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

Problem Statement

Automatic Speech Recognition (ASR) hallucinations—fluent transcriptions generated without basis in the audio input—pose significant risks to downstream applications, potentially injecting offensive, stereotypical, or factually incorrect content. While traditional error metrics like Word Error Rate (WER) and semantic distances (e.g., BERTScore, SeMaScore) serve as general performance proxies, they are insufficient for distinguishing hallucinations from other error types, particularly in the absence of ground-truth reference transcripts. Existing detection methods often rely on oracle settings (requiring reference text) or synthetic data, leaving a gap in robust, reference-free detection for real-world spontaneous speech. This study addresses the challenge of detecting hallucinations in the Whisper large v3 model using human-annotated real-speech data, evaluating three distinct paradigms: text-based metrics, Large Language Model (LLM) prompting, and internal decoder state probing.

Methodology

The research utilizes the HALAS dataset, which contains human-annotated hallucination labels for predictions from seven ASR models on Earnings-22 audio data. The study focuses on the Whisper large v3 model, where approximately 23.8% of utterances are marked as hallucinations. The evaluation employs 5-fold cross-validation, stratified by hallucination rate and audio duration, using ROC AUC, F1 score, precision, and recall as primary metrics.

The investigation proceeds through three detection frameworks:

1. Text-Based Detection:

   • Oracle Features: Metrics requiring ground truth, including WER, Character Error Rate (CER), Insertion Error Rate (IER), Length Ratio (LenR), BERTScore, SeMaScore, and a Common Hallucinated Phrase (CHP) detector.
   • Reference-Free Features: Metrics operating solely on the prediction, including 5-gram Repetition rate (Rep), Stopword Ratio (StopW), GPT-2 Perplexity (PPL), Characters Per Second (CPS), wav2vec Forced Alignment confidence (Align), and a Naive CHP detector.
   • Classifiers: Logistic Regression, Random Forest, and XGBoost were trained on these features.

2. LLM-Based Detection:

   • Zero-shot baselines using GPT-4o mini and Gemini 2.0/3.0 Flash were established.
   • The approach was iteratively enhanced via prompt engineering: upgrading the reasoning model, injecting domain-specific pathology knowledge (Whisper's non-speech hallucination list), adding few-shot examples, and adapting prompts for strictly reference-free settings.

3. Internal State Probing:

   • Initial attempts using Whisper's built-in confidence heuristics (Average Log-probability, Compression Ratio, No-Speech Probability) proved ineffective.
   • The study instead probed the intermediate decoder representations. Embeddings were extracted from Self-Attention (SA), Cross-Attention (CA), and the final output (D) layers across the 32 decoder layers.
   • To handle variable sequence lengths, mean and max-pooling were applied to the sequence and its deltas (step-by-step differences).
   • A Bidirectional Long Short-Term Memory (BLSTM) classifier was trained on these sequential embeddings to detect hallucinations without reference transcripts.

4. Late Fusion:

   • A meta-classifier (Logistic Regression) was constructed to combine the outputs of the best-performing text-based (XGBoost) and internal-state (BLSTM) detectors, along with audio duration as a feature.

Key Results

Text-Based Detection

• Oracle vs. Reference-Free: Oracle metrics achieved strong performance (e.g., BERTScore AUC: 82.3%, XGBoost F1: 62.8%). However, reference-free text features alone suffered severe degradation. The best reference-free XGBoost model dropped to an F1 of 37.7%, with a significant collapse in recall (15.5% for Logistic Regression), indicating that text-based detection is highly dependent on ground-truth anchors.
• Feature Importance: Tree-based ensembles (XGBoost) significantly outperformed linear baselines, confirming that hallucination detection benefits from non-linear interactions between features.

LLM-Based Detection

• Performance: Out-of-the-box LLMs struggled (F1 ~41%). While prompt engineering (domain knowledge, few-shot examples) improved performance to an F1 of 58.7%, LLMs failed to surpass the lightweight XGBoost text classifier (F1 62.8%) despite the massive computational overhead.
• Reference-Free Limitation: In strictly reference-free settings, LLM performance collapsed to an F1 of 32.8%, mirroring the degradation seen in text-only methods.

Internal State Probing

• Layer Analysis: Hallucination signals were found to be encoded across intermediate layers rather than localized to a single layer or the final EOS token. Mean-pooling of deltas at layer 15 achieved the highest AUC (82%) among linear probes.
• BLSTM Performance: The BLSTM classifier utilizing the full decoding sequence from the final output (D) layers achieved an AUC of 87.6% and an F1 score of 65.5%.
• Significance: This represents the strongest reference-free performance in the study, outperforming the best text-feature models (F1 62.8%) without requiring ground-truth transcripts.

Cross-Paradigm Fusion

• Complementarity: The paradigms captured partially non-overlapping signals (agreement 0.64–0.73). Text classifiers maximized recall (0.73) but suffered in precision, while internal state detectors offered the most balanced performance.
• Fusion Results: A late-fusion meta-classifier combining text and internal state probabilities achieved the highest overall performance with an F1 score of 68.3% and an AUC of 90.0%.
• Limitations: While fusion improved metrics, it forfeited the "reference-free" capability of the internal state detector alone, as the text component requires ground truth.

Significance and Claims

The paper claims that while text-based and LLM-based methods are effective in oracle settings, they are fundamentally constrained in zero-shot deployments due to the lack of ground-truth anchors. The primary contribution is the demonstration that probing the internal states of the Whisper decoder provides a highly competitive, reference-free solution for hallucination detection.

The study highlights that hallucination traits are encoded systemically across intermediate decoding layers, not just in final outputs. By utilizing a BLSTM on these internal states, the authors achieve robust detection without reference transcripts. Furthermore, the research establishes that a late-fusion approach can maximize performance by leveraging the complementary nature of text and internal-state signals, though this comes at the cost of requiring reference data.

The authors conclude that internal state probing offers a promising avenue for real-world ASR hallucination detection, applicable to other autoregressive encoder-decoder models, while noting that extending this approach to alternative architectures and developing specialized lightweight LLMs for hallucination detection remain important directions for future work.