{"id":"b89210b1-64bc-450a-a4c6-f0aaa85412ab","arxiv_id":"2506.12073","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Neural LCS uses learned phoneme and word similarity instead of exact matches to align dysfluent speech to intended text, and it outperforms DTW and Hard LCS on simulated benchmarks.","lead":"This paper presents a machine-learning method that lines up the words a person with a speech disorder actually says with the sentence they were trying to say, even when sounds are added, dropped, or mispronounced. It reports better alignment and segmentation than current tools, though the strongest evidence comes from artificially generated speech and only a small amount of real patient data.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"PPA evaluation has no described ground-truth labels, so the real-data claim is unverified; Table 4 also contradicts 'significantly outperforms' on deletion.","rationale":"The paper proposes a sensible supervised neural aligner and demonstrates clear gains over Hard LCS and DTW on its own simulated benchmarks, with code released. That part of the contribution is reasonably supported. The central clinical claim, however, rests on the PPA experiment, which lacks any described ground truth, so the most load-bearing concern is that the real-data result is unverifiable and may be circular. The reader's weakest_assumption is broader, focusing on simulation realism; I partially agree because the simulation-transfer worry is real, but the concrete, checkable problem is the missing label provenance for PPA. In addition, Table 4 shows the STA model underperforms on deletion in the LLM-disorder evaluation, directly contradicting the abstract's overbroad segmentation superiority claim. The missing comparison to SSDM/CSA and [4] further weakens the 'state-of-the-art' wording. These issues justify keeping the conditional verdict rather than moving to accept: the method may well work, but the evidence for real clinical data is not yet established. No change beyond the reader's already-conditional recommendation is needed, so the verdict remains unchanged.","tokens_in":8556,"tokens_out":4263,"duration_ms":55516,"concrete_test":"Ask the authors to release or fully describe the PPA gold-standard boundary annotations (source, protocol, and inter-annotator agreement), then recompute Table 4's PPA Boundary Loss with per-type metrics and confidence intervals; if the STA model no longer beats YOLO-Stutter within inter-annotator agreement, or if the labels originate from the same simulation/alignment pipeline, the real-data claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3.2 reports Boundary Loss on PPA Speech (21 ms vs. 17 ms) but never states what the 'actual boundaries' are. No manual annotation protocol, number of dysfluent regions, or inter-annotator agreement is described; the only information is that 38 participants read the grandfather passage. Without a gold standard, the BL numbers are uninterpretable, and if the boundaries came from the same injection/alignment pipeline used to train the model, the comparison with YOLO-Stutter is circular. Furthermore, Table 4 on the LLM-disorder set shows the STA model is worse than YOLO-Stutter on deletion (27 ms vs. 10 ms), so the abstract's blanket 'significantly outperforms ... dysfluent speech segmentation' does not hold for one of the four dysfluency types. The paper also never compares against SSDM/CSA [5,6] or the weakly-supervised aligner [4], which are the closest prior dysfluent aligners; 'state-of-the-art' is asserted without a direct comparison. Together, the real-data claim in the abstract is not supported by a verifiable evaluation.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Neural LCS, a siamese-network-based soft alignment method for dysfluent text-text and speech-text alignment, and an associated STA model that combines CTC-based phoneme decoding with Neural LCS to segment dysfluent speech. The authors construct two simulated resources: a text-text corpus built by injecting dysfluencies into VCTK text according to hand-defined CMU phoneme similarity categories, and an LLM disorder corpus generated by Claude-generated dysfluent text synthesized with VITS TTS, where VITS internal phoneme frame boundaries serve as ground truth. They evaluate alignment accuracy against DTW and Hard LCS on simulated text, and boundary loss against YOLO-Stutter on LLM disorder audio and on PPA speech from 38 participants. The central claim is that Neural LCS significantly outperforms state-of-the-art models in both alignment accuracy and dysfluent speech segmentation.","tokens_in":8811,"tokens_out":3739,"duration_ms":42493,"significance":"If validated, the proposed soft, context-aware LCS formulation would be a genuinely useful contribution to disordered speech analysis: it addresses partial alignment and phonetic similarity in a linguistically grounded way, and the released code and simulated corpora could benefit the community. The paper also makes a laudable effort to generate large-scale dysfluent training data. However, the current evidence is weakened by circular evaluation on the text-text benchmark, an underspecified PPA gold standard, an overbroad abstract claim contradicted by the deletion result, and the absence of comparisons with the closest prior dysfluent aligners. The core idea is plausible, but the experimental support is not yet commensurate with the claims.","major_comments":[{"comment":"The text-text evaluation is circular: the training labels are generated by injecting dysfluencies using the hand-defined CMU phoneme categories in Table 1, and the test set is generated by the same rule. The model is therefore being evaluated on its ability to reproduce the very simulation rule it was trained on, so the 72.55% phoneme-level accuracy on text-text data does not by itself demonstrate generalization to real dysfluent speech. The authors should validate on an independently annotated corpus of dysfluent transcripts, or at minimum report a human-annotated subset, and clearly separate 'simulation-rule reconstruction' from 'alignment accuracy on real dysfluencies'.","section":"§2.2.1, Table 1, §3.3.1, Table 3"},{"comment":"The PPA Speech evaluation does not define the ground-truth boundaries used to compute Boundary Loss. No manual annotation protocol, number of dysfluent regions, or inter-annotator agreement is reported, so the 17 ms versus 21 ms comparison is uninterpretable. If the 'actual boundaries' were obtained from the same CTC/Neural LCS pipeline used to train the STA model, the comparison with YOLO-Stutter would be circular. The authors must describe how PPA boundaries were obtained and ideally report per-dysfluency-type results on PPA as well.","section":"§3.3.2, Table 4"},{"comment":"The abstract's blanket claim that Neural LCS 'significantly outperforms state-of-the-art models in ... dysfluent speech segmentation' is contradicted by Table 4 on the LLM disorder set: the STA model has a Boundary Loss of 27 ms for deletions versus 10 ms for YOLO-Stutter, i.e., it is substantially worse on one of the four dysfluency types. The authors should qualify the claim, explain why deletion detection is worse, and report aggregate metrics that include deletion or per-type significance tests.","section":"§3.3.2, Table 4 (LLM disorder rows)"},{"comment":"The paper asserts state-of-the-art performance but never compares against the closest prior dysfluent aligners: SSDM/CSA [5, 6] and the weakly-supervised forced aligner of [4]. The Introduction explicitly discusses SSDM/CSA and motivates Neural LCS as an improvement over CSA, yet no experimental comparison is provided. Given that SSDM/CSA is the direct predecessor and one of the authors is a co-author of that work, this omission is particularly consequential for the 'state-of-the-art' claim. The authors should add these baselines to Tables 3 and 4 or justify their exclusion with quantitative evidence.","section":"§2.1, §3.3.1"},{"comment":"No variance or significance testing is reported. The abstract uses 'significantly outperforms,' but the tables show single point estimates without error bars, confidence intervals, or statistical tests (e.g., paired bootstrap or multiple-seed runs). The authors should report standard deviations across seeds or bootstrap intervals, and perform a significance test for the headline comparisons in Tables 3 and 4.","section":"§3.3, Tables 3 and 4"}],"minor_comments":[{"comment":"In the Vowel row, 'IH IY' should read 'IH, IY' to be consistent with the other comma-separated lists.","section":"§2.2.1, Table 1"},{"comment":"The text reads 'wax2vec2.0 feature extractor'; this should be 'wav2vec2.0'.","section":"§2.4"},{"comment":"The section title 'Speech-text alignment' appears twice (in §3.3.1 and §3.3.2), but §3.3.1 actually reports text-text alignment results; the first title should be renamed, e.g., 'Text-text alignment'.","section":"§3.3.1"},{"comment":"The PPA Speech description states 'approximately one hour of speech in total' for 38 participants reading the grandfather passage; please clarify the duration per participant and whether all recordings were analyzable.","section":"§3.1"},{"comment":"The notation '110.7×10^4 sentences' is unconventional; use '1.107M' or '1,107,000' for readability.","section":"Table 2"},{"comment":"Reference [7] cites a Wikipedia page for the longest common subsequence; it would be more appropriate to cite a peer-reviewed source such as Hirschberg (1977), which is already [26].","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The manuscript's author list overlaps substantially with the authors of SSDM [5] and YOLO-Stutter [8], yet neither SSDM/CSA is benchmarked in the experiments. This is a scope concern that the editor may wish to raise explicitly, as the 'state-of-the-art' claim is unsupported without that comparison. The PPA data are co-authored with the UCSF clinical group, so an annotation protocol and per-type metrics should be straightforward to provide; their absence is surprising and needs to be addressed. The text-text circularity is the most serious correctness risk and should be resolved with independent annotation before acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my read. The paper's real contribution is a supervised, context-aware aligner for dysfluent text and speech: a siamese T5 encoder that predicts per-token alignment labels (0/1/2), trained with focal loss, plus two new simulated corpora (text-text from VCTK with rule-based injections, and LLM+TTS audio). That's a plausible engineering step forward, and the open code is a plus. On the paper's own text-text and LLM-text benchmarks, it beats Hard LCS and DTW by large margins, which is credible.\n\nThe soft spots are in the evaluation. First, the text-text test set is generated from the same hand-defined CMU phoneme categories used to build the training labels, so the test distribution partly encodes the training rule. That doesn't make the result meaningless, but it limits what the phoneme-level accuracy numbers mean. Second, the LLM disorder audio gets its 'true' alignments from VITS's internal alignment, which is the same TTS pipeline that generated the audio; that's synthetic ground truth. Third, the PPA result in Table 4 is uninterpretable as reported: no ground-truth boundary annotation protocol, no number of dysfluent regions, no inter-annotator agreement. Without that, a 4 ms boundary-loss gap over YOLO-Stutter proves little. Fourth, the paper never compares against the closest prior aligners, SSDM/CSA and the weakly-supervised aligner in [4], so 'state-of-the-art' is asserted rather than shown. Finally, the abstract says 'significantly outperforms' with no significance test, and Table 4 itself shows the STA model is worse on deletion (27 ms vs 10 ms), which the paper acknowledges in prose but the abstract overstates.\n\nNone of these are fatal to the idea. The architecture is sensible and the simulated corpora are useful assets if released with injection probability metadata. But the central real-data claim needs a proper gold-standard evaluation, and the missing baseline comparison needs to be filled.\n\nMy recommendation: this deserves a serious referee. It's a legitimate engineering contribution with a real clinical bottleneck in view. Send it out, but the referee should require a comparison against SSDM/CSA, error bars, and a manually annotated real PPA evaluation before accepting. I would not cite it in its current form.","headline":"Sensible supervised dysfluent aligner with useful new simulated corpora, but evaluation is too circular and the PPA ground truth is unverified; worth reviewing, not accepting as-is.","tokens_in":9392,"tokens_out":2531,"would_cite":false,"duration_ms":28704,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A new neural aligner, Neural LCS, learns phoneme similarity to map dysfluent speech onto intended text, outperforming exact-match baselines in alignment accuracy and boundary loss.","keywords":["dysfluent speech alignment","forced alignment","longest common subsequence","phoneme similarity","speech segmentation","primary progressive aphasia","dysfluency detection","neural sequence alignment"],"falsifier":"Run the speech-text alignment model against the exact-match baseline on a corpus of patient audio with clinician-verified dysfluent boundaries and phoneme-level transcriptions; if the neural aligner does not maintain substantially lower boundary loss and higher alignment accuracy, the central claim fails.","tokens_in":8341,"feed_emoji":"🗣️","tokens_out":9082,"duration_ms":76256,"temperature":0.7,"pith_summary":"To diagnose neurodegenerative speech disorders, software must map what a patient actually said onto the text they meant to say; the mismatches are the dysfluencies. This paper argues that existing aligners fail because they require exact token matches, so they cannot pair acoustically similar phonemes such as P and B. Neural LCS replaces exact matching with a learned soft alignment that marks each unit as aligned (1), dysfluent within an aligned region (0), or missing (2). On simulated text-text and LLM-generated dysfluent speech, the method lifts phoneme-level alignment accuracy to 72.55% (text-text) and 90.96% (LLM text). It also segments audio of real primary progressive aphasia patients with a lower boundary loss (17 ms) than a strong stutter-detection baseline (21 ms), which matters for automated diagnosis.","feed_headline":"Neural LCS lifts dysfluent speech alignment to 91% accuracy","feed_subtitle":"For diagnosing speech disorders, the aligner also cuts boundary error on real patient audio from 21 ms to 17 ms.","key_machinery":"The central object is Neural LCS, a neural generalization of the longest common subsequence aligner. Instead of requiring exact token equality, it uses a siamese (shared-weights) network: a transformer text encoder processes the reference and the dysfluent sequences, the resulting representations are concatenated, and a 1D convolutional network plus multilayer perceptron with softmax emits per-position alignment labels (0, 1, or 2). Contextual attention lets a K align with a G or a vowel with a similar vowel, which is what partial and substitution-heavy dysfluencies require. For speech input, a self-supervised speech feature extractor with connectionist temporal classification decoding first produces the dysfluent phoneme sequence, and Neural LCS then aligns that sequence to the reference; the simulation pipeline supplies training labels by injecting phoneme-category-based dysfluencies into English text and by generating natural dysfluent text with an LLM and audio with an end-to-end TTS model.","core_discovery":"The paper's central claim is that dysfluent speech alignment should be treated as a learned, context-sensitive version of the longest common subsequence problem rather than as exact string matching. The Neural LCS model encodes reference and dysfluent sequences with a shared transformer encoder, concatenates the two representations, and applies a convolutional classifier to predict a label sequence in which 1 means aligned, 0 marks a dysfluent unit inside an aligned segment, and 2 marks a missing phoneme or word. It is trained with a class-imbalance-weighted loss to handle the fact that most units are aligned. The authors report that this soft aligner outperforms dynamic time warping and hard LCS on both phoneme and word levels, and that the speech-to-text variant built on connectionist temporal classification plus Neural LCS lowers the boundary loss of dysfluent segments on both simulated disordered speech and real primary progressive aphasia speech.","pith_inferences":["Because the real-patient evaluation reports only an aggregate boundary loss with no manual annotation or per-type breakdown, a clinician-annotated comparison is the open test of whether the simulated-data gains transfer; the numbers are suggestive but not conclusive.","If learned phoneme similarity is the driver of the gains, replacing the hand-built phoneme categories with embeddings fit on real disordered speech should further improve substitution alignment, an experiment the paper leaves to future work.","The alignment objective is not tied to English: with a suitable tokenizer and speech encoder it could be applied to non-English dysfluent speech, but the current simulated data and phoneme categories are English-only."],"forward_implications":["Phoneme-level alignment of dysfluent text rises from 24.78% to 72.55% on simulated text-text data and from 43.53% to 90.96% on LLM-generated text, so downstream dysfluency counts should become far more reliable.","On real primary progressive aphasia audio, the speech-text model segments dysfluent regions with 17 ms boundary loss versus 21 ms for the stutter-detection baseline, indicating clinically usable segmentation.","The 0/1/2 labeling makes each alignment output a dysfluency annotation, so insertion and deletion detection do not require a separate post-processing stage.","Repetition accuracy stays near 96% across different training mixes of dysfluency types, while substitution accuracy responds to the proportion of substitutions in the training data."],"supporting_citations":[{"why":"Defines the longest common subsequence problem that Neural LCS generalizes, and serves as the exact-match baseline.","marker":"[26]"},{"why":"Supplies dynamic time warping as the global-alignment baseline in the accuracy comparison.","marker":"[27]"},{"why":"Serves as the open-source dysfluency boundary detection baseline for the speech-text segmentation comparison.","marker":"[8]"},{"why":"Defines the primary progressive aphasia population whose real recordings are used for evaluation.","marker":"[22]"},{"why":"Supplies the language model used to generate natural dysfluent text for the text-speech simulated corpus.","marker":"[29]"},{"why":"Supplies the end-to-end TTS model that produces dysfluent audio and internal phoneme time-frame alignments.","marker":"[30]"},{"why":"Provides the transformer feature encoder used inside the siamese alignment network.","marker":"[33]"},{"why":"Provides the class-imbalance-weighted loss used to train the alignment label predictor.","marker":"[34]"},{"why":"Supplies connectionist temporal classification decoding for mapping audio frames to a dysfluent phoneme sequence.","marker":"[16]"},{"why":"Supplies the self-supervised speech feature extractor used by the speech-to-text model.","marker":"[35]"}],"fun_headline_variants":["Neural LCS: learned context-aware alignment for dysfluent speech","Neural LCS outdoes DTW and hard LCS on dysfluent speech","Learned LCS improves dysfluent speech alignment and segmentation","Context-sensitive LCS aligns dysfluent speech with higher accuracy","Neural LCS: better alignments for diagnosing speech disorders"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simulated labels used for training and evaluation behave like real dysfluent speech; if the simulation rules do not match how people with primary progressive aphasia actually speak, the reported gains may not appear on clinical audio.","fun_headline_variants_meta":{"raw":{"variants":["Neural LCS: learned context-aware alignment for dysfluent speech","Neural LCS outdoes DTW and hard LCS on dysfluent speech","Learned LCS improves dysfluent speech alignment and segmentation","Context-sensitive LCS aligns dysfluent speech with higher accuracy","Neural LCS: better alignments for diagnosing speech disorders"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000699,"raw_usage":{"total_tokens":3120,"prompt_tokens":872,"completion_tokens":2248,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":488,"completion_tokens_details":{"reasoning_tokens":2155}},"tokens_in":488,"tokens_out":2248,"duration_ms":18859,"temperature":1.0,"reasoning_tokens":2155,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:38:46.671695+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the speech-text alignment model against the exact-match baseline on a corpus of patient audio with clinician-verified dysfluent boundaries and phoneme-level transcriptions; if the neural aligner does not maintain substantially lower boundary loss and higher alignment accuracy, the central claim fails.","supporting_citations":[{"cited_title":"Phone-to-audio alignment without text: A semi-supervised approach,","cited_arxiv_id":null,"evidence_quote":"Defines the longest common subsequence problem that Neural LCS generalizes, and serves as the exact-match baseline."},{"cited_title":"Towards hierarchical spoken lan- guage disfluency modeling,","cited_arxiv_id":null,"evidence_quote":"Serves as the open-source dysfluency boundary detection baseline for the speech-text segmentation comparison."},{"cited_title":"Wavlm-ctc-hugginface,","cited_arxiv_id":null,"evidence_quote":"Supplies the language model used to generate natural dysfluent text for the text-speech simulated corpus."},{"cited_title":"Conditional variational autoencoder with adversarial learning for end-to-end text-to-speech,","cited_arxiv_id":null,"evidence_quote":"Supplies the end-to-end TTS model that produces dysfluent audio and internal phoneme time-frame alignments."},{"cited_title":"Dynamic-programming approach to continuous speech recognition,","cited_arxiv_id":null,"evidence_quote":"Provides the transformer feature encoder used inside the siamese alignment network."},{"cited_title":"Cmu phoneme dictionary","cited_arxiv_id":null,"evidence_quote":"Provides the class-imbalance-weighted loss used to train the alignment label predictor."},{"cited_title":"During inference, we apply greedy decoding on CTC emis- sion matrix","cited_arxiv_id":null,"evidence_quote":"Supplies the self-supervised speech feature extractor used by the speech-to-text model."}],"review_version":1}