Limitations

    The benchmark uses read speech synthesized through a TTS→ASR cascade rather than spontaneous, conversational, or natively recorded speech.

Abstract

from arXiv · show

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.

1 Introduction

Spoken hallucination detection is understudied, especially for low-resource languages, because speech pipelines add noise beyond text processing. The paper introduces a trilingual benchmark and evaluates detection across text, transcripts, and audio.

  • Spoken hallucination detection remains scarce, particularly for low-resource languages, despite speech systems producing potentially harmful fabricated phrases.
  • The benchmark spans English, Russian, and Kazakh with controlled fabrication, contradiction, and context-inconsistency hallucinations at mild, moderate, and severe levels.
  • Transcript-based detection generally outperforms direct audio processing, with TTS–ASR degradation most pronounced for Kazakh.
  • 12,013 aligned news samples are provided as original or hallucinated text, synthesized speech, and ASR transcripts.
  • A real-world split of 290 fact-checked fake news stories supports transfer evaluation and Russian provenance analysis of synthetic-benchmark confounds.
  • The study compares fine-tuned multilingual encoders and zero-shot multimodal audio-language decoders across text and audio settings.

2 Benchmark Construction

The benchmark combines controlled multilingual hallucinations with speech-pipeline representations and a real-world fake-news evaluation split. Its construction explicitly addresses generation quality, provenance confounding, and language-dependent transcription distortion.

  • 2.1 Synthetic Subset: 12,013 news samples cover English, Russian, and Kazakh articles with original and hallucinated versions across text, synthesized audio, and ASR transcripts.
  • 2.1 Synthetic Subset: Hallucinations are generated by predefined fabrication, contradiction, and context-inconsistency types at mild, moderate, and severe levels.
  • 2.1 Synthetic Subset: Generation prompts include self-assessment, followed by external validation from two judge models from families disjoint from the generators.
  • 2.2 Real-World Evaluation Split: 290 fact-checked false items form a real-world evaluation split, with 225 collected natively in Russian and 65 natively in Kazakh.
  • 2.2 Real-World Evaluation Split: Real-world items include matched truthful negatives, are excluded from training and development, and are rendered through the same TTS→ASR pipeline in three modalities.
  • 2.3 Speech Synthesis and Transcription: WER/CER are 7.30%/2.74% for English, 9.08%/4.74% for Russian, and 34.04%/16.90% for Kazakh.
  • 2.3 Speech Synthesis and Transcription: Kazakh’s elevated transcription distortion reflects TTS–ASR cascade effects, phonetic changes, proper names, and agglutinative morphology.
  • 2.4 Quality Evaluation: LLM vs. Human: 94.2% of samples were mostly intelligible, while 61.7% had minor and 38.3% major semantic discrepancies relative to source text.

3 Empirical Evaluation

The evaluation compares fine-tuned multilingual encoders and zero-shot decoder models across original text, ASR transcripts, and direct audio, including transfer to real-world misinformation. Transcript-based processing generally performs better than direct audio, while binary degradation follows ASR error and synthetic-trained encoders transfer strongly to real-world fakes.

  • RQ1–RQ3: The study evaluates original and ASR-transcribed text using binary and multiclass hallucination detection, alongside zero-shot decoder comparisons.The multiclass tasks distinguish hallucination type and severity.
  • RQ1: Text-based: Fine-tuned encoders achieve binary F1 of 0.52–0.89, while type and severity classification remain harder at 0.14–0.68 and 0.15–0.66.Performance is generally higher on original text than ASR transcripts, though not uniformly.
  • RQ1: Text-based: Binary-task degradation for mDeBERTa and ReMBERT increases with per-language ASR error, whereas this relationship is inconsistent for type and severity tasks.The strongest degradation occurs in low-resource Kazakh.
  • RQ2: Audio-based: Transcript-based inputs generally outperform direct audio across five zero-shot decoder models spanning 1.5B to 33B parameters.Qwen2.5-Omni performs strongest overall, particularly for type classification, while audio-based binary detection has no consistent leader.
  • RQ3: Real-world transfer: On real-world data, original-text macro-F1 reaches 0.819 versus 0.786 for Russian and 0.883 versus 0.863 for Kazakh on ReMBERT’s real-world versus synthetic tests.Transcript-based real-world macro-F1 remains .753–.835, with larger O→T degradation in Kazakh.
  • RQ3: Provenance: Synthetic binary F1 overstates factuality sensitivity because Russian provenance analysis finds model-dependent reliance on machine-written style.mDeBERTa remains more selective than ReMBERT on faithful machine rewrites of truthful articles.

4 Insights and Challenges

The experiments identify transcription quality, audio reasoning, language resources, task granularity, and provenance as central challenges. Transcript representations are generally more reliable, but synthetic binary evaluation can conflate factuality with machine-style signals.

  • Transcription quality: Binary detection degrades on ASR transcripts relative to original text, with degradation increasing alongside ASR error for mDeBERTa and ReMBERT.The relationship is less consistent for fine-grained tasks.
  • Zero-shot audio reasoning: Transcript-based inputs generally outperform direct audio, while smaller zero-shot models remain near chance across modalities.Zero-shot multilingual audio reasoning is fragile and sensitive to model scale and task granularity.
  • Low-resource conditions: Kazakh has the highest ASR error and pronounced downstream degradation, consistent with compounded TTS→ASR noise in low-resource conditions.This language-specific gap makes low-resource speech processing especially challenging.
  • Task granularity: Binary detection is substantially easier than distinguishing hallucination types and severity levels.Fine-grained understanding requires richer supervision or stronger multimodal grounding.
  • Provenance: Real-world macro-F1 matches or exceeds synthetic-test performance, but faithful LLM rewrites reveal a model-dependent provenance component.The real-world split reduces the original provenance confound because both classes originate from human-written content.
  • Scope boundary: The benchmark isolates content-level hallucinations rendered via TTS rather than spontaneous conversational speech.The real-world split partially relaxes this constraint by evaluating naturally occurring misinformation.

5 Conclusion

The paper introduces a multilingual benchmark and real-world evaluation split for comparing hallucination detection across text, transcripts, and audio. Results show modality- and resource-dependent performance, strong synthetic-to-real transfer, and provenance-related confounding, motivating noise- and provenance-robust detection.

  • Contribution: The benchmark contains 12,013 aligned English, Russian, and Kazakh news samples plus fact-checked Russian and Kazakh fakes for evaluation.Samples support comparison across original text, synthesized speech, and ASR transcripts.
  • Findings: Original text generally outperforms ASR transcripts, with pronounced degradation in low-resource Kazakh, while transcript inputs generally outperform direct audio.These results indicate persistent modality and language-resource effects.
  • Findings: Synthetic training transfers strongly to real-world fakes, while Russian provenance analysis identifies both veracity-related and model-dependent machine-style signals.The analysis exposes a confound in synthetic hallucination evaluation.
  • Future challenges: The paper identifies speech-pipeline noise, low-resource degradation, and fine-grained classification as challenges for future work.It points toward noise-robust multimodal modeling and provenance-robust detection objectives.

Limitations

The study’s conclusions are bounded by its speech setting, language and domain coverage, synthetic generation process, and real-world evaluation design. Several dataset-label and model-evaluation choices also constrain interpretation and generalization.

  • Read TTS→ASR speech excludes spontaneous, conversational, and natively recorded speech, limiting direct transfer to natural-speech deployment.Observed degradation jointly reflects synthesis and recognition artifacts.
  • WER is not directly comparable across typologically different languages, and agglutinative Kazakh can inflate word-level errors relative to CER.The study therefore reports CER alongside WER and relates ASR error to downstream degradation.
  • Results cover only English, Russian, and Kazakh in the news domain, so generalization to other languages, domains, or speaking styles remains uncertain.
  • LLM-generated hallucinations, LLM judging, and partial human inspection may leave residual model-specific artifacts.
  • Synthetic binary labels correlate text provenance with veracity, although type and severity tasks within hallucinated items are unaffected.The study quantifies this confound using human-written fakes and faithful LLM rewrites.
  • The real-world split is an evaluation set with uneven Russian–Kazakh composition, no English, possible translation artifacts, non-random fact-checked items, and register shift from synthetic newswire.Part of the high detection rate may reflect stylistic distribution shift.
  • Decoder results are zero-shot and lack audio fine-tuning, so reported audio performance is a lower bound for adapted multimodal models.

Ethical Considerations

The dataset uses publicly available news and fact-checking sources while limiting redistribution of copyrighted or fabricated content. The release documents model use and restricts the resource to research-oriented detection work.

  • Original article texts are not redistributed; the release provides source URLs, generated rewrites, labels, transcripts, audio metadata, and prompts.Unavailable pages are represented by outlet and publication metadata.
  • Copyrighted material is used under the stated limitations of Kazakhstan’s copyright law, and the corpus contains no private or sensitive personal data.
  • The real-world split contains publicly debunked misinformation, but releases URLs and metadata rather than the fake texts themselves.
  • Because the resource contains deliberately fabricated and contradictory content, it is intended solely for developing and evaluating detection systems under a research-only license.The paper identifies potential misuse as disinformation.
  • Generative models and independent LLM judges are used for hallucination generation and partial quality evaluation, with this use disclosed under ACL policy.

A Generation and Evaluation Prompts

Appendix A documents the prompt templates, model assignments, language variants, quality reports, and constraints used to generate and evaluate controlled hallucinations. Context inconsistency receives distinct structural safeguards and evaluation dimensions.

  • A Generation and Evaluation Prompts: The appendix records canonical templates, language-specific deltas, superseded prompts, preprocessing prompts, and runtime placeholders.Table 6 lists placeholders used in the templates.
  • A Generation and Evaluation Prompts: Generation uses GPT-3.5-Turbo and GPT-4 for English, Gemini 2.0 Flash-Lite for Russian and Kazakh, while Claude Haiku and DeepSeek Chat serve as judges.
  • A.1 Hallucination Generation: Fabrication and contradiction prompts provide the original article, specify a severity level, require formatted full-text output, and request quality or deviation reports.The prompts also require strict output-language and formatting behavior.
  • A.1 Hallucination Generation: Contradiction prompts explicitly extract narrative and factual information, generate contradictions, and evaluate factual accuracy, credibility, fluency, coherence, relevance, satisfaction, and type adequacy.
  • A.1 Hallucination Generation: Russian and Kazakh variants localize extraction, evaluation scales, explanations, and full-text markers while adding a requirement to preserve similar article length and structure.
  • A.1 Hallucination Generation: Context inconsistency alters logic, narrative structure, entities, causality, or timeline while preserving explicit numeric values and keeping length within ±15%.Its quality report includes Internal Consistency and Logical Coherence rather than the factuality dimensions used by the other types.
  • A.1 Hallucination Generation: Context-inconsistency prompts require contradiction and logical-flaw identification, full rewritten output, and ratings across internal consistency, logical coherence, fluency, coherence, relevance, satisfaction, and type adequacy.
  • A.1 Hallucination Generation: The appendix distinguishes context inconsistency from the other two types by forbidding numeric changes, imposing the ±15% length constraint, and substituting structural quality dimensions.

B Training Details

Training uses fixed fine-tuning settings for multilingual encoders, with a specialized mDeBERTa-v3 configuration and stratified data splits. ASR training and selection are language-specific, especially for Kazakh.

  • Encoder Training: Encoders use effective batch size 16, learning rate 2e-5, three epochs, maximum sequence length 512, fp16, checkpointing, and development macro-F1 early stopping.The run uses seed 42 and patience 1 for early stopping.
  • ASR Selection: Whisper-large-v3 is used for English and Russian, while fine-tuned wav2vec2 is selected for Kazakh because it nearly halves Kazakh error on identical audio.
  • Encoder Training: mDeBERTa-v3 instead uses fp32, eager attention, learning rate 1e-5 with 10% warmup, and physical batch 8 with accumulation 2.These changes address fp16 instability while preserving effective batch size 16.
  • Data Splits: 80/10/10 data splits are stratified by language and label.

C Models

The study documents the models used across detection, generation, and speech processing, together with their modality support.

  • Model Inventory: Table 8 lists detection, generation, and speech models and records each model’s modality support.The related materials distinguish audio and text support for the study’s models.

D Full LLM-Based Quality Evaluation

The paper reports comprehensive LLM-based quality evaluation across hallucination type, severity, and language, alongside model and judge-agreement documentation.

  • Quality Evaluation: Complete LLM-based quality evaluation is reported by hallucination type, severity, and language.The evaluation is discussed in Section 2.4 and summarized in Table 10.
  • Models: Table 8 documents the models used in the study and their modality support.
  • Judge Agreement: Table 9 reports per-dimension agreement between the original and independent judges.Agreement is presented for each quality dimension.

E Real-World Split Details

The real-world split combines fact-checked fake news with controlled truthful negatives and evaluates quality using multiple dimensions. Fake items originate in Russian and Kazakh and are translated across languages.

  • Fake News Collection: 290 fake items come from factcheck.kz, including 225 natively Russian and 65 natively Kazakh items.The collection was downloaded in October 2024 after removing 15 exact duplicates.
  • Translation: Each fake item is machine-translated into the other language and manually reviewed by a fluent bilingual author.All 290 translations were judged adequate, with no post-editing required.
  • Quality Evaluation: LLM-as-a-judge quality scores cover hallucination type, severity, and language on a 1-to-5 scale.The validation sample contains 216 items, with eight items per cell.
  • Evaluation Dimensions: The quality-evaluation legend defines symbols for context inconsistency, fabrication, and contradiction, plus abbreviations for evaluation dimensions.The listed dimensions include adequacy of type, fluency, internal consistency, logical coherence, relevance, coherence, credibility, factual accuracy, and satisfaction.
  • Truthful Negatives: Truthful negatives are generated with Gemini 2.0 Flash-Lite under instructions to preserve facts and introduce no unsupported information.
Loading 2608.24707v1…