{"id":"c0ae8eb4-af66-4524-85bb-148eeaee196a","arxiv_id":"2504.14915","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"StableQuant adaptively clips outliers per layer before choosing quantization scales, keeping WER within 0.3% at 8-bit precision on HuBERT and wav2vec2.0.","lead":"The paper introduces StableQuant, a method for compressing speech AI models that chooses each layer's number range separately. It reports a 4x smaller model and 2x faster inference while keeping speech recognition errors nearly unchanged.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central W8A8 result depends on Stage 1's isolated per-layer sensitivity selection, which is unvalidated; a joint-selection ablation is needed to confirm the claim.","rationale":"The paper's core contribution is empirical; the adaptive layer selection is the main new mechanism, so its validity is load-bearing. The isolated-sensitivity assumption is not demonstrated. However, the reported numbers are internally consistent, and the method is plausible; a joint-selection ablation would settle the concern. This does not change the reader's CONDITIONAL verdict but reinforces it, because the requested conditions already include additional validation and stronger baselines.","tokens_in":7854,"tokens_out":11070,"duration_ms":104270,"concrete_test":"For each model, run a greedy joint selection under W8A8: start with all layers max-quantized (the Table I All-Layer setting), then iteratively add to S the layer whose percentile clipping yields the largest WER improvement on dev-clean, re-evaluating jointly after each addition, until no layer improves WER by more than gamma. Compare the final test-clean WER and the selected S against StableQuant's Stage 1 output. If joint selection gives a different S or test-clean WER more than 0.3% below baseline, then isolated Stage 1 selection is load-bearing; if S and WER match, the heuristic is adequate.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (Stage 1) selects clipping layers S by applying max-scale W8A8 quantization to one layer at a time and comparing WER on the dev set to the full-precision baseline (lines 2-9). This treats per-layer quantization error as separable: a layer whose isolated WER increase exceeds gamma=0.25% is added to S. Under full W8A8 quantization, errors from different layers interact through residual connections and the Transformer stack, so the set of layers that are problematic jointly need not equal the set that are problematic in isolation. The threshold gamma is arbitrary and tuned on dev-clean, and there is no comparison against a joint selection procedure. Since S determines which layers receive percentile clipping in Stage 2, a suboptimal S could change the reported W8A8 WER (HuBERT 2.35 vs 2.16, wav2vec2.0 3.07 vs 2.78) and invalidate the \"<0.3% drop\" claim.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes StableQuant, a post-training quantization method for speech foundation models. The method first identifies 'sensitive' layers by quantizing each layer's activations individually at 8 bits and measuring the WER increase relative to an FP32 baseline (Stage 1, Algorithm 1); it then searches a percentile clipping threshold for the selected layers and an MSE-based scale for all layers (Stage 2, Algorithm 2). Experiments on HuBERT-Large and wav2vec2.0 fine-tuned for ASR on LibriSpeech report W8A8 WER of 2.35% and 3.07% versus baselines of 2.16% and 2.78%, model size reduction from 1205 MB to 353 MB for HuBERT-Large, and roughly a 2x inference speedup under TensorRT. The paper concludes that StableQuant achieves less than 0.3% WER degradation at 8-bit quantization for both models.","tokens_in":7971,"tokens_out":7182,"duration_ms":63194,"significance":"If the results are reproducible, the paper addresses an underexplored and practically important problem: post-training quantization for speech foundation models with CNN feature extractors. The empirical demonstration that all-layer quantization is much harder than Transformer-only quantization (Table I) is a useful observation, and the proposed combination of layer selection and percentile clipping is simple and plausible. The authors provide a public code repository link, which aids reproducibility. The main limitations are the lack of validation of the layer-selection assumption, the absence of error bars or sensitivity analyses, and insufficient detail on comparison baselines; these prevent the 'optimal' and '<0.3% drop' claims from being fully established.","major_comments":[{"comment":"The selection of the clipping-layer set S is made by quantizing one layer at a time while leaving all other layers in full precision (line 4 sets only Nq(al) to the quantized version). This treats per-layer quantization error as separable and assumes that the layers most sensitive in isolation are also the layers that matter after all weights and activations are quantized to W8A8. In a Transformer-CNN stack with residual connections, quantization errors can interact, so this assumption is not self-evident. The paper does not report an ablation against any joint selection procedure (e.g., greedy forward selection at W8A8, clipping all CNN layers, or leave-one-out at W8A8). Because S directly determines which layers receive percentile clipping in Stage 2, this is load-bearing for the central W8A8 claim; please add such an ablation.","section":"III-B, Algorithm 1 (Stage 1)"},{"comment":"The threshold gamma=0.25% and the percentile grid P in [0,0.5] with step 0.01 are selected on the dev-clean set, and all reported WER numbers come from a single run. The headline result is close to the threshold: for wav2vec2.0 the W8A8 WER increase is 0.29%, just under the 0.3% bound, and for HuBERT it is 0.19%, comparable to gamma. A sensitivity analysis over gamma, the calibration set, and the histogram bin count B, or repeated runs with different dev/calibration splits, is needed to establish that the '<0.3%' claim is robust rather than a consequence of hyperparameter selection.","section":"III-B, IV-B"},{"comment":"The columns 'Percentile', 'MSE', and 'Entropy' are presented as conventional calibration baselines, but the caption says the table applies 'the proposed layer-adaptive calibration strategies.' It is not clear whether these baselines use the same selected layer set S as StableQuant, whether they also perform a cut-off ratio search, or whether they represent standard single-scale calibration applied uniformly to all layers. As written, the comparison conflates the calibration objective with the layer-selection mechanism, which is exactly the proposed contribution. Please specify the exact protocol for each baseline.","section":"IV-B, Table II"},{"comment":"The AdaRound comparison reports a single WER of 2.66% with no indication of the bit width, whether activation quantization is included, what calibration data were used, or whether the same TensorRT pipeline was applied. The claim that StableQuant 'surpasses the quantized results with AdaRound' therefore cannot be fully evaluated. Please detail the AdaRound setup and, ideally, report both methods under identical W8A8 conditions.","section":"IV-B, Table III"}],"minor_comments":[{"comment":"The histogram H_{D,clip}(X) is not defined as normalized, and the MSE expression in line 10 weights squared errors by histogram counts; please clarify the normalization and state how the bin count B was chosen.","section":"III-B, Algorithm 2"},{"comment":"The abstract and conclusion claim 'optimal quantization performance regardless of the network architecture type,' but the experiments include only two SFMs sharing a CNN+Transformer structure; please temper 'regardless' or add evidence for more architectures.","section":"Abstract and Conclusion"},{"comment":"The numeric inference times are not given in the caption or the text; reporting exact milliseconds values would make the 'roughly twofold' speedup claim easier to verify.","section":"IV-C, Fig. 2"},{"comment":"Section II-A begins without a clear lead-in sentence immediately after the heading, and Table I uses the abbreviation 'TR-Layer' without definition; minor editing would improve readability.","section":"II-A and Table I"}],"recommendation":"major_revision","confidential_remarks":"This is a solid empirical paper with a clear central claim, but the joint-selection ablation and baseline-detail clarifications are necessary before publication. I recommend major revision; no ethical concerns, though the 'first work to apply PTQ to SFMs' claim should be moderated or supported with a focused survey."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"StableQuant is a genuinely useful empirical recipe, not a deep theoretical contribution. The core result—W8A8 quantization of HuBERT-Large and wav2vec2.0 keeps WER degradation under 0.3% on test-clean while cutting model size to a quarter and roughly doubling inference speed—is consistent with the tables and strikes me as credible. The method is a sensible combination: use a dev set to find which CNN layers are sensitive to naive max-scale quantization, then for those layers search a percentile clipping cutoff and MSE-optimal scale, and quantize the rest with MSE-calibrated scales only. That the selected clipping layers are always the first few conv layers matches the distributional story in Figure 1. I credit the authors for reporting test-clean results, since dev-clean is used for model selection and test-clean is held out.\n\nThe weakest part is Stage 1's isolated per-layer selection. Algorithm 1 quantizes one layer at a time and flags layers whose isolated WER increase exceeds gamma. It is not obvious that the set selected this way is the correct set under full quantization; errors can interact through residual connections in the Transformer stack, though for front-end conv layers the interaction may be weaker. The paper does not compare against a joint selection procedure or a brute-force choice. This is a real gap, but not a fatal one: the selected set is small and consistent, and the central W8A8 claim would need only a modest correction if a different set were better. Minor concerns: no error bars or multiple runs; only AdaRound as an external PTQ baseline, and the baselines in Table II may not have been searched with the same effort; and \"first work to apply PTQ to SFMs\" is a strong claim with no serious literature search to back it. The gamma threshold is also arbitrary and tuned on the same dev set.\n\nThis paper is for people doing quantization for speech models, not for theorists. It deserves a serious referee; the empirical recipe is easy to evaluate and the main result is useful. I would accept it for peer review and ask for a joint-selection ablation, error bars, broader baselines, and a softened novelty claim.","headline":"Useful empirical recipe for PTQ of speech foundation models with a believable W8A8 result; the isolated per-layer sensitivity selection needs a joint-selection ablation before I'd trust it beyond W8A8.","tokens_in":8615,"tokens_out":2264,"would_cite":true,"duration_ms":20534,"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 layer-by-layer calibration method lets 8-bit speech foundation models keep ASR accuracy within 0.3% word error rate while cutting model size to about a quarter.","keywords":["post-training quantization","speech foundation models","layer-adaptive calibration","outlier clipping","automatic speech recognition","model compression","HuBERT","wav2vec2.0"],"falsifier":"On the same LibriSpeech dev-clean set, apply the full percentile-clipping and MSE calibration of Stage 2 to every layer, bypassing the Stage-1 selection; if the resulting W8A8 or W6A6 WER is no worse than StableQuant's selected-layer result, then the isolated-sensitivity selection is not responsible for the method's gains.","tokens_in":7535,"feed_emoji":"🎙️","tokens_out":8847,"duration_ms":72734,"temperature":0.7,"pith_summary":"The paper claims that speech foundation models resist standard post-training quantization because their convolutional feature extractors produce activation distributions that are wider and more outlier-prone than the Transformer layers, and that a layer-adaptive calibration fixes this. The proposed method, StableQuant, first identifies the layers where naive quantization hurts WER, then for those layers clips the activation histogram at a searched cut-off percentile before choosing the quantization scale by MSE. On HuBERT and wav2vec2.0 fine-tuned for ASR, the paper reports WER within 0.3% of full precision at 8-bit weight/activation quantization, a model-size drop from 1205 MB to 353 MB for HuBERT-Large, and roughly doubled inference speed. This matters because it suggests speech encoders can be compressed for deployment without fine-tuning or student networks.","feed_headline":"Speech AI models shrink 4x with 8-bit quantization, WER barely moves","feed_subtitle":"Layer-adaptive outlier clipping keeps HuBERT and wav2vec2.0 near full precision while doubling inference speed.","key_machinery":"The load-bearing machinery is the two-stage calibration in Algorithms 1 and 2. Stage 1 selects a clipping set S by temporarily quantizing each layer alone with percentile scaling and keeping any layer whose WER increase exceeds gamma = 0.25%. Stage 2, for layers in S, clips the activation histogram at a candidate cut-off percentile p between 0 and 0.5 (grid step 0.01) and computes the scale factor as the bin-center value that minimizes the mean squared error between original clipped bins and their quantized-dequantized reconstructions; layers outside S use the same MSE search without clipping. This separates outlier removal from scale selection, so the quantization range is driven by reconstruction error over the dominant distribution rather than by the full dynamic range.","core_discovery":"The central claim is that the failure of PTQ on speech foundation models comes from the CNN front-end, not from the Transformer stack: the paper measures CNN activation ranges up to about 100 times wider than Transformer activations, with strong layer-to-layer differences in both HuBERT and wav2vec2.0. StableQuant's recipe is to treat layers differently: quantize each layer alone with a naive percentile scale, flag every layer whose WER increase exceeds a 0.25% threshold, and then restrict outlier clipping to those flagged layers, searching the cut-off percentile p in [0, 0.5] at 0.01 steps and selecting the scale that minimizes MSE on the clipped histogram. On the ASR task, this yields 2.35% WER for HuBERT-Large at W8A8 (baseline 2.16%) and 3.07% for wav2vec2.0 (baseline 2.78%), with graceful degradation at W6A6 where conventional percentile, MSE, and entropy calibrations collapse to WER above 42%.","pith_inferences":["An implied next test is to run StableQuant on a third speech foundation model with a different front-end, such as WavLM; if the under-0.3% WER drop does not survive, the 'regardless of architecture' claim would need to be narrowed.","The paper's Stage-1 sensitivity test quantizes one layer at a time; a natural extension is a joint or greedy subset selection that quantizes several layers together, since quantized errors from different layers may interact and change which layers most need clipping.","The cut-off percentile and the gamma threshold are tuned on LibriSpeech dev-clean; a robustness check would be whether the same settings transfer to noisy or out-of-domain speech without recalibration.","Because StableQuant only changes the scale-calibration step, it is complementary to weight-rounding optimizers such as AdaRound; the paper compares against AdaRound but does not test the combination, which may push W6A6 accuracy higher."],"forward_implications":["At W8A8, HuBERT-Large with StableQuant needs 353 MB versus 1205 MB full precision and still beats full-precision HuBERT-Base, which stands at 6.33% WER, with 2.35% WER.","At W6A6, the same method keeps HuBERT at 3.98% and wav2vec2.0 at 6.10% WER, where percentile, MSE, and entropy calibration each exceed 42%, so 6-bit deployment becomes usable.","Because the calibration is post-training and uses only the dev set, no student network, distillation, or fine-tuning is needed to compress the models.","Inference time roughly doubles when both weights and activations are quantized to 8 bits and run through TensorRT on an RTX 3090, across audio lengths from 1 to 600 seconds.","The same two-stage recipe works for both HuBERT and wav2vec2.0, which use different self-supervised objectives and slightly different front-ends, supporting the claim that the method is not architecture-specific."],"supporting_citations":[{"why":"Supplies the wav2vec2.0 architecture with its CNN feature extractor that StableQuant quantizes.","marker":"[19]"},{"why":"Supplies the HuBERT architecture, the second evaluation model.","marker":"[20]"},{"why":"Represents the LLM-oriented SmoothQuant PTQ approach that the paper argues does not transfer directly to speech foundation models.","marker":"[17]"},{"why":"Provides the AdaRound layer-wise weight-rounding PTQ baseline that StableQuant is compared against.","marker":"[26]"},{"why":"Supplies the LibriLight corpus used to pre-train the two speech foundation models.","marker":"[23]"},{"why":"Supplies the LibriSpeech corpus used for ASR fine-tuning, calibration, and evaluation on dev-clean and test-clean.","marker":"[24]"},{"why":"Provides the TensorRT 8-bit inference engine used to measure the quantized models' inference speed.","marker":"[25]"}],"fun_headline_variants":["Per-layer quantization: 4x smaller speech AI, 2x faster, <0.3% WER drop","Adaptive per-layer PTQ: 4x smaller, 2x faster, <0.3% WER drop","Speech models 4x smaller with layer-adaptive quantization, WER drop <0.3%","Per-layer outlier clipping: 4x smaller speech, <0.3% WER loss"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method rests on the assumption that a layer's sensitivity measured when only that layer is quantized predicts which layers are harmful when all layers are quantized together, even though quantized errors from different layers interact.","fun_headline_variants_meta":{"raw":{"variants":["Per-layer quantization: 4x smaller speech AI, 2x faster, <0.3% WER drop","Adaptive per-layer PTQ: 4x smaller, 2x faster, <0.3% WER drop","Speech models 4x smaller with layer-adaptive quantization, WER drop <0.3%","Per-layer outlier clipping: 4x smaller speech, <0.3% WER loss"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001641,"raw_usage":{"total_tokens":6521,"prompt_tokens":946,"completion_tokens":5575,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":5464}},"tokens_in":562,"tokens_out":5575,"duration_ms":34254,"temperature":1.0,"reasoning_tokens":5464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:36:50.161902+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On the same LibriSpeech dev-clean set, apply the full percentile-clipping and MSE calibration of Stage 2 to every layer, bypassing the Stage-1 selection; if the resulting W8A8 or W6A6 WER is no worse than StableQuant's selected-layer result, then the isolated-sensitivity selection is not responsible for the method's gains.","supporting_citations":[{"cited_title":"SmoothQuant: Accurate and efficient post-training quantization for large language models,","cited_arxiv_id":null,"evidence_quote":"Represents the LLM-oriented SmoothQuant PTQ approach that the paper argues does not transfer directly to speech foundation models."},{"cited_title":"Libri-light: A benchmark for asr with limited or no supervision,","cited_arxiv_id":null,"evidence_quote":"Supplies the LibriLight corpus used to pre-train the two speech foundation models."},{"cited_title":"8-bit inference with tensorrt,","cited_arxiv_id":null,"evidence_quote":"Provides the TensorRT 8-bit inference engine used to measure the quantized models' inference speed."}],"review_version":1}