Hallucination Benchmark for Speech Foundation Models

TL;DR

The paper tackles the problem that Word Error Rate (WER) inadequately captures the variety of errors in automatic speech recognition, especially hallucinations where fluent yet contextually unfounded transcriptions resemble valid language. It introduces SHALLOW, a four-dimensional benchmark that separately assesses lexical, phonetic, morphological, and semantic hallucinations with targeted metrics, enabling fine-grained diagnostics of ASR systems. Through extensive evaluation across diverse architectures and datasets, SHALLOW reveals distinct error profiles not visible through WER, including how model design, acoustic conditions, and domains influence specific failure modes. The framework is validated on synthetic and real data, supports targeted improvements in high-stakes applications like healthcare or law, and offers reproducible protocols and open resources for the community.

Abstract

Hallucinations in automatic speech recognition (ASR) systems refer to fluent and coherent transcriptions produced by neural ASR models that are completely unrelated to the underlying acoustic input (i.e., the speech signal). While similar to conventional decoding errors in potentially compromising the usability of transcriptions for downstream applications, hallucinations can be more detrimental due to their preservation of syntactically and semantically plausible structure. This apparent coherence can mislead subsequent processing stages and introduce serious risks, particularly in critical domains such as healthcare and law. Conventional evaluation metrics are primarily centered on error-based metrics and fail to distinguish between phonetic inaccuracies and hallucinations. Consequently, there is a critical need for new evaluation frameworks that can effectively identify and assess models with a heightened propensity for generating hallucinated content. To this end, we introduce SHALLOW, the first benchmark framework that systematically categorizes and quantifies hallucination phenomena in ASR along four complementary axes: lexical, phonetic, morphological, and semantic. We define targeted metrics within each category to produce interpretable profiles of model behavior. Through evaluation across various architectures and speech domains, we have found that SHALLOW metrics correlate strongly with word error rate (WER) when recognition quality is high (i.e., low WER). Still, this correlation weakens substantially as WER increases. SHALLOW, therefore, captures fine-grained error patterns that WER fails to distinguish under degraded and challenging conditions. Our framework supports specific diagnosis of model weaknesses and provides feedback for model improvement beyond what aggregate error rates can offer.

Figures (7)

• Figure 1: SHALLOW benchmark, with its four dimensions: lexical, phonetic, morphological, and semantic scores.
• Figure 2: t-SNE projection of SHALLOW metrics, synthetic data.
• Figure 3: Spearman Correlation of SHALLOW metrics across different WER values.
• Figure 4: Distribution of hallucination scores across categories for each SHALLOW metric. Each subplot shows box plots of metric values per hallucination type. For most metrics, scores peak in their intended category. The PF metric is lowest on phonetic samples, reflecting successful detection of phonetic proximity.
• Figure 5: Spearman correlation of hallucination scores, synthetic data.