Hallucinations in Neural Automatic Speech Recognition: Identifying Errors and Hallucinatory Models

TL;DR

This work defines hallucinations in neural ASR as fluent outputs that are semantically disconnected from the input, showing that standard metrics like WER fail to reveal them. It introduces a perturbation-based, test-time evaluation framework that does not require training data to assess a model's susceptibility to hallucinations, complemented by a semantic-fluency analysis combining cosine similarity and language-model perplexity. Through dataset-noise experiments (label mismatches) on LibriSpeech 360 and CommonVoice, the authors characterize how error types correlate with noise patterns and demonstrate a detection algorithm that differentiates hallucinatory from phonetic errors. The findings highlight practical implications for ASR reliability and security in deployed systems and point to future work on generalization across languages and attenuation of hallucinatory outputs.

Abstract

Hallucinations are a type of output error produced by deep neural networks. While this has been studied in natural language processing, they have not been researched previously in automatic speech recognition. Here, we define hallucinations in ASR as transcriptions generated by a model that are semantically unrelated to the source utterance, yet still fluent and coherent. The similarity of hallucinations to probable natural language outputs of the model creates a danger of deception and impacts the credibility of the system. We show that commonly used metrics, such as word error rates, cannot differentiate between hallucinatory and non-hallucinatory models. To address this, we propose a perturbation-based method for assessing the susceptibility of an automatic speech recognition (ASR) model to hallucination at test time, which does not require access to the training dataset. We demonstrate that this method helps to distinguish between hallucinatory and non-hallucinatory models that have similar baseline word error rates. We further explore the relationship between the types of ASR errors and the types of dataset noise to determine what types of noise are most likely to create hallucinatory outputs. We devise a framework for identifying hallucinations by analysing their semantic connection with the ground truth and their fluency. Finally, we discover how to induce hallucinations with a random noise injection to the utterance.

Figures (6)

• Figure 1: Word Error Rate (WER) distribution across evaluation dataset of assessed models. We exclude 0 WER scores since they are not considered error and for histogram clarity. Although the WER scores alone are not informative with regard to the types of errors produced by the models, the WER distribution in UR and RR model shows increase of highly erroneous outputs.
• Figure 2: The cosine similarity and perplexity kernel distribution across 5 models, showing clear distinction between phonetic errors and hallucinations.
• Figure 3: Comparison between cosine similarity distributions before and after perturbations in baseline and RR model. The higher mean of cosine similarity metric in RR model after perturbation suggests an increase of the oscillations, due to the noise from repetitive labels.
• Figure 4: The error distribution in baseline and UU model before and after running the hallucination detection algorithm. The lower peak of the distribution near 0 cosine similarity indicates hallucinatory outputs. The increase in hallucination ration in Unique-Unique dataset is not reflected in its WER results.
• Figure 5: Ratio of detected hallucinations per model per evaluation dataset.