PaperlayerView exact revision on arXiv

Source-linked AI summary

The Anatomy of an ASR Hallucination

Hamees Sayed, Apoorv Singh, Kumar Aman, Akshat Mandloi

arXiv:2609.04404v1cs.CL

TL;DR

ASR hallucinations raise the question of where transcript output stops being adequately constrained by audio. The paper combines layerwise measurements and interventions in two independently trained Conformer-Large recognizers; across evaluated models and shifts, it identifies a terminal-stage dependency for grounded output, while the intervention outputs are not fluent fabrications.

  • Problem

    The paper asks where along an ASR network output stops being grounded in continuous acoustic evidence.

    Source: §1 · p. 1

  • Method

    The authors compare two Conformer-Large recognizers under environmental and L1-background shifts using skip-layer interventions, representation measurements, and probes of head-readability and grapheme information.

    Source: §6 · p. 9

  • Results

    Across evaluated models and shifts, bypassing the final block raises DR by +94 to +99 pp (CTC) and +42 to +88 pp (RNN-T), versus ≤8.5 pp for middle-block skips.

    Source: §1 · p. 2, §6 · p. 9

  • Takeaways & Limitations

    The findings identify terminal-stage grounding as a mechanistic precondition for hallucination, not the complete causal origin of naturally occurring fluent fabrications.

    Source: §7 · p. 9

    The results concern two English Conformer-Large checkpoints and do not establish that the localization generalizes to other architectures, languages, or training pipelines.

    Source: §8 · p. 9

Abstract

from arXiv · show

ASR systems sometimes produce fluent text that is unrelated to the speech they receive. We view these hallucinations as one possible consequence of a broader grounding failure, in which the transcript is no longer adequately guided by the audio. To understand where this failure becomes possible, we study two independently trained Conformer-Large recognizers - one CTC and one RNN-T - under environmental degradation and speaker-background shift. In both models, the final encoder stage emerges as a critical boundary: bypassing the final block causes divergence on nearly every utterance, whereas bypassing middle blocks has little effect. At this same stage, the representations become more compact, text becomes readable by the trained decoder, and grapheme information becomes explicit. Importantly, the intervention produces garbled or repetitive output rather than fluent fabrication. Our result therefore identifies a mechanistic precondition for hallucination - the failure to produce adequately grounded output - not the complete origin of naturally occurring hallucinations. Together, the results reveal a consistent terminal-stage dependency for grounded recognition across two decoder families and multiple distribution shifts.

1 Introduction

The paper frames semantic hallucination as one possible manifestation of broader grounding failure and asks where ASR output stops being constrained by audio. Across two decoder families, skip-layer interventions localize a critical dependency to the terminal encoder stage, while related analyses characterize that stage and the intervention outputs.

1 Introduction

  • Grounding failure is loss of adequate acoustic constraint; semantic hallucination is its fluent, intelligible subset, while repetition, empty output, and garbling are other failures.The interventions test a mechanistic precondition for hallucination, not whether they themselves produce fluent fabrication.Source: §1 · p. 1
  • Bypassing the final encoder block raises divergence rate by +94 to +99 pp (CTC) and +42 to +88 pp (RNN-T), while bypassing blocks in L=2–13 moves it by ≤8.5 pp.WER rises in parallel under final-block bypass.Source: §1 · p. 2
  • Across the final one to two blocks, effective rank drops 18–44%, grapheme identity becomes linearly readable, and the trained head transitions to usable text as acoustic properties decay.The CTC grapheme probe rises from 0.81 to 0.99.Source: §1 · p. 2
  • With content-heterogeneous batches, replacing the RNN-T predictor output with its batch mean yields DR of 83–92% and WER of 367–512%, producing repetition collapse rather than fluent fabrication.Corpus-order batches can be content-homogeneous, allowing each utterance’s content to leak through the batch mean.Source: §1 · p. 2
  • A strict LLM judge confirms fluent off-topic fabrication in only 0.005 (CTC) and 0.069 (RNN-T) of flagged baseline outputs; intervention outputs are classified into five semantic outcome categories.The categories are fabrication, on-topic error, repetition/collapse, empty output, and other failure.Source: §1 · p. 2

2 Method

The study uses two independently trained Conformer-Large recognizers and evaluates them on environmental and speaker-background shifts. It combines divergence and WER measurements with probes, representation analyses, and interventions to examine layerwise grounding and head-readiness.

2.1 Models

  • The two pretrained Conformer-Large checkpoints are independently trained CTC and RNN-T models with different depths, objectives, and partially different training data; cross-model agreement is treated as replication, not proof of a shared encoder cause.The models have 18 and 17 encoder blocks, respectively.Source: §2.1 · p. 2

2.2 Datasets and decoding

  • WildASR supplies environmental-degradation and accent splits, while L2-ARCTIC supplies clean read English from speakers across six L1 backgrounds; decoding is greedy and uses normalized text.Causal and lens experiments use 200 utterances per WildASR split and 80 per L1 for the L2-ARCTIC lens.Source: §2.2 · pp. 2–3

2.3 The divergence flag and its rate (DR)

  • The deterministic divergence rate flags utterances through insertion, complete 4-gram divergence, or combined length and character-error thresholds, but it does not assess fluency.The authors retain the name DR because validation finds the flag broader than strict hallucination.Source: §2.3 · p. 3

2.4 Analysis toolkit

  • The encoder lens substitutes a layer’s post-normalization output into the trained decode path to measure layerwise WER; the comparable RNN-T quantity is its across-layer discontinuity, not absolute WER.The RNN-T substitution path carries measurable machinery overhead.Source: §2.4 · pp. 3–4
  • Effective rank estimates the number of active variance directions, while participation ratio and TwoNN serve as robustness checks.Per-utterance effective rank is length-sensitive and reported only in the repository.Source: §2.4 · p. 4
  • Skip-layer ablation replaces a block’s output with its input to test necessity, while the transducer predictor ablation replaces predictor outputs with zeros or a batch-channel mean.The batch mean retains a residual batch-averaged signal while discarding each utterance’s own emission history.Source: §2.4 · p. 4
  • Linear CKA compares matched-layer representations, while linear probes test whether silence, shift type, graphemes, SNR, and L1 background are readable from layer outputs.The grapheme target uses the CTC model’s final-layer argmax as a frame-aligned proxy.Source: §2.4 · pp. 3–4

3 Divergence behaviour and metric validation

Across seven WildASR conditions, both models show a shared difficulty ordering, while cross-model disagreement rises with environmental degradation. Validation shows the divergence flag mostly captures severe grounded errors rather than fluent fabrication.

3.1 Baseline decoding behaviour

  • 0.3–2.6% baseline DR across conditions; both models find clean and phone-codec easiest, with clipping, noise-gap, far-field, and reverberation hardest.Source: §3.1 · p. 4, Table 1 · p. 5
  • 66.9% clean to 95.3% noise-gap cross-model disagreement, while accent disagreement is 40.8%.The table reports per-condition mean WER, DR, and hypothesis disagreement on the full WildASR splits.Source: §3.1 · p. 4, Table 1 · p. 5

3.2 What does the flag actually measure?

  • 184/185 CTC and 111/116 RNN-T flags came from rule (b); exceptions include repetition collapse and one-word references.The authors treat rule (b) as the operative detector.Source: §3.2 · p. 4
  • Only 1/185 CTC and 8/116 RNN-T flagged outputs were judge-confirmed fluent fabrications; flags mostly captured short-reference on-topic errors.Illustrative cases include repetition and missed fluent topic shifts that retain some 4-gram overlap.Source: §3.2 · p. 4, Table 2 · p. 5
  • DR or WER increases alone do not establish semantic hallucination; human-annotated intervention outputs remain future work.Source: §3.2 · p. 5

4 Localising the failure on WildASR

On WildASR, bypassing terminal encoder blocks sharply raises divergence, while middle-block skips have little effect; judged outputs are garbled or repetitive rather than fluent fabrications. The terminal region also compresses representations and first yields head-readable text and rising grapheme readability, while predictor-ablation effects depend on batch composition.

4.1 Causal skip-layer ablation

  • CTC L=17 skips raise DR +94 to +99 pp and RNN-T final-block skips +42 to +88 pp; middle-band skips move DR by at most 8.5 pp.The terminal-stage effect is distinct from a generic increase in recognition errors.Source: §4.1 · p. 5

4.2 What do the intervention outputs actually contain?

  • No intervention output was judged a fluent fabrication; final-block skips yielded CTC garbling (87% other-failure) and RNN-T repetition (98% collapse).These outcomes identify a failure to produce grounded output, not the complete origin of naturally occurring hallucinations.Source: Table 3 · p. 6, §4.2 · p. 6

4.3 Effective rank: the terminal compression

  • Effective rank peaks at L=15 then drops 24–41% (CTC) and 18–44% (RNN-T); CTC compression spans L=16–17, while RNN-T concentrates it in the final block.Far-field and reverberation conditions sit well below other conditions from mid-stack; their peak gaps are 37–41 ER for CTC and 77–86 ER for RNN-T.Source: Figure 2 · p. 6, §4.3 · p. 6

4.4 Encoder lens: when the head can read the representation

  • The trained head yields no usable text before the terminal output: CTC remains garbled (WER 108–515%, DR ≥97%), while RNN-T becomes usable at L=16.The lens locates where the trained head can read text, not by itself where information is lost; RNN-T substitution adds WER 2.3–12.5 pp over production decoding.Source: §4.4 · pp. 6–7

4.5 Probes: acoustic to symbolic

  • Shift-type readability peaks at L=1 (0.978 CTC, 0.981 RNN-T) then falls to 0.38 and 0.42 at the final layer, while grapheme readability rises with depth.CTC grapheme accuracy jumps from 0.811 at L=15 to 0.919 at L=16 and 0.990 at L=17; the grapheme target is a model-derived proxy, not ground truth.Source: §4.5 · p. 7

4.6 Predictor ablation and batch-composition sensitivity

  • With shuffled content-heterogeneous batches, batch-mean replacement yields DR 83–92% and WER 367–512% across conditions; corpus-order homogeneity leaks utterance content.The injected mean retains batch-averaged information, so it does not isolate the predictor’s internal-LM contribution; zero/blank replacements saturate at DR 94–100% and WER ∼3,000%.Source: Table 4 · p. 7, §4.6 · pp. 7–8

5 Generalization to L1-background shift

On clean read speech, L1 background affects recognition difficulty, while the depth at which usable text appears remains stable across accent groups. Accent and environmental shifts leave different representational signatures.

5.1 Per-accent decoding

  • WER spans 2.4 (CTC, Hindi) to 18.1 (CTC, Vietnamese), a 7.5× spread from L1 background alone, while DR ranges 0.7–18.8%.Hindi is easier for both models than WildASR-clean, and the models agree on the easiest and hardest accents while differing in mid-ranking order.Source: §5.1 · p. 8

5.2 The terminal-stage discontinuity per accent

  • Across all eight groups and both models, only the final layer yields usable lens text, with final-layer lens WER tracking accent difficulty (CTC 2.1–20.5; RNN-T 7.2–39.4).The location of the discontinuity is stable across groups, while its fidelity varies with accent; RNN-T final-layer lens WER should be compared across groups, not to Table 5.Source: Figure 3 · p. 8, §5.2 · p. 8
  • Accent identity is most decodable at L=4–8 (CTC up to 0.97; RNN-T 0.98) and remains readable at the CTC final layer (0.70; chance 1/8), unlike the environmental-shift probe peaking at L=1.The eight-way probe includes US-native and L2-European comparison groups with different corpora and recording conditions, so their separability partly reflects channel.Source: §5.2 · p. 8

5.3 Two shift types, two encoder responses

  • Unlike spectral distortion’s rank concentration, accent leaves effective-rank profiles nearly unchanged while remaining linearly decodable deep in the encoder; the authors make no claim about representational rotation.At the peak layer, effective-rank spread is ≈28 ER for CTC (153–181) and ≈17 for RNN-T (227–244), versus 37–86-point condition gaps under spectral distortion.Source: §5.3 · pp. 8–9

6 Discussion

The terminal one to two encoder blocks are where causal, geometric, lens, and probe evidence converge on loss of grounded, head-readable output. Environmental and accent shifts differ upstream, but the terminal-stage discontinuity recurs across evaluated models and shifts.

6 Discussion

  • The terminal one to two blocks define the operative boundary: bypassing them raises DR by +97/+71 pp on average, versus ≤8.5 pp for mid-band skips, alongside rank compression, a lens discontinuity, and a grapheme-probe jump.In CTC, effective rank bottoms at L=16 while head-readability arrives at L=17, and skipping L=16 already substantially raises DR.Source: §6 · p. 9
  • The predictor batch-mean ablation collapses heterogeneous-batch decoding into repetition, while corpus-order batches can leak utterance content; it neither isolates the predictor’s internal-LM contribution nor tests decoder-side mitigation.The localization claim instead rests on within-encoder evidence measured with the predictor intact.Source: §6 · p. 9
  • Environmental degradation concentrates rank, whereas L1-background shift preserves rank while leaving accent identity decodable; the terminal discontinuity is stable across evaluated cells, not claimed universally.The authors explicitly limit the claim to the evaluated models and shifts.Source: §6 · p. 9

7 Conclusion

The paper localizes a mechanistic precondition for ASR hallucination, rather than the complete origin of naturally occurring fluent hallucinations.

7 Conclusion

  • The terminal encoder stage is a recoverability boundary: bypassing it disrupts grounded output, yielding garbling or repetition, and localizes a precondition—not the full cause—of natural fluent hallucinations.The same stage compresses effective rank, first yields head-readable text, and completes the acoustic-to-symbolic transition; 0 of 354 judged outputs were fluent fabrications.Source: §7 · p. 9

8 Limitations

The evidence is scoped to two English Conformer-Large checkpoints and has limitations in measurement, sampling, interventions, and decoding conditions.

8 Limitations

  • Results cover two English Conformer-Large checkpoints from one toolkit; agreement across them does not establish the localization for other architectures, languages, sizes, or training pipelines.The checkpoints differ in objectives, depths, and partially different training data.Source: §8 · pp. 9–10
  • The deterministic 4-gram-divergence flag has baseline precision of 0.005–0.069 for fluent off-topic fabrication, population recall is uncertain, and semantic classifications rely on one LLM judge without human validation.Source: §8 · p. 9
  • Per-utterance encoder effective rank is length-sensitive and unadjusted, while causal and lens subsets use the first 200 utterances per split rather than random samples.WildASR’s consecutive degradation variants give corpus-order subsets unusual internal structure; matched or regression-controlled analyses require re-collection.Source: §8 · p. 9
  • Skip-layer ablation establishes necessity under one perturbation, and batch-mean predictor replacement retains residual batch signal; RNN-T lens WER also carries 2.3–12.5 pp substitution overhead.Greedy decoding is used throughout; tested RNN-T beam-search changes were limited to clean and reverberation conditions.Source: §8 · p. 10

9 Ethical Considerations

The authors note both dual-use concerns and fairness implications of their findings, while reporting that the work used released artifacts and involved no human subjects.

9 Ethical Considerations

  • Layer-localized findings could inform adversarial degradation of deployed models, although the interventions require white-box access.The authors also note that compression calibrated on clean, high-resource speech may disproportionately degrade non-native accents and degraded acoustic channels.Source: §9 · p. 10
  • The authors encourage compression work to evaluate on shifted and accent-diverse benchmarks because affected populations already experience worse ASR.They identify potential fairness implications when rank-based pruning is calibrated on clean, high-resource speech.Source: §9 · p. 10
  • The experiments used publicly released artifacts under their release terms, collected no new data, and involved no human subjects.The authors state that they followed the ACL Code of Ethics.Source: §9 · p. 10

A Reproducibility details

The reproducibility details specify the checkpoints, probe setup, decoding controls, and evaluation subsets used in the experiments.

A Reproducibility details

  • Experiments used two NVIDIA NeMo Conformer-Large checkpoints, with single-layer probes trained for 1,500 AdamW steps and stratified class balancing.Probe settings included learning rate 5 × 10−3, weight decay 10−4, and a cap of 8,000 frames per class; the grapheme target was the CTC model’s layer-17 argmax.Source: A Reproducibility details · p. 11
  • For the RNN-T encoder lens, the CUDA-graph greedy decoder was disabled so the substitution hook would not be silently bypassed.The method used NeMo’s change_decoding_strategy.Source: A Reproducibility details · p. 11
  • Evaluation used full splits for Tables 1 and 5, and specified subsets for lens, ablation, CKA, and beam-search analyses.The appendix also reports Whisper EnglishTextNormalizer for WER, released validation materials and scripts, and experiments run on one NVIDIA H200 GPU.Source: A Reproducibility details · p. 12

B LLM-judge validation details

The flag’s precision was assessed against a strict LLM judge, and the appendix points to layer-wise probe curves for both models.

B LLM-judge validation details

  • Rule precision was 0.005 for CTC and 0.069 for RNN-T when the LLM judge was treated as reference.The judged samples included all flagged utterances plus 700 unflagged utterances per model; population recall was reported with stratum-reweighted intervals.Source: B LLM-judge validation details · p. 12
  • The two RNN-T judge-only misses were fluent topic-changing hypotheses that retained some 4-gram overlap with the reference.The judge’s mean self-reported confidence on flag-only disagreements was 1.0.Source: B LLM-judge validation details · p. 12

C Layer-wise probe curves

  • Figure 4 shows four probe targets per encoder layer for both models.The probes are described as layer-wise curves in the appendix.Source: C Layer-wise probe curves · p. 12

D Lens WER overlaid on encoder ER

The appendix relates the terminal-stage lens and effective-rank changes to accent behavior, cross-model similarity, and alternative dimensionality measures.

D Lens WER overlaid on encoder ER

  • Lens WER first falls below 100% at the final block in both models, co-located with the ER decline; CTC’s ER decline starts at L=16, before head-readability.The CTC ER minimum precedes head-readability by one block, so the changes are described as co-located at the terminal stage, not simultaneous.Source: Figure 5 · p. 12, D Lens WER overlaid on encoder ER · p. 12

E Accent probes and per-accent effective rank

F Cross-model geometry (CKA)

  • At L=16, CKA is highest for far-field (0.69) and reverberation (0.70), with no confidence-interval overlap against other conditions.Similarity declines with depth in both independently trained models; the authors treat this as a replication observation from two runs, not a population claim or direct measure of subspace overlap.Source: Figure 8 · p. 13, F Cross-model geometry (CKA) · p. 13

G Alternative dimensionality measures: PR and TwoNN

  • PR drops 34–62% peak-to-final and tracks ER, whereas TwoNN generally increases from L=15 to the final layer; localization results are measure-independent.The authors interpret this pattern as variance concentrating into fewer directions while the point manifold retains its intrinsic dimensionality.Source: G Alternative dimensionality measures:… · p. 13, Table 7 · p. 14

Related papers

Loading 2609.04404v1…