Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

The evolution of ASR technology has transitioned from traditional statistical models like HMM-GMM to deep neural network architectures, culminating in transformer-based models such as Whisper. These models leverage large-scale pretraining on vast datasets, enabling remarkable generalization across languages and acoustic environments. Despite these advances, a critical issue persists: hallucinations, where models generate plausible but false transcriptions for non-speech or noisy inputs. Early solutions relied on confidence thresholds or heuristic filters, which proved insufficient as hallucinations can exhibit high confidence scores. Fine-tuning approaches, though effective, are resource-intensive and risk degrading overall accuracy. Recent research has begun exploring internal model mechanisms, aiming to understand and manipulate the representations within the neural network. This paper builds upon this trend, focusing on the linear structure of hallucination signals embedded in the residual activations and sparse latent features, proposing a novel, parameter-free intervention method that operates during inference, thus offering a scalable and practical solution.

The core challenge addressed is the high hallucination rate of Whisper when processing non-speech audio, which leads to the generation of irrelevant or fabricated transcriptions. Existing filtering heuristics are inadequate because hallucinated outputs often have high confidence scores, making them hard to detect and filter in real-time. Fine-tuning models to reduce hallucinations is costly and can impair recognition performance. Moreover, the lack of understanding of the internal representations that encode hallucination signals limits the development of more effective, generalizable solutions. Therefore, the fundamental problem is how to reliably identify and suppress hallucination-prone inputs by exploiting the internal states of the model without retraining or fine-tuning, ensuring both robustness and efficiency.

The key innovations include: 1) revealing that hallucination-related information is linearly separable in both raw activations and SAE latent spaces, validated through layer-wise AUC analysis; 2) designing two parameter-free steering strategies—activation space steering and SAE latent space steering—that intervene during inference without modifying model weights; 3) demonstrating that sparse, discriminative features can be targeted for intervention, enabling precise suppression of hallucinations while preserving speech recognition accuracy. These contributions collectively provide a new paradigm for internal model control, combining interpretability, efficiency, and robustness, which surpasses prior post-processing or fine-tuning methods.

• �� Data collection: Extract residual stream activations from each encoder layer of Whisper, applying average pooling to obtain fixed-size representations.
• �� SAE training: Train sparse autoencoders on these activations across diverse datasets, enforcing sparsity via L1 regularization to produce disentangled, interpretable latent features.
• �� Discriminability assessment: Use logistic regression classifiers to evaluate the linear separability of hallucination signals at each layer, computing AUC scores to identify the most discriminative layers.
• �� Feature importance analysis: Derive importance scores from classifiers to select top-k SAE latent features that encode hallucination signals.
• �� Steering vector construction: For activation space, compute the difference vector between hallucinating and non-hallucinating sets; for SAE, create a sign-flipped sparse vector based on importance scores.
• �� Inference-time intervention: Apply the steering vectors to residual activations or SAE latent representations, either additively or multiplicatively, to steer the internal states away from hallucination regimes.
• �� Evaluation: Measure hallucination rate reduction and WER impact across multiple datasets and model variants, optimizing hyperparameters (α, k) via grid search on non-speech data.

The experimental setup involves testing on two Whisper variants—small and large-v3—using datasets such as FSD50k, MUSAN, WHAM! for non-speech, and LibriSpeech, FLEURS, AISHELL-1 for speech. The datasets are split to prevent data leakage, with hyperparameter tuning performed solely on non-speech training data. The primary metrics include hallucination rate (HR) and WER/CER. The classifiers trained on internal representations validate the linear separability of hallucination signals. Hyperparameters for steering (α, k) are optimized through grid search, with evaluation on held-out non-speech test sets to assess generalization. Ablation studies compare the effectiveness of activation versus SAE-based steering, analyzing the impact of different layers, feature dimensions, and intervention strengths. Results demonstrate consistent hallucination reduction across datasets and models, with minimal degradation in recognition accuracy.

Quantitative results show that SAE-based steering reduces hallucination rates from over 70% to below 15% across datasets, with WER increases less than 1%. Layer-wise analysis confirms that the most discriminative features are concentrated in the final encoder layers, with AUC scores exceeding 0.95. The sparse latent features, typically around 50-100 dimensions, contain the majority of hallucination signals, enabling targeted interventions that outperform activation steering in stability and cross-domain robustness. The methods generalize well across model sizes and languages, demonstrating their practical applicability and scalability.

The proposed internal representation steering techniques can be integrated into existing ASR pipelines to enhance robustness against non-speech noise and background interference. They are particularly suitable for real-time applications such as voice assistants, live captioning, and automated transcription services, where minimizing false positives is critical. The parameter-free nature allows easy deployment without retraining, making it accessible for industry adoption. Additionally, the insights gained from internal representation analysis can inform future model design, leading to inherently more robust architectures. Long-term, these methods could contribute to the development of self-correcting, explainable speech recognition systems capable of operating reliably in diverse, noisy environments.

The effectiveness of the methods hinges on the linear separability of hallucination signals, which may not hold in highly ambiguous or adversarial scenarios. The interventions are currently limited to the final encoder layer, potentially missing opportunities for multi-layer or hierarchical control. Computational overhead from SAE encoding and latent space manipulation could hinder real-time deployment in resource-constrained settings. The generalization to low-resource languages and unseen acoustic conditions requires further validation, and the approach may need adaptation for different model architectures or modalities. Future work should address these limitations by developing adaptive, multi-layer strategies, optimizing computational efficiency, and exploring broader applicability.