{"id":"e9da66d4-7db2-430e-93d5-ed6a28bb71d4","arxiv_id":"2506.03554","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"MS-Wavehax, a multi-stream extension of the Wavehax vocoder, achieves the best throughput in low-latency CPU streaming and near non-causal quality with one frame of lookahead.","lead":"This paper proposes MS-Wavehax, a compact neural vocoder for low-latency streaming speech synthesis, and compares its speed and quality against several baselines on a CPU. It reports that MS-Wavehax achieves the highest throughput below 80 ms latency while matching the quality of larger models, using only 2.4% of HiFi-GAN's parameters.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ONNX STFT/iSTFT replacement confounds the throughput comparison; MS-Wavehax's low-latency advantage may be an implementation artifact.","rationale":"The reader's weakest assumption concerned generalizability across different CPUs and inference backends; that is a fair external-validity concern. The stress test identifies an internal confound that is more immediate and more load-bearing: the ONNX conversion itself, specifically the replacement of STFT/iSTFT with convolution-based implementations, dramatically changes MAC counts in Table 1 and therefore likely changes the RTF ranking in Figure 2. The paper uses this ranking to support the central claim of highest throughput under low-latency conditions. If the ranking is an artifact of the conv-based STFT replacement, then the claim about the proposed architecture does not hold as stated. The proposed concrete test would settle whether this concern lands. Since the reader already issued a CONDITIONAL verdict, and the stress test reinforces that conditionality without moving it to reject, the verdict remains UNCHANGED; however, the stated reason for conditionality should be updated to include this implementation confound along with the generalizability issue.","tokens_in":8621,"tokens_out":3697,"duration_ms":38541,"concrete_test":"Measure streaming RTF on the same AMD EPYC 7302 with ONNX Runtime using a custom STFT/iSTFT operator (or equivalently with PyTorch eager using the original torch.stft and torch.istft), keeping all model weights and chunk sizes fixed; recompute the Figure 2 ranking for chunk sizes that give less than 80 ms latency. If Vocos or Wavehax ties or beats MS-Wavehax, the throughput claim is an artifact of the conv-based STFT replacement.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central throughput claim rests on the ONNX Runtime experiments in Section 3.2. However, Table 1 shows that to support ONNX conversion, torch.stft and torch.istft were replaced with convolution-based implementations, and this replacement changes the MAC counts dramatically across models: Vocos jumps from 1.348B to 46.96B, Wavehax from 1.787B to 13.03B, while MS-Wavehax rises only from 1.576B to 2.291B. Thus the low-latency RTF ranking may be determined not by the multi-stream architecture or the 2D-versus-1D convolution properties that the im2col analysis in Section 3.2 and Table 2 investigates, but by the relative overhead of the conv-based STFT replacement. The paper does not break down RTF by layer or ablate the STFT implementation, so the claimed 'highest throughput under low-latency conditions' is not isolated from this implementation choice. Since the abstract and conclusions generalize to resource-constrained environments rather than specifically to ONNX with conv-based STFT, the central claim is not yet established.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes multi-stream Wavehax (MS-Wavehax), a compact neural vocoder that extends Wavehax with four-stream subscale synthesis using trainable analysis/synthesis filters, and also introduces a multi-stream version of Vocos. The authors evaluate streaming synthesis on a single CPU with ONNX Runtime, reporting real-time factors (RTF) as a function of chunk size, and present a theoretical im2col-based analysis of why 1D ConvNeXt-based Vocos overtakes 2D Wavehax variants at larger chunk sizes. Subjective MOS evaluations in analysis-synthesis and text-to-speech tasks are reported for the baseline and multi-stream models, including causal MS-Wavehax with and without one-frame lookahead. The central claims are that MS-Wavehax offers high speech quality under causal and non-causal conditions, achieves the highest throughput at low latency (<80 ms), and requires only 2.4% of HiFi-GAN V1's parameter count.","tokens_in":8864,"tokens_out":3574,"duration_ms":40565,"significance":"If the central claims hold, this is a practically valuable result: a vocoder of 0.332M parameters that nearly matches non-causal quality with one frame of lookahead would be directly useful for on-device streaming TTS and voice conversion. The paper also provides a useful methodological discussion of how streaming overhead, parameter loading, and cache behavior interact with 1D versus 2D convolution design, and the im2col matrix-size analysis in Section 3.2 and Table 2 is a clear, concrete contribution. The experimental setup uses public corpora (JVS, JSUT), a standard MOS protocol with 30 listeners, and detailed hyperparameters, which are strengths. However, as detailed below, the throughput comparison is confounded by the ONNX STFT/iSTFT replacement, and the subjective comparison lacks causal baselines and statistical testing, so the main claims are not yet fully established.","major_comments":[{"comment":"The central throughput claim is confounded by the ONNX STFT/iSTFT replacement. Table 1 shows that replacing torch.stft/torch.istft with convolution-based implementations changes per-second MACs from 1.348B to 46.96B for Vocos (a 35x increase), from 1.787B to 13.03B for Wavehax (a 7.3x increase), and from 1.576B to 2.291B for MS-Wavehax (a 1.45x increase). Since Figure 2 reports RTF measured with ONNX Runtime, the ranking at small chunk sizes may reflect the relative overhead of the conv-based STFT implementation rather than the multi-stream architecture or the 2D-versus-1D convolution behavior analyzed in the im2col discussion. The manuscript does not provide a per-layer RTF breakdown or an ablation that holds the STFT implementation fixed across all models. To support the claim that MS-Wavehax has the highest streaming throughput, the authors should either report RTF with a common STFT/iSTFT implementation (e.g., native PyTorch or a backend optimized for all models), or explicitly isolate and compare the STFT overhead per model.","section":"Section 3.2, Table 1, Figure 2"},{"comment":"The MOS evaluation does not include causal baselines. The non-causal set includes HiFi-GAN, iSTFTNet, Vocos, and Wavehax, but the only causal models evaluated are MS-Wavehax with lookahead 0 and 1. Thus the claim that MS-Wavehax 'delivers high speech quality under causal conditions' cannot be compared against causal variants of any other vocoder. For example, a causal HiFi-GAN or causal Vocos would be a natural baseline to determine whether the quality gap under causality is specific to MS-Wavehax or a general property of causal streaming. The authors should add at least one strong causal baseline or soften the claim accordingly.","section":"Section 4.2, Figure 4"},{"comment":"No statistical significance testing is reported for the MOS results. The figure shows 95% confidence intervals, but with 30 listeners and 10 samples per model, overlapping intervals do not establish equivalent or superior quality. The manuscript states that MS-Wavehax 'achieved the highest MOS in both tasks' and that 'the performance gap between MS-Wavehax with LA=1 and LA=0 disappeared in TTS,' but these conclusions are not supported by any pairwise significance test (e.g., matched bootstrap, Wilcoxon, or ANOVA with a post-hoc test). The authors should add an appropriate significance test or explicitly present the comparison as descriptive rather than inferential.","section":"Section 4.2, Figure 4"},{"comment":"Two load-bearing design choices are justified only by 'internal experiments' that are never shown: the selection of trainable analysis/synthesis filters over PQMF and DWT, and the replacement of layer normalization with batch normalization. No objective metric values, architectures, or ablations are provided for these choices, which makes the proposed method difficult to reproduce and means the reader cannot assess whether these choices are robust or tuned to the specific corpora used. The authors should report the numerical results of these internal experiments or omit the claims and treat the choices as fixed design decisions.","section":"Section 2.2 and Section 2.3"},{"comment":"The generalizability of the throughput analysis to 'resource-constrained environments' is supported by measurements on a single AMD EPYC 7302 CPU with a 16 MiB L3 cache, which is a server-class processor rather than a typical edge or mobile CPU. The paper acknowledges cache and SIMD dependencies in the discussion, but no data is provided for a range of hardware (e.g., ARM mobile cores, lower-power x86, or devices with smaller caches). The practical recommendations about chunk sizes and 1D vs 2D convolution design should either be explicitly scoped to this one CPU or validated on at least one additional representative device.","section":"Section 3.2"}],"minor_comments":[{"comment":"The sentence 'These spectrograms are then concatenated with the projected mel-spectrogram features, resulting in a 12-channel representation' appears twice verbatim; one copy should be removed.","section":"Section 2.2"},{"comment":"The table caption lists 'MACs' for PyTorch and ONNX, but the figure and text never clarify whether the PyTorch numbers use the native torch.stft/torch.istft or the same convolution-based replacement; this should be stated explicitly for both columns.","section":"Section 3.2 and Table 1"},{"comment":"The explanation for MS-Vocos underperforming Vocos is speculative ('We speculate that estimating inter-stream dependencies is particularly challenging...') and is not tested. If this claim is retained, it would benefit from a simple diagnostic, such as varying the latent channel count in MS-Vocos.","section":"Section 4.2"},{"comment":"There are several typographical artifacts, such as 'V ocoder' in the title and repeated 'V ocos' spacing in the references; a final proofreading pass would be helpful.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The main issue is that the throughput claim may be an artifact of the ONNX conv-based STFT replacement, and the MOS comparisons lack causal baselines and significance testing. These are fixable with additional experiments rather than being fatal to the approach. The paper would be suitable for publication if the authors isolate the STFT implementation effect, add causal baselines, and either add significance tests or carefully qualify the quality claims. The im2col analysis and the compact MS-Wavehax design are genuinely useful contributions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"You should know two things about this one. First, MS-Wavehax looks like a real step forward for on-device streaming vocoders: 0.33M parameters, multi-stream with trainable filters, and MOS that ties or beats much larger baselines in both analysis-synthesis and TTS. Second, the headline throughput result is weaker than it appears, because the ONNX experiments replace torch.stft/istft with convolution-based implementations, and that replacement inflates MACs very unevenly (Vocos goes from 1.3B to 47B, Wavehax from 1.8B to 13B, while MS-Wavehax goes from 1.6B to 2.3B). So the low-latency RTF ranking in Figure 2 may be telling you more about STFT implementation overhead than about the multi-stream architecture or the 2D-vs-1D convolution properties the paper analyzes in Section 3.2. The im2col analysis covers residual blocks only, not the STFT/iSTFT layers, so it does not currently support the conclusion that MS-Wavehax's advantage is architectural. The stress-test note is right to flag this.\n\nWhat the paper does well: the streaming adaptations (caching, batch norm for streaming) are sensible and clearly described; the chunk-size sweep is useful; the MOS evaluation is reasonably careful, with 95% CIs and both A/S and TTS tasks; and the authors are transparent about the ONNX STFT replacement in Table 1. The improvement of MS-Wavehax over Wavehax is also non-trivial and the explanation via local spectral consistency is plausible.\n\nSoft spots beyond the STFT confound: no significance tests for MOS (the CIs overlap in several places, so claims like 'superior quality' are a bit strong); only one CPU tested, though the paper acknowledges hardware dependence; and the 'internal experiments' that justify trainable filters and BN are referenced but not shown. None of these are fatal for a conference paper, but they should be addressed.\n\nWho is this for? Anyone working on streaming TTS or voice conversion on CPUs will want to read it. It deserves a serious referee, but the referee should ask for an ablation or at least a per-layer RTF breakdown that separates STFT/iSTFT cost from the rest of the model. If the authors can do that, the empirical contribution becomes much more solid. As is, I would not fully trust the 'highest throughput under low-latency' claim, but I would trust the model's quality and compactness.","headline":"MS-Wavehax is a genuinely compact, high-quality streaming vocoder, but the central throughput claim is partly confounded by an uneven ONNX STFT replacement across models.","tokens_in":658,"tokens_out":1051,"would_cite":true,"duration_ms":29291,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a multi-stream extension of the Wavehax vocoder delivers the best throughput among compared neural vocoders at sub-80 ms streaming latency on a single CPU, with near-non-causal quality at one frame of lookahead…","keywords":["neural vocoder","streaming synthesis","low-latency audio","multi-stream decomposition","Wavehax","CPU-only inference","subscale synthesis","real-time speech synthesis"],"falsifier":"Run the same streaming RTF comparison on a low-end mobile CPU (e.g., an ARM Cortex-A series core with small L1/L2 caches and no large shared L3) using the same ONNX Runtime pipeline and chunk sizes of 1, 4, 8, and 16 frames. If Vocos already matches or beats MS-Wavehax at chunk sizes below 8 frames, or if MS-Wavehax fails to keep the lowest RTF below 80 ms latency, then the paper's central 'highest throughput under low-latency' claim does not generalize beyond the tested EPYC configuration. A second check: switch the convolution backend from im2col-based GeMM to direct convolution on the same EPYC CPU and see whether the RTF crossover between MS-Wavehax and Vocos shifts or disappears, which would confirm or falsify the proposed memory-bandwidth mechanism.","tokens_in":8466,"feed_emoji":"🎙️","tokens_out":6933,"duration_ms":70828,"temperature":0.7,"pith_summary":"This paper argues that a high-fidelity neural vocoder can be made small and fast enough for real-time streaming on a single CPU by decomposing the waveform into sub-bands. It introduces multi-stream Wavehax (MS-Wavehax), which extends the aliasing-free Wavehax vocoder with multi-stream synthesis that splits the signal into four subscale waveforms. The paper claims that MS-Wavehax achieves the highest streaming throughput among the compared vocoders at low latency (chunk sizes under 80 ms) and nearly matches its own non-causal quality with just one frame of lookahead, while using about 2.4% of HiFi-GAN V1's model size. It also identifies why the choice between 1D and 2D convolutions and the streaming chunk size dominate efficiency, through an im2col-based memory-bandwidth analysis. If correct, this gives a practical recipe for deploying neural vocoders on resource-constrained devices without sacrificing audio quality.","feed_headline":"A 0.33M-parameter vocoder wins low-latency CPU streaming","feed_subtitle":"MS-Wavehax keeps top speech quality at just 2.4% of HiFi-GAN V1's size, with one 10 ms frame of lookahead.","key_machinery":"The load-bearing mechanism is multi-stream synthesis with trainable filters combined with streaming adaptations. A signal analysis filter (length 63) decomposes the input waveform into four subscale signals; each subscale goes through STFT with 20 ms frames and a 10 ms shift, producing subscale spectrograms that are concatenated with projected mel-spectrogram features into a 12-channel representation. A series of 2D convolutional layers and residual blocks processes this representation, and a final pointwise convolution outputs eight spectrograms (real and imaginary parts of four subscale signals), which are converted back to the time domain via iSTFT and merged by a synthesis filter. Streaming feasibility comes from ring-buffer caching applied to all convolutional, STFT, and iSTFT layers, and from replacing layer normalization with batch normalization, since batch norm does not require computing statistics over the whole sequence at inference. The throughput analysis works through im2col matrix sizes: 2D convolutions produce input matrices that grow with chunk size T and the frequency dimension, increasing main-memory traffic at larger chunks, while 1D convolutions have large but constant filter matrices that are reused from cache.","core_discovery":"On the paper's own terms, the central discovery is that multi-stream decomposition of the harmonic-prior-based Wavehax architecture yields a vocoder that is both subjectively high-quality and uniquely suited to low-latency CPU streaming. MS-Wavehax splits the input waveform with an analysis filter into four subscale signals, converts each to a complex spectrogram, concatenates them with mel-spectrogram features, processes them with 2D convolutions, and recombines the streams via a synthesis filter. In listening tests for both analysis-synthesis and text-to-speech, non-causal MS-Wavehax and causal MS-Wavehax with one-frame lookahead score highest among the vocoders. With 0.332 million parameters and about 1.576 GMACs per second of audio in PyTorch, it achieves the best real-time factor at small chunk sizes and requires only 2.4% of HiFi-GAN V1's model size. The paper further shows that as chunk size grows, memory transfer in 2D convolutions becomes the bottleneck, which is why the small 2D-convolution-based MS-Wavehax wins at low latency while the 1D-convolution-based Vocos wins at large chunk sizes.","pith_inferences":[],"forward_implications":["For chunk sizes corresponding to less than 80 ms of additional latency, MS-Wavehax gives the lowest real-time factor among the tested vocoders in a single-thread CPU environment, making it suitable for interactive real-time applications.","With one frame (10 ms) of lookahead, causal MS-Wavehax matches its non-causal quality in listening tests, indicating that near-zero-future-context streaming is feasible without audible degradation.","The model size of MS-Wavehax is about 2.4% of HiFi-GAN V1's parameter count (0.332M versus 13.82M), so it can fit into on-device memory budgets where larger vocoders are impractical.","The im2col memory-bandwidth analysis predicts that 2D-convolution vocoders lose throughput relative to 1D-convolution vocoders as chunk size grows, providing a concrete design rule for choosing streaming chunk sizes.","Replacing layer normalization with batch normalization and adding caching does not harm quality and, in the paper's internal comparisons, even improves objective metrics.","The relationship between latency and throughput depends on hardware cache hierarchy and memory bandwidth, so the optimal chunk size and vocoder architecture should be chosen according to the target device.","The paper's analysis is performed on a single AMD EPYC 7302 CPU using ONNX Runtime, and the practical recommendations assume this behavior generalizes to other resource-constrained CPUs and inference backends.","The multi-stream decomposition's benefit may extend beyond efficiency: partitioning the spectrogram into smaller sub-spectrograms makes spectral structures more locally consistent, which the paper argues improves how 2D convolutional kernels capture dependencies."],"supporting_citations":[{"why":"Base architecture Wavehax, which MS-Wavehax extends with multi-stream synthesis and the harmonic prior","marker":"[12]"},{"why":"Source of multi-stream synthesis with trainable filters, the core decomposition technique","marker":"[13]"},{"why":"Vocos architecture and training setup, the 1D-convolution comparison baseline","marker":"[7]"},{"why":"HiFi-GAN V1 baseline that MS-Wavehax is compared against in size, throughput, and quality","marker":"[3]"},{"why":"iSTFTNet baseline, another iSTFT-based vocoder compared in the latency-throughput analysis","marker":"[4]"},{"why":"Multi-band iSTFT extension, which motivates splitting the final layer into four sub-spectrograms","marker":"[27]"},{"why":"Ring-buffer caching mechanism used to make all convolutional, STFT, and iSTFT layers streaming-friendly","marker":"[28]"},{"why":"Batch normalization replaces layer normalization to enable streaming inference without sequence-wide statistics","marker":"[30]"},{"why":"JVS corpus used for the multi-speaker analysis-synthesis evaluation","marker":"[31]"},{"why":"Harvest algorithm for F0 extraction, which provides the harmonic prior input to Wavehax and MS-Wavehax","marker":"[32]"}],"fun_headline_variants":["Tiny 0.33M vocoder tops CPU streaming quality and speed","Multi-stream Wavehax: 2.4% size, top quality, low latency","CPU streaming: 0.33M-param vocoder beats all in quality","Sub-10ms lookahead, 40x smaller, top speech quality","Vocoder for edge: 0.33M params, highest quality, real-time"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The central throughput claim assumes that the im2col-based memory-bandwidth behavior measured on a single AMD EPYC 7302 CPU with ONNX Runtime generalizes to other resource-constrained devices; if a target device's cache hierarchy or convolution backend changes the balance between data transfer and compute, the reported throughput ranking and recommended chunk sizes could change.","fun_headline_variants_meta":{"raw":{"variants":["Tiny 0.33M vocoder tops CPU streaming quality and speed","Multi-stream Wavehax: 2.4% size, top quality, low latency","CPU streaming: 0.33M-param vocoder beats all in quality","Sub-10ms lookahead, 40x smaller, top speech quality","Vocoder for edge: 0.33M params, highest quality, real-time"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000952,"raw_usage":{"total_tokens":4073,"prompt_tokens":973,"completion_tokens":3100,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":589,"completion_tokens_details":{"reasoning_tokens":2992}},"tokens_in":589,"tokens_out":3100,"duration_ms":23447,"temperature":1.0,"reasoning_tokens":2992,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:59:23.233300+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same streaming RTF comparison on a low-end mobile CPU (e.g., an ARM Cortex-A series core with small L1/L2 caches and no large shared L3) using the same ONNX Runtime pipeline and chunk sizes of 1, 4, 8, and 16 frames. If Vocos already matches or beats MS-Wavehax at chunk sizes below 8 frames, or if MS-Wavehax fails to keep the lowest RTF below 80 ms latency, then the paper's central 'highest throughput under low-latency' claim does not generalize beyond the tested EPYC configuration. A second check: switch the convolution backend from im2col-based GeMM to direct convolution on the same EPYC CPU and see whether the RTF crossover between MS-Wavehax and Vocos shifts or disappears, which would confirm or falsify the proposed memory-bandwidth mechanism.","supporting_citations":[{"cited_title":"iSTFTNet: Fast and Lightweight Mel-Spectrogram V ocoder Incorporating Inverse Short-Time Fourier Transform,","cited_arxiv_id":null,"evidence_quote":"Base architecture Wavehax, which MS-Wavehax extends with multi-stream synthesis and the harmonic prior"},{"cited_title":"iSTFTNet2: Faster and More Lightweight iSTFT-Based Neural V ocoder Using 1D- 2D CNN,","cited_arxiv_id":null,"evidence_quote":"Source of multi-stream synthesis with trainable filters, the core decomposition technique"},{"cited_title":"Our analysis revealed that streaming throughput de- pends on overhead from data and parameter loading as well as computational complexity","cited_arxiv_id":null,"evidence_quote":"Vocos architecture and training setup, the 1D-convolution comparison baseline"},{"cited_title":"throughput We analyze the relationship between latency and throughput via block streaming synthesis using several neural vocoders","cited_arxiv_id":null,"evidence_quote":"HiFi-GAN V1 baseline that MS-Wavehax is compared against in size, throughput, and quality"},{"cited_title":"Wavehax and MS-Wavehax utilized F0 for generating prior signals, whereas other models concate- nate it with the mel-spectrogram, resulting in a 101-dimensional input feature","cited_arxiv_id":null,"evidence_quote":"iSTFTNet baseline, another iSTFT-based vocoder compared in the latency-throughput analysis"},{"cited_title":"High Qual- ity Streaming Speech Synthesis with Low, Sentence-Length- Independent Latency,","cited_arxiv_id":null,"evidence_quote":"Multi-band iSTFT extension, which motivates splitting the final layer into four sub-spectrograms"},{"cited_title":"A ConvNet for the 2020s,","cited_arxiv_id":null,"evidence_quote":"Ring-buffer caching mechanism used to make all convolutional, STFT, and iSTFT layers streaming-friendly"},{"cited_title":"Multi-band MelGAN: Faster Waveform Generation for High-Quality Text-to-Speech,","cited_arxiv_id":null,"evidence_quote":"Batch normalization replaces layer normalization to enable streaming inference without sequence-wide statistics"},{"cited_title":"Fre-GAN: Adversar- ial Frequency-Consistent Audio Synthesis,","cited_arxiv_id":null,"evidence_quote":"JVS corpus used for the multi-speaker analysis-synthesis evaluation"},{"cited_title":"Harvest: A High-Performance Fundamental Fre- quency Estimator from Speech Signals,","cited_arxiv_id":null,"evidence_quote":"Harvest algorithm for F0 extraction, which provides the harmonic prior input to Wavehax and MS-Wavehax"}],"review_version":1}