Pith. sign in

REVIEW 3 major objections 5 minor 20 references

Identifying Hearing Difficulty Moments in Conversational Audio

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Prompted audio language models can detect Hearing Difficulty Moments in conversational audio, reaching an F1 of 0.87 with ten examples, outperforming both a fine-tuned Wav2Vec 2.0 classifier (0.76) and an ASR hotword heuristic (0.39).

desk verdict A useful, narrow empirical paper with a plausible but not fully pinned-down headline result: prompted audio LMs beat fine-tuned Wav2Vec on a newly defined hearing-difficulty detection task, subject to unverified assumptions about Gemini's training data. read the letter →

arxiv 2507.23590 v1 pith:VXQFEXVI submitted 2025-07-31 cs.SD eess.AS

Jack Collins , Adrian Buzea , Chris Collier , Alejandro Ballesta Rosen , Julian Maclaren , Richard F. Lyon , Simon Carlile This is my paper · ORCID
classification cs.SDeess.AS
keywords hearingdifficultymomentsaudiolanguagemodelsfew-shotpromptingdialogueactclassificationsignal-non-understandingconversationalassistivetechnologycontinuous
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper defines a Hearing Difficulty Moment as a moment when a conversational participant is struggling to understand what was said, and asks whether such moments can be detected continuously from audio alone. Using 298 hand-filtered 'signal-non-understanding' utterances drawn from Switchboard and MRDA, the authors compare an ASR hotword heuristic, a fine-tuned Wav2Vec 2.0 classifier, and prompted or LoRA-fine-tuned audio language models. Their central finding is that a prompted Gemini 1.5 Pro model, given four seconds of preceding audio and just ten labeled examples, reaches an average F1 of 0.87, clearly above Wav2Vec 2.0's 0.76 and the heuristic's 0.39. Removing the audio and giving the model only transcripts collapses performance to the heuristic level, which the authors take as evidence that non-semantic acoustic cues carry the signal. The paper argues this opens a path toward realtime hearing assistance and passive screening of hearing difficulty in daily life.

What carries the argument

The load-bearing mechanism is few-shot prompting of a multimodal audio language model with a disposable classification vocabulary: the model is told to answer only with the tokens 'P' or 'N', and the softmax over those two next-token log probabilities becomes a continuous hearing-difficulty confidence signal for each four-second audio window. The prompt deliberately teaches both semantic cues (clarification phrases like 'What?' or 'Sorry?') and non-semantic Lombard-effect cues (raised pitch, energy shift, prolonged vowels, strained voice), and the paper attributes the audio-versus-text gap to those acoustic cues. The supporting evaluation machinery is Monte Carlo cross-validation over five splits with conversation-level leakage prevention and a ten-to-one negative-to-positive ratio.

What would settle it

Run the same 10-shot protocol on a held-out set of conversation audio that can be shown not to be in the audio language model's pretraining corpus, such as newly recorded dyads with self-reported hearing difficulty; if F1 falls from 0.87 toward the Wav2Vec level, the reported gap is explained by memorization rather than cross-modal reasoning.

Watch

Extended reading notes

Core claim

The paper's central claim is that general-purpose audio language models, prompted or lightly fine-tuned, can detect Hearing Difficulty Moments in conversational audio better than both a transcription-based hotword heuristic and a fine-tuned state-of-the-art ASR model. The authors construct the detection task by mapping DAMSL 'signal-non-understanding' act tags from 1,200 Switchboard and MRDA conversations onto four-second audio frames, then measure F1 with Monte Carlo cross-validation. Gemini 1.5 Pro in a 10-shot prompting configuration attains F1 0.87, with 2-shot at 0.85 and 0-shot at 0.75; Wav2Vec 2.0 transfer learning attains 0.76, Gemini 2.0 Flash with LoRA 0.77, and the ASR hotword heuristic 0.39. A text-only variant of Gemini 1.5 Pro also scores 0.39, which the paper interprets as showing that the uplift depends on the audio modality and its non-semantic cues rather than on keywords in the transcript.

Load-bearing premise

The comparison assumes the evaluated audio language models were not pre-trained on the exact Switchboard and MRDA conversations used in the test splits, and it further assumes the filtered signal-non-understanding labels genuinely mark hearing difficulty rather than other kinds of confusion.

Editorial extensions

If this is right

  • A real-time hearing-aid pipeline could take the last four seconds of audio, compute the 'P'-token log probability at one-second hops, and trigger noise reduction or speech enhancement when the signal crosses a threshold.
  • Transcript-only methods appear to cap at the hotword heuristic's F1 of 0.39, so any practical deployment needs access to raw audio, not just ASR output.
  • With only ten labeled examples, a large audio language model already surpasses a task-specific fine-tuned ASR model, implying the marginal cost of adding new hearing-difficulty labels is low.
  • LoRA fine-tuning a smaller audio language model reaches F1 0.77, so the approach is not confined to the largest models, though it still trails the 10-shot larger model.
  • Because positive events are much rarer in real conversation than in the 1:10 test set, serving systems will need threshold adjustment or voice-activity-detection gating to control false positives.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • Going beyond the paper: the jump from 0.75 at 0-shot to 0.85 at 2-shot suggests one positive and one negative example suffice to teach a large audio language model a new perceptual category; the same P/N token trick could be tried for other rare conversational states such as confusion, surprise, or alignment.
  • Going beyond the paper: if the 0.87 F1 survives a contamination audit, the result implies that general-purpose audio language models can substitute for task-specific fine-tuned classifiers in low-resource audio event detection, which would change dataset construction from collecting thousands of hours to collecting a few dozen examples.
  • Going beyond the paper: because the text-only version collapses to 0.39, live-captioning and transcript-based assistive systems are likely blind to hearing difficulty that is expressed acoustically before any clarification phrase, so a practical hearing assistant would need to monitor microphone audio directly.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper introduces a new task, Hearing Difficulty Moment detection in conversational audio, defined as identifying utterances where a speaker struggles to understand what was said. The authors compare four approaches: an ASR-hotword heuristic, fine-tuned Wav2Vec 2.0, prompted Gemini 1.5 Pro in 0-shot, 2-shot, and 10-shot configurations, and LoRA fine-tuned Gemini 2.0 Flash, all evaluated on a dataset derived from the Switchboard and MRDA corpora with 298 manually filtered positive examples. They report average F1 across 5 Monte Carlo cross-validation splits, with Gemini 1.5 Pro 10-shot achieving 0.87 versus 0.76 for Wav2Vec and 0.39 for the baseline. The authors conclude that audio language models, through multimodal reasoning, excel at this task and that non-semantic acoustic cues are key, as the text-only variant performs at baseline level.

Significance. If the empirical comparison holds, the paper makes a useful contribution by defining a practically motivated audio event detection task and demonstrating that general-purpose audio language models can perform it with few-shot prompting, avoiding task-specific fine-tuning. The work is clearly written and includes several good practices: conversation-level train/test splits to prevent leakage, Monte Carlo cross-validation, a continuous-time labeling scheme, precision-recall analysis, and an explicit discussion of real-world limitations such as class imbalance and always-on deployment. However, the significance is currently constrained by three issues: the evaluation rests on a small, subjectively filtered set of 298 positive examples; the central comparison uses closed models on public corpora without a contamination audit; and the statistical reporting is too thin to support the abstract's claim of significant outperformance. These gaps are fixable and do not undermine the internal logic of the task formulation, but they must be addressed before the central claim can be accepted.

major comments (3)
  1. [Section 3.1 / Section 4, Table 1] The central claim of superiority for Gemini models is vulnerable to pretraining contamination. SWDA and MRDA are standard, publicly available dialogue-act corpora, and Gemini 1.5 Pro and 2.0 Flash are closed models with undisclosed training data. If these models memorized the audio or transcripts of these corpora, the few-shot F1 scores of 0.85 and 0.87 could reflect retrieval of memorized examples rather than generalizable hearing-difficulty detection. The text-only 0-shot result (0.39) provides some evidence against simple transcript memorization, but it does not rule out audio-side memorization or prompting-triggered recall. I request a contamination analysis: for example, an evaluation on a held-out corpus not plausibly in the pretraining distribution (e.g., newly recorded or non-public conversational audio), a membership-inference test, or a comparison against an open-weights audio language model with known pretraining data. Without such evidence, the abstract's claim that audio language models "excel at this task" is not empirically established.
  2. [Section 4, Table 1 and statistical claims] The statistical reporting is insufficient to support the headline comparison. Table 1 reports only mean F1 over five MCCV splits, with no standard deviations, confidence intervals, or per-split values. The asterisks mark only two comparisons (Gemini 1.5 Pro audio 0-shot vs. text-only, and 2-shot vs. the preceding method), yet the abstract states that audio language models "significantly outperforming" the Wav2Vec baseline; the 10-shot versus Wav2Vec contrast (0.87 vs. 0.76) is not marked as significant. I ask for per-split results, confidence intervals for all methods, and a clear statement of which pairwise differences are statistically significant after correction for multiple comparisons. The decision threshold used to compute F1 must also be specified for each method, since the reported F1 values are meaningless without it.
  3. [Section 3.1, dataset construction] The manual filtering of 522 "signal-non-understanding" utterances down to 298 hearing-difficulty examples is a subjectively defined step with no validation. The paper does not provide the exclusion criteria beyond examples such as "non-understanding due to semantics," nor any inter-annotator agreement. This step defines the target concept itself, so its reliability directly affects the meaningfulness of all downstream F1 scores. I request a detailed annotation rubric and, ideally, a second annotator's labels on a subset to quantify agreement. This is not a circularity problem, but it is a validity concern for the proposed task definition.
minor comments (5)
  1. [Section 2.3] The description of how Gemini's next-token logits are converted into a continuous "P" probability is underspecified; Figure 1 shows a green curve, but the scaling, smoothing, and thresholding are not defined. Please add a precise algorithmic description.
  2. [Section 2.4] The LoRA fine-tuning description reports rank 8 and learning rate multiplier but omits other hyperparameters essential for reproduction, such as target modules, dropout, and optimizer settings. Please provide a complete configuration or a link to a released configuration file.
  3. [Section 3.1] The paper says the dataset comprises "1,200 long conversations" from SWDA and MRDA, but both corpora contain 1,199 conversations each; please state the exact number of conversations used after any filtering, and clarify how many come from each corpus and how speaker overlap across conversations is handled.
  4. [Section 1] The claim that "no prior work that we could find has explored the use of audio input for dialogue act prediction" is strong and may be outdated; consider softening it to "to our knowledge" and citing recent audio-language-model work on dialogue understanding.
  5. [Section 5, paragraph on imbalance] The discussion correctly notes that the 1:10 training ratio is much more balanced than the estimated 1:1000 occurrence rate in natural conversation. It would strengthen the paper to report precision at the expected serving distribution, or an operating-point analysis, rather than only balanced F1.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: all reported F1 scores are evaluated against external human-annotated dialogue-act labels, and no equation, fitted parameter, or self-citation chain reduces the predictions to the inputs.

full rationale

The paper's derivation chain is empirical rather than formal: it defines Hearing Difficulty Moments via human-annotated 'signal-non-understanding' act tags from SWDA and MRDA, then compares the F1 of several classifiers on held-out conversations. No quantity is defined in terms of the target result, and no fitted parameter is renamed as a prediction. The few-shot prompting examples are drawn from training conversations and tested on conversation-disjoint splits, which is standard practice rather than circularity. The prompt's mention of words like 'What?' and 'pardon' is a task description, not an output that is forced by construction, and the text-only control (F1 0.39) shows that simply seeing the same cues does not guarantee high performance. The main validity threat is pretraining contamination of the closed Gemini models on the public evaluation corpora, but that is an external-validity concern about generalization, not an internal circular-reasoning defect. Because the load-bearing claim is an empirical comparison against independent labels, the paper contains no significant circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The central claim rests on seven modeling and sampling choices and five domain assumptions, most importantly the validity of dialogue-act labels as hearing difficulty and the absence of pretraining contamination. No new physical entities are introduced.

free parameters (7)
  • Audio context window = 4 seconds
    Chosen by authors to capture acoustic context; follows sound classification practice but no context-length sweep is reported.
  • Minimum positive event onset = 0.4 seconds
    Positive timesteps are sampled only after the event has been occurring for at least 0.4 seconds, a hand-set threshold affecting positive sampling.
  • Negative sampling ratio = 10:1 negative to positive
    Evaluation uses ten random negatives per positive, which differs from the estimated 1-in-1000 real-world rate; authors acknowledge this affects false-positive rates.
  • Number of MCCV splits = 5
    Monte Carlo cross-validation with 5 splits; more splits would give tighter variance estimates.
  • Wav2Vec training hyperparameters = lr 1e-5, batch size 8, 30-50 epochs
    Hand chosen with no reported systematic search; the Wav2Vec F1 of 0.76 depends on these choices.
  • Gemini 2.0 Flash LoRA settings = rank 8, lr multiplier 0.5, 14 epochs
    Fixed by authors; the LoRA fine-tuning result (F1 0.77) may depend on these choices.
  • Few-shot example count and balance = 2-shot and 10-shot, balanced P/N
    Few-shot performance scales with these counts; the strongest result is at 10-shot.
assumptions (5)
  • domain assumption Signal-non-understanding dialogue act tags in SWDA and MRDA, after manual filtering, are valid proxies for hearing difficulty moments.
    Section 3.1: labels are derived from act tags and manually refined by excluding semantic non-understanding; the mapping is subjective.
  • domain assumption Four seconds of preceding audio is sufficient context to detect a hearing difficulty moment.
    Section 3.1: fixed window chosen from sound-classification practice; no context-length sweep.
  • domain assumption Gemini models have no prior exposure to the evaluation corpora.
    Implicit in Section 4 comparisons; no contamination audit is provided.
  • standard math Monte Carlo cross-validation with the Nadeau-Bengio corrected t-test gives valid statistical inference for these comparisons.
    Sections 3.2 and 4 rely on standard statistical methodology, though fold-dependence assumptions are not examined.
  • domain assumption Audio-only prosodic cues, such as the Lombard effect, generalize across environments and speakers.
    Section 2.3 prompt and Section 5; authors acknowledge limited environmental variance in training data.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Identifying Hearing Difficulty Moments in Conversational Audio." pith.science (2026). https://pith.science/paper/VXQFEXVI

@misc{pith2026250723590,
  author       = {Pith},
  title        = {Pith review of: Identifying Hearing Difficulty Moments in Conversational Audio},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/VXQFEXVI}},
  note         = {Machine review of arXiv:2507.23590}
}
read the original abstract

Individuals regularly experience Hearing Difficulty Moments in everyday conversation. Identifying these moments of hearing difficulty has particular significance in the field of hearing assistive technology where timely interventions are key for realtime hearing assistance. In this paper, we propose and compare machine learning solutions for continuously detecting utterances that identify these specific moments in conversational audio. We show that audio language models, through their multimodal reasoning capabilities, excel at this task, significantly outperforming a simple ASR hotword heuristic and a more conventional fine-tuning approach with Wav2Vec, an audio-only input architecture that is state-of-the-art for automatic speech recognition (ASR).

Figures

Figures reproduced from arXiv: 2507.23590 by the authors.

Figure 1
Figure 1. Example of the continuous output from the Gemini 1.5 Pro 10-shot [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Hearing Difficulty Moments are propagated from the utterance level [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Distribution of lengths of Hearing Difficulty Moments in the dataset. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Examples consist of 4 seconds of sampled context audio under the [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Precision-recall curves compared from a single fold for Wav2Vec and [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

20 extracted references · 14 canonical work pages

  1. [1]

    Geneva, Switzer- land: WHO, 2021, pp

    World Health Organization, World Report on Hearing . Geneva, Switzer- land: WHO, 2021, pp. 1, 139. [Online]. Available: https://www.who.int/ publications/i/item/9789240020481

    arXiv 2021

  2. [2]

    Dialogue act classification using a Bayesian approach,

    S. Grau, E. Sanchis, M. J. Castro, and D. Vilar, “Dialogue act classification using a Bayesian approach,” inProc. 9th Int. Conf. Speech and Computer (SPECOM), St. Petersburg, Russia, 2004, pp. 495–499. [Online]. Available: https://www. isca-archive.org/specom_2004/grau04_specom.html

    work page 2004

  3. [3]

    Dialogue Act Recognition inSynchronousandAsynchronousConversations,

    M. Tavafi, Y. Mehdad, S. Joty, G. Carenini, and R. Ng, “Dialogue Act Recognition inSynchronousandAsynchronousConversations,” in Proc. 14th Annu. Meet. Spe- cial Interest Group on Discourse and Dialogue (SIGDIAL), Metz, France, 2013, pp. 117–121. [Online]. Available:https://aclanthology.org/W13-4017

    work page 2013

  4. [4]

    Sequential Short-Text Classification with Re- current and Convolutional Neural Networks,

    J. Y. Lee and F. Dernoncourt, “Sequential Short-Text Classification with Re- current and Convolutional Neural Networks,” inProc. 2016 Conf. North Amer- 9 ican Chapter Assoc. Comput. Linguistics: Human Lang. Technol. (NAACL- HLT), San Diego, CA, USA, 2016, pp. 515–520. [Online]. Available: https: //aclanthology.org/N16-1062/

    work page 2016

  5. [5]

    Dialogue Act Classification in Domain-Independent Conversations Using a Deep Recurrent Neural Network,

    H. Khanpour, N. Guntakandla, and R. D. Nielsen, “Dialogue Act Classification in Domain-Independent Conversations Using a Deep Recurrent Neural Network,” in Proc. Interspeech, Osaka, Japan, 2016, pp. 2012–2021. [Online]. Available:https: //aclanthology.org/C16-1189

    work page 2016

  6. [6]

    Using Context Information for Dialog Act ClassificationinaDNNFramework,

    Y. Liu, K. Han, Z. Tan, and Y. Lei, “Using Context Information for Dialog Act ClassificationinaDNNFramework,” in Proc. Interspeech, Copenhagen, Denmark, 2017, pp. 2170–2178, doi: 10.18653/v1/D17-1231

    work page doi:10.18653/v1/d17-1231 2017

  7. [7]

    Hearing Loss Detection From Facial Expressions in One- On-One Conversations,

    Y. Yin et al., “Hearing Loss Detection From Facial Expressions in One- On-One Conversations,” in Proc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), Seoul, Korea, 2024, pp. 5460–5464, doi: 10.1109/I- CASSP48485.2024.10446324

    arXiv 2024

  8. [8]

    Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages,

    Y. Zhanget al., “Google USM: Scaling Automatic Speech Recognition Beyond 100 Languages,” arXiv preprint arXiv:2303.01037, 2023. [Online]. Available:https: //arxiv.org/abs/2303.01037

    arXiv 2023

Show all 20 references
  1. [9]

    wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations,

    A. Baevski, H. Zhou, A. Mohamed, and M. Auli, “wav2vec 2.0: A Framework for Self-Supervised Learning of Speech Representations,” in Proc. Conf. Neural Inf. Process. Syst. (NeurIPS), Virtual Event, 2020, pp. 12449– 12460. [Online]. Available: https://proceedings.neurips.cc/pape...

    2020

  2. [10]

    Librispeech: An ASR corpus based on public domain audio books,

    V. Panayotov, G. Chen, D. Povey, and S. Khudanpur, “Librispeech: An ASR corpus based on public domain audio books,” inProc. IEEE Int. Conf. Acoust., Speech Signal Process. (ICASSP), South Brisbane, QLD, Australia, 2015, pp. 5206–5210, doi: 10.1109/ICASSP.2015.7178964

    2015

  3. [11]

    Gemini 1.5: Unlocking multimodal understand- ing across millions of tokens of context,

    Gemini Team, Google, “Gemini 1.5: Unlocking multimodal understand- ing across millions of tokens of context,” Google AI Technical Re- port, 2024. [Online]. Available: https://storage.googleapis.com/deepmind- media/gemini/gemini_v1_5_report.pdf

    2024

  4. [12]

    LoRA:Low-RankAdaptationofLargeLanguageModels,

    E.J.Hu et al., “LoRA:Low-RankAdaptationofLargeLanguageModels,” in Proc. Int. Conf. Learn. Represent. (ICLR), Virtual Event, 2022. [Online]. Available: https://arxiv.org/pdf/2106.09685

    2022 arXiv

  5. [13]

    Our next-generation models are here: Gemini 2.0 and Gemini 2.0 Flash,

    D. Hassabis and S. Pichai, “Our next-generation models are here: Gemini 2.0 and Gemini 2.0 Flash,” The Keyword (Google Blog), 2024. [Online]. Avail- able: https://blog.google/technology/google-deepmind/google-gemini-ai-update- december-2024

    2024

  6. [14]

    Switchboard SWBD-DAMSL Shallow- Discourse-Function Annotation Coders Manual, Draft 13,

    D. Jurafsky, E. Shriberg, and D. Biasca, “Switchboard SWBD-DAMSL Shallow- Discourse-Function Annotation Coders Manual, Draft 13,” Inst. of Cognitive Sci., Univ. of Colorado, Boulder, CO, USA, Tech. Rep. 97-02, 1997. [Online]. Available: https://web.stanford.edu/~jurafsky/ws97/...

    1997

  7. [15]

    The ICSI Meeting Recorder Dialog Act (MRDA) Corpus,

    E. Shriberg, R. Dhillon, S. Bhagat, A. Stolcke, and K. Wurtz, “The ICSI Meeting Recorder Dialog Act (MRDA) Corpus,” in Proc. HLT-NAACL 2004: Short Papers, Boston, MA, USA, 2004, pp. 97–100. [Online]. Available:https: //aclanthology.org/W04-2319/

    2004

  8. [16]

    Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech,

    A. Stolckeet al., “Dialogue Act Modeling for Automatic Tagging and Recognition of Conversational Speech,”Computational Linguistics, vol. 26, no. 3, pp. 339–374,

  9. [17]

    Coding Dialogs with the DAMSL Annotation Scheme,

    M. Core and J. Allen, “Coding Dialogs with the DAMSL Annotation Scheme,” in Working Notes of the AAAI Fall Symp. on Communicative Action in Humans and 10 Machines, Boston, MA, USA, 1997, pp. 28–35. [Online]. Available:https://www. eecis.udel.edu/~carberry/CIS-885/Papers/Core-A...

    1997

  10. [18]

    A Dataset and Taxonomy for Urban Sound Research,

    J. Salamon, C. Jacoby, and J. P. Bello, “A Dataset and Taxonomy for Urban Sound Research,” in Proc. 22nd ACM Int. Conf. Multimedia, New York, NY, USA, 2014, pp. 1041–1044, doi: 10.1145/2647868.2655045

    2014

  11. [19]

    Inference for the Generalization Error,

    C. Nadeau and Y. Bengio, “Inference for the Generalization Error,”Mach. Learn., vol. 52, no. 3, pp. 239–281, 2003, doi: 10.1023/A:1024068626366 11

    2003 doi

  12. [2000]

    Available:https://aclanthology.org/J00-3003

    [Online]. Available:https://aclanthology.org/J00-3003

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.