From Text Metrics to Model Internals: A Study of Whisper ASR Hallucination Detection

Researchers developed multiple approaches to detect hallucinations in OpenAI's Whisper ASR model, where the system generates fluent but unfounded transcriptions. The study found that probing the model's internal decoder states outperformed text-based and LLM-based detection methods, with a hybrid approach combining text metrics and internal representations achieving the best overall performance.

This research addresses a critical reliability problem in automatic speech recognition systems: hallucinations that sound plausible but lack any acoustic basis. As ASR models become increasingly deployed in production environments—from medical transcription to legal proceedings—detecting these errors is essential for maintaining system trustworthiness and preventing downstream application failures.

The study benchmarked three detection paradigms against real-world human-annotated speech data. Text-based classifiers using evaluation metrics achieved high recall rates but suffered significant performance degradation without reference transcripts, limiting their practical applicability. LLM-based approaches improved precision through domain-specific prompt engineering but remained less competitive than simpler alternatives, suggesting that leveraging larger language models doesn't automatically translate to better hallucination detection in specialized domains.

The most significant finding reveals that hallucination characteristics are encoded within Whisper's intermediate decoder layers, enabling accurate detection without ground-truth references. This breakthrough is particularly valuable because production systems often lack reference transcripts for comparison. The late-fusion meta-classifier combining text metrics with internal-state probing achieved superior performance by exploiting complementary signals: surface-level textual properties alongside deep model representations.

This research has immediate implications for developers deploying Whisper in safety-critical applications. The discovery that internal decoder states contain detectable hallucination signals suggests that robust error detection doesn't require external LLMs or reference data. Organizations can implement lightweight monitoring layers directly on model outputs. Future work should explore whether these detection patterns generalize across different ASR architectures and whether similar hallucination detection strategies apply to other generative models beyond speech recognition.

• →Probing Whisper's internal decoder representations outperformed text-based and LLM-based detection methods for identifying hallucinations.
• →Text classifiers using evaluation metrics achieved high recall but required reference transcripts, limiting real-world applicability.
• →Hallucination features are encoded across multiple intermediate decoding layers, enabling detection without ground-truth references.
• →Late-fusion approach combining text metrics and internal states achieved best overall detection performance.
• →LLM-based detection with domain prompting improved precision but remained less competitive than lightweight text methods.