Hallucinations in Neural Automatic Speech Recognition: Identifying Errors and Hallucinatory Models (2401.01572v1)

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.

• The paper introduces a novel framework that uses a perturbation-based method to identify hallucinations in neural ASR outputs.
• It reveals that traditional metrics like WER do not differentiate between hallucinatory and non-hallucinatory models.
• The study highlights the influence of dataset noise on inducing errors, emphasizing the need for further research on mitigation strategies.

Introduction

Automatic Speech Recognition (ASR) systems have significantly improved with the advent of neural architectures. However, one of the lesser-discussed challenges they face is the output of 'hallucinations'. Hallucinations within the context of neural models refer to outputs that are coherent and fluently structured yet have no semantic connection to the input utterance. This issue could potentially mislead users, introducing concerns about the system's reliability and credibility.

Hallucination Detection

A novel aspect of this paper is its investigation into ASR hallucinations, developing a framework to detect and analyze them. An intriguing finding of the paper is that commonly used metrics like Word Error Rate (WER) fail to discern between models prone to hallucinations and those that are not. To tackle this, the researchers propose a perturbation-based method, applied during testing, which does not require access to the original training data. This method is adept at identifying models susceptible to hallucinations even when they have similar WERs to non-hallucinatory models.

Understanding Hallucination-Prone Models

By inducing hallucinations through the injection of random noise into utterances, the paper examines how various factors like dataset noise contribute to ASR errors, elucidating the connection between the types of dataset noises and the likelihood of producing hallucinatory outputs. A comprehensive analysis of these outputs' semantic content and fluency helps to distinguish them from other ASR errors like phonetic substitutions or oscillations, which involve repeated n-grams.

Implications and Future Research

While the paper makes significant headway into understanding and detecting ASR hallucinations, it also points out limitations and necessary further research. The created methodologies are currently applied to English ASR, implying the need for broader linguistic applications. Furthermore, while the paper delineates hallucinations, it does not address their mitigation — a critical next step in enhancing ASR system security and performance.