{"id":"79712015-bf90-4335-ab11-06da44bc5301","arxiv_id":"2411.10362","paper_version":3,"verdict":"REJECT","confidence":"HIGH","novelty_score":2.0,"correctness_risk":"high","formal_verification":"none","parameter_count":0,"one_line_summary":"ASR, a large language model, and smart glasses are combined into an interaction loop model whose performance is quantified only with standard textbook metrics, with no empirical validation.","lead":"This paper proposes a conceptual interaction loop that links speech recognition, a large language model, and smart glasses to show answers in a wearable display. It provides only theoretical formulas and a toy script run, with no real data or working end-to-end system.","discovery_kind":"incremental","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own Section 5 reports a 'list index out of range' error in the LLM module, directly contradicting the claim that the implemented cycle runs successfully; without a working end-to-end run or any data, the central feasibility claim is unsupported.","rationale":"Reader's verdict is REJECT, and I agree. The most load-bearing weakness is not the additivity of latency delays (Section 3), though that is also questionable; it is that the paper's own experimental section disproves the feasibility claim. The code is openly provided, which is creditworthy, and the modular description is a reasonable conceptual sketch. However, the reported run stops at the LLM decoding stage, so the interaction cycle is never closed. The paper also concedes in the Limitation section that real-time synchronization and delay management may face technical obstacles. Because Section 4 states no data are obtainable, the formulas (WER, FER, perplexity, Dtotal) are never applied to any output; they are textbook metrics, not a validated model. The Summary's claim of a 'major advance' is therefore unsupported. My proposed test is to rerun the provided code; if the error reproduces, the central claim fails on the paper's own terms. This is an internal inconsistency, not a matter of external consensus. I partially agree with the reader: the latency additivity issue is real but secondary; the decisive problem is the failed implementation and missing evaluation.","tokens_in":10281,"tokens_out":3629,"duration_ms":31096,"concrete_test":"Execute the three scripts from the GitHub repository (ASR, LLMs, Smart Glasses) in sequence in a clean Python 3.13 environment using a short sample audio file; if the LLM script reproduces the 'list index out of range' error and no text is passed to the display module, the paper's claim of a successful end-to-end run is refuted. Alternatively, if the error does not reproduce, the claim would be partially supported, though the absence of any quantitative evaluation would remain.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that the ASR-LLMs-Smart Glasses cycle model is feasible and its performance quantifiable depends on the implemented system actually completing the interaction loop. Section 5 states that the code 'can run successfully,' but the immediately following description reports that the LLM part raised a 'list index out of range' error during decoding, and that the LLMs part 'requires further optimization of the decoding logic.' Section 4 explicitly says that no data can be obtained because the architecture does not exist in current industry or academia, so no WER, FER, perplexity, or Dtotal is ever computed for the proposed pipeline. The formulas in Section 3 are standard definitions from the literature, not evaluations of this model. Thus the manuscript's own evidence contradicts the feasibility claim: the pipeline halts at the LLM stage, and the quantitative metrics are never instantiated. This is not a disagreement with external consensus; it is an internal mismatch between the claimed successful run and the reported error.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an 'ASR-LLMs-Smart Glasses' interaction cycle model in which user speech is captured, transcribed by an automatic speech recognition pipeline, semantically processed by a large language model, and rendered on smart glasses, with the user's response closing the loop. The contribution is presented as a conceptual architecture plus mathematical formulas for accuracy (WER, FER, perplexity), coherence, and latency (Dtotal = Tfe + Tam + Tlm). The manuscript also describes a Python implementation and points to GitHub code. Section 4 explicitly states that no data can be obtained because the architecture does not yet exist in industry or academia, and Section 5 reports that the LLM module fails with a 'list index out of range' error during decoding. No measured values of WER, FER, perplexity, or latency are reported for the proposed pipeline.","tokens_in":10475,"tokens_out":4524,"duration_ms":41679,"significance":"If the framework were validated, it could serve as a useful modular decomposition of a voice-driven smart glasses system, and the public release of the code is a positive feature. I found no circularity or fitted-parameter issue: the equations are standard textbook definitions, and the latency expression is an asserted summation. However, the significance is not established by the manuscript. The central feasibility claim rests on an implementation that the paper itself reports as failing at the LLM stage, and the quantitative performance claims are never instantiated with any data. The paper is better read as a proposal or position statement than as a research article reporting validated results, and in its current form it does not support the abstract's claim that the model's performance is quantified.","major_comments":[{"comment":"The section's opening sentence claims that the code 'completely implements the connection architecture of speech recognition, automatic language generation and smart glasses display, and can run successfully,' but the immediately following description reports a 'list index out of range' error in the LLM decoding stage and states that 'the LLMs part requires further optimization of the decoding logic.' Because the LLM component is the central semantic link between ASR output and the smart glasses display, the implemented pipeline does not in fact complete the proposed interaction cycle. This internal contradiction directly undermines the paper's central feasibility claim.","section":"Section 5, Results"},{"comment":"The experiment section explicitly states that 'since this architecture is not a model that already exists and is practiced in the current industry and academia, it is impossible to analyze the model by obtaining data.' Accordingly, none of the formulas in Section 3—WER, FER, perplexity, or Dtotal—is instantiated with a measured value for the proposed pipeline. These formulas are standard definitions from the ASR and NLP literature, not evaluations of this model, so the abstract's claim that mathematical formulas are used to quantify the model's performance is not supported. The quantitative feasibility claim would require at least one end-to-end run on real or simulated speech input with reported metrics.","section":"Section 4, Experiment"},{"comment":"The total latency expression Dtotal = Tfe + Tam + Tlm is asserted without justification. It presupposes that the pipeline stages operate strictly sequentially, with no pipelining, parallel decoding, caching, or end-to-end neural models, and it omits both the LLM inference time and the smart glasses parsing/rendering time even though those stages are central to the proposed cycle. The manuscript gives no argument or measurement supporting this additive decomposition, so the quantitative latency claims do not follow even if the qualitative stage-by-stage description is accepted.","section":"Section 3, Proposed Algorithms"}],"minor_comments":[{"comment":"The section heading 'ASR-LLMs-Small Glass' appears to be a typo for 'Smart Glass'; throughout the manuscript there are numerous missing spaces and grammatical errors, such as 'This researchproposesthe interaction loopmodel' in the abstract, which should be corrected.","section":"Section 2, heading"},{"comment":"Figure 2 is referenced as showing the experimental results but does not appear in the manuscript, so the reader cannot inspect the claimed output or verify the reported behavior.","section":"Section 5, Figure 2"},{"comment":"The citation '(2017)' in Section 2.3 lacks an author and is not formatted consistently with the other citations; several other references also have incomplete or inconsistent formatting.","section":"Section 2.3, references"},{"comment":"The section describes the ASR part as having 'successfully simulated audio capture' and the smart glasses modules as simulated; the manuscript should state clearly that these are mock modules rather than a functional hardware integration.","section":"Section 5, implementation scope"},{"comment":"Section 7 introduces composite metrics named coherence (C), accuracy (A), error rate (E), delay rate (D), and efficiency (Ef), but no formulas or definitions are given for these quantities anywhere in the paper.","section":"Section 7, Summary"}],"recommendation":"reject","confidential_remarks":"The manuscript is not suitable for publication as a research article in its current form because the core feasibility claim is contradicted by the paper's own experimental report and no data are provided. The GitHub code release is a positive element, but the implementation is not a working end-to-end system. The authors could consider reframing the work as a position or vision paper, and they would need to complete a real prototype and report measured metrics before resubmitting as a research contribution."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I read the paper. My take: it is a student-level position piece, not a research result. The one experiment it reports actually fails at the LLM module, and Section 4 says no data were available, so none of the formulas are ever applied. The feasibility claim is therefore unsupported.\n\nWhat is good: the paper assembles a coherent pipeline description from the ASR, LLM, and smart glasses literatures, and it states the standard metrics (WER, FER, perplexity) correctly. It also deserves credit for reporting the \"list index out of range\" error in Section 5 rather than hiding it, and for publishing the code on GitHub. The limitation section is honest about integration challenges.\n\nThe problems: Section 5 says the code \"can run successfully\" but then describes a crash in the LLM decoding stage. Those two statements cannot both be true. Section 4 explicitly says no data can be obtained because the architecture does not exist in practice, so the mathematical formulas are never instantiated; they are just textbook definitions. The latency formula Dtotal=Tfe+Tam+Tlm assumes purely sequential stages and ignores pipelining, caching, or end-to-end models; that is a real simplification, but it is a minor concern compared to the lack of any quantitative evaluation. The summary's claim that this is \"a major advance in the field of human-computer interaction\" is not supported by the content.\n\nWho is this for? Possibly a workshop poster or a student project report. The writing is clear enough that the author has a basic understanding of the components. But as a research paper, it has no new method, no data, and no validated implementation.\n\nRecommendation: I would not send this to peer review. It does not meet the bar for a serious referee. The author should either fix the implementation, collect actual WER/FER/perplexity numbers on a real pipeline, or reframe the manuscript as a survey of ASR-LLM-glasses integration. If it is resubmitted after that, it could become a modest workshop contribution, but the current version is not a research contribution.","headline":"An honest but unsupported conceptual sketch: the only run of its own code fails at the LLM stage, so the feasibility claim does not hold.","tokens_in":10993,"tokens_out":3070,"would_cite":false,"duration_ms":27857,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper proposes an 'ASR-LLMs-Smart Glasses' interaction cycle and quantifies its accuracy, coherence, and latency with mathematical formulas.","keywords":["automatic speech recognition","large language models","smart glasses","human-computer interaction","interaction cycle model","word error rate","perplexity","latency"],"falsifier":"Run any modern streaming speech-to-text system combined with a language model and head-mounted display, and compare the measured time from speech onset to displayed output against the sum of individually timed stage durations; if the measured total is materially smaller than the sum because frames are processed while the user is still speaking or text is rendered before the model finishes, the linear latency identity fails.","tokens_in":10093,"feed_emoji":"🥽","tokens_out":9198,"duration_ms":70573,"temperature":0.7,"pith_summary":"This paper proposes a wearable interaction cycle in which a user's speech is transcribed by an automatic speech recognition (ASR) system, expanded into a response by a large language model (LLM), and shown as text on smart glasses, with the user's reaction starting the next loop. The author argues that this 'ASR-LLMs-Smart Glasses' loop is a feasible way to make human-computer interaction seamless, and that its performance can be judged before a physical product exists. To that end the paper decomposes the pipeline into stages and assigns numerical measures: signal-to-noise ratio for capture, word error rate for recognition, frame error rate and perplexity for acoustic and language modeling, and an additive latency formula for delay. The contribution is a reusable conceptual and mathematical scaffold rather than an empirical validation; the paper's own implementation is a small Python simulation in which the LLM stage fails with an index error. A sympathetic reader would take the pith to be a checklist and a set of formulas for designing and budgeting voice-driven wearable assistants.","feed_headline":"Three formulas quantify the voice-to-glasses interaction cycle","feed_subtitle":"Accuracy, coherence, and latency each get a formula for a speech-to-glasses loop.","key_machinery":"The central object is the interactive cycle model itself, a four-stage ASR chain (speech capture, feature extraction, acoustic modeling, language modeling) linked to the LLM's transformer pipeline (tokenization via byte pair encoding, embedding, positional encoding, multi-head self-attention, softmax) and to a smart glasses display pipeline (data parsing, text formatting, rendering, display). Its load-bearing identities are the formulas $WER=(S+D+I)/N$, $FER=\\text{incorrectly classified frames}/\\text{total frames}$, $\\mathrm{Perplexity}=\\left(\\prod_{i=1}^{N} 1/P(w_i \\mid w_{i-1}, \\ldots, w_1)\\right)^{1/N}$, and the additive latency law $D_{\\mathrm{total}}=T_{\\mathrm{fe}}+T_{\\mathrm{am}}+T_{\\mathrm{lm}}$. These formulas do the work of converting an architectural proposal into quantities a designer could in principle compute, budget, or optimize before building hardware.","core_discovery":"The central claim is that the interaction loop 'ASR-LLMs-Smart Glasses' is feasible as a unified human-computer interaction model, and that its performance can be quantified with mathematical formulas organized around accuracy, coherence, and latency. The paper defines the pipeline as four ASR stages—speech capture, feature extraction, acoustic modeling, and language modeling—followed by the LLM's transformer operations (input reception, byte-pair-encoding tokenization, embedding, positional encoding, multi-head self-attention, softmax decoding), then smart glasses data parsing, text formatting, rendering, and display. Accuracy is captured by $WER=(S+D+I)/N$, frame error rate, and perplexity; coherence is tied to capture clarity and perplexity; latency is reduced to $D_{\\mathrm{total}}=T_{\\mathrm{fe}}+T_{\\mathrm{am}}+T_{\\mathrm{lm}}$, the sum of feature-extraction, acoustic-model, and language-model times. The author presents these results as theoretical, intended to test and evaluate feasibility in the absence of deployed products, with a Python 3.13 IDLE implementation shared through a public repository.","pith_inferences":["Because modern speech recognizers and language models generate hypotheses incrementally, the linear delay sum $D_{\\mathrm{total}}=T_{\\mathrm{fe}}+T_{\\mathrm{am}}+T_{\\mathrm{lm}}$ is likely an upper bound on real-world latency rather than a description of it; if so, the paper's feasibility case is conservative.","The paper's metrics stop at ASR accuracy and coherence; a natural extension is to score the LLM's response relevance and the glasses' display readability, for instance with task-completion or gaze-based measures, to close the loop's evaluation.","The feedback half of the cycle—user reaction time and the next utterance—is left outside all formulas; explicit terms for those would turn the static pipeline into a repeating control loop and would give a stability condition for smooth interaction.","A testable extension would replace the four-stage ASR chain with a single end-to-end acoustic-to-text model; the paper's taxonomy would then need to collapse $T_{\\mathrm{fe}}$, $T_{\\mathrm{am}}$, and $T_{\\mathrm{lm}}$ into one learned pass, and WER plus perplexity effects would be observable."],"forward_implications":["Improving capture quality (higher SNR), acoustic modeling (lower FER), and language modeling (lower perplexity) each directly lowers the ASR word error rate $WER=(S+D+I)/N$.","The latency identity $D_{\\mathrm{total}}=T_{\\mathrm{fe}}+T_{\\mathrm{am}}+T_{\\mathrm{lm}}$ gives an additive budget: cutting any one stage's time by a fixed amount cuts total delay by that amount, so hardware accelerators can be targeted stage by stage.","Coherence, though qualitative, is proxied through perplexity, so designers can pick an LLM whose perplexity on transcribed speech is low to get more coherent display text.","The architecture's modularity means the three components can be developed and replaced independently, as the paper's separate Python scripts for ASR, LLM, and smart glasses suggest.","The model implies a full working prototype is constructible with current technology and that its performance should be measurable with the paper's formulas, making the proposal an engineering blueprint rather than a purely conceptual diagram."],"supporting_citations":[{"why":"Defines the Word Error Rate formula that anchors the accuracy metric.","marker":"Ali & Renals, 2018"},{"why":"Introduces the multi-head self-attention transformer that the LLM stage assumes.","marker":"Vaswani, 2017"},{"why":"Justifies using mathematical formulas to explore operational boundaries of LLM-based interaction models.","marker":"Bubeck et al., 2023"},{"why":"Surveys smart glasses interaction methods and defines the display-based output stage.","marker":"Lee & Hui, 2018"},{"why":"Demonstrates a prior AR-glasses-plus-ChatGPT application, supporting practicality of the combined loop.","marker":"Lee et al., 2023"},{"why":"Discusses device-directed speech detection with LLMs, grounding the speech-capture analysis.","marker":"Wagner et al., 2024"},{"why":"Compares sequence-to-sequence models for speech recognition, informing the ASR implementation.","marker":"Prabhavalkar et al., 2017"},{"why":"Defines delay rate as the time difference between speech input and text output, grounding the latency formula.","marker":"Wang et al., 2022"}],"fun_headline_variants":["Voice-to-glasses: three formulas, one loop","Quantifying the ASR-LLM-glasses interaction","Math for speech-to-smart-glasses","Theoretical formulas for a voice-driven display","Accuracy, coherence, latency: the voice-glasses trio"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole quantified argument rests on the assumption that the interaction can be split into strictly sequential stages whose delays add exactly as $D_{\\mathrm{total}}=T_{\\mathrm{fe}}+T_{\\mathrm{am}}+T_{\\mathrm{lm}}$; if real systems pipeline or overlap decoding, the delay formula no longer describes the system.","fun_headline_variants_meta":{"raw":{"variants":["Voice-to-glasses: three formulas, one loop","Quantifying the ASR-LLM-glasses interaction","Math for speech-to-smart-glasses","Theoretical formulas for a voice-driven display","Accuracy, coherence, latency: the voice-glasses trio"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000435,"raw_usage":{"total_tokens":2209,"prompt_tokens":932,"completion_tokens":1277,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":548,"completion_tokens_details":{"reasoning_tokens":1203}},"tokens_in":548,"tokens_out":1277,"duration_ms":12079,"temperature":1.0,"reasoning_tokens":1203,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T19:41:46.943072+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run any modern streaming speech-to-text system combined with a language model and head-mounted display, and compare the measured time from speech onset to displayed output against the sum of individually timed stage durations; if the measured total is materially smaller than the sum because frames are processed while the user is still speaking or text is rendered before the model finishes, the linear latency identity fails.","supporting_citations":[],"review_version":1}