{"id":"2cd2d5d4-a46b-42f5-93f6-9d1d29234f03","arxiv_id":"2509.09155","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"HISPASpoof is a new large Spanish synthetic-speech dataset for detection and attribution, with evidence that English-trained detectors fail on Spanish and Spanish training helps.","lead":"This paper introduces HISPASpoof, a large Spanish dataset for detecting and attributing AI-generated speech. It shows that detectors trained on English generalize poorly to Spanish, while training on HISPASpoof substantially improves detection.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Detection experiments may exploit recording-condition mismatch between real corpora and clean TTS audio, not synthesis artifacts.","rationale":"The reader's weakest assumption—unexamined recording-condition mismatch between real and synthetic audio—is precisely the most load-bearing vulnerability. The paper's central contribution is a benchmark and the demonstration that training on it improves detection; if the benchmark contains trivial channel-level cues, the performance numbers are not evidence of deepfake detection ability, and the cross-lingual conclusions are also confounded. This is not an external-consensus disagreement but an internal validity threat: Table II sampling-rate disparities and the use of three different real-speech corpora with distinct acoustic conditions make the channel artifact almost inevitable unless explicitly controlled. The paper provides no such control or analysis. The proposed concrete test—applying a common band-limiting/noise-matching transform to both classes—would directly reveal whether the low EERs in Table VIII persist when channel cues are removed. Since the reader already assigned CONDITIONAL and identified this same assumption, my stress-test reinforces that verdict without changing it; the concern is serious but addressable by additional experiments or revised claims, so REJECT is not warranted yet.","tokens_in":12199,"tokens_out":7516,"duration_ms":94845,"concrete_test":"Re-train the best-performing detector (e.g., Spec-ResNet) on HISPASpoof after applying a common narrowband transform to every signal in both real and synthetic classes: resample all audio to 16 kHz and low-pass filter at 4 kHz (or, alternatively, add a measured room impulse response and background noise to the synthetic class to match real-corpus channel statistics). Evaluate on the same U_HIS test set. If EER rises from <1% to above ~15% while ASVspoof2019 results remain similar, the original gains are largely attributable to recording-condition mismatch rather than synthesis artifacts. A complementary check is to train a channel-only classifier (e.g., on average log-mel spectrum or estimated SNR) and see if it achieves EER below 5% on the same split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that training on HISPASpoof substantially improves Spanish deepfake detection rests on the assumption that real and synthetic speech differ primarily in synthesis artifacts. In §III-C, real speech is drawn from VoxPopuli (16 kHz), CIEMPIESS (16 kHz), and crowdsourced Latin American corpora (48 kHz), while synthetic speech is generated cleanly by six TTS systems at 16–44.1 kHz (Table II). No resampling, channel normalization, or noise augmentation is described. The detection test set U_HIS includes unseen speakers and unseen generators, but real and synthetic signals systematically differ in background noise, microphone, bandwidth, and recording environment. A detector can therefore separate classes by learning channel statistics (e.g., spectral tilt, noise floor, high-frequency energy) rather than synthesis artifacts. This is consistent with LFCC-GMM—limited to 4 kHz features—achieving 1.57% EER and Spec-ResNet 0.72% EER on unseen speakers/generators (Table VIII), which is unusually low for a cross-generator detection task. The same confound affects cross-lingual comparisons: ASVspoof2019 real audio is studio-quality VCTK, so poor Spanish generalization may reflect channel shift rather than language mismatch. Without controlling for channel cues, the reported detection improvements do not establish real-world synthetic speech detection ability, weakening the dataset's validity as a benchmark.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces HISPASpoof, a large-scale Spanish-language dataset for synthetic speech detection and attribution. It contains real speech from six Spanish accents (taken from VoxPopuli, CIEMPIESS, and crowdsourced Latin American corpora) and synthetic speech from six zero-shot TTS systems, with a total of about 535k signals. The authors evaluate five detection methods (LFCC-GMM, MFCC-ResNet, Spec-ResNet, PaSST, Wav2Vec2-AASIST) under four training conditions: English ASVspoof2019, multilingual ODSS, the Spanish subset of ODSS, and HISPASpoof. They report EERs on test sets with unseen speakers and unseen generators, and also report closed-set and open-set attribution results. The central claims are that English-trained detectors generalize poorly to Spanish and that training on HISPASpoof substantially improves Spanish synthetic speech detection.","tokens_in":12521,"tokens_out":3793,"duration_ms":42082,"significance":"If the claims hold, HISPASpoof is a valuable public resource: it is the first large-scale Spanish dataset supporting both detection and attribution, it covers multiple accents and modern zero-shot TTS systems, and the paper provides baseline results for several representative detectors. The dataset release and the use of seen/unseen speaker-generator splits are strengths. However, the validity of the benchmark depends critically on whether the detection improvements measure synthesis artifacts rather than incidental recording-condition differences between the real and synthetic speech, and on the statistical reliability of the reported results given the small unseen-speaker/generator test set.","major_comments":[{"comment":"The real and synthetic speech are not matched for recording conditions. Real speech comes from VoxPopuli (16 kHz), CIEMPIESS (16 kHz radio), and crowdsourced Latin American corpora (48 kHz), while synthetic speech is generated cleanly at 16–44.1 kHz. No resampling, filtering, channel normalization, or noise augmentation is described. A detector could therefore succeed by exploiting bandwidth, noise floor, or channel statistics rather than synthesis artifacts. This concern is reinforced by the unusually low EERs on the unseen test set in Table VIII (LFCC-GMM 1.57%, Spec-ResNet 0.72%), especially for a cross-generator task. To support the central claim, the authors should add channel-matched controls, e.g., resampling all audio to a common rate, adding noise/channel augmentation, and reporting performance on low-band versus full-band features or on a real-speech-only/channel-shift control.","section":"§III-C, Tables I–II, Table VIII"},{"comment":"All detection and attribution results are reported as single-run EER or accuracy values with no confidence intervals, error bars, or significance tests. The HISPASpoof unseen test set is built from only six unseen speakers and two held-out generators (Section III-C), so the reported values such as 0.72% EER are point estimates with potentially large uncertainty. The paper should provide error bars across multiple training runs or bootstrap over speakers/generators, and should report per-speaker and per-generator breakdowns for the key comparisons, at least for the HISPASpoof-trained models in Table VIII.","section":"§V, §VII, Tables V–X"},{"comment":"The open-set threshold δ is selected on a 10% held-out portion of the test set, but the manuscript does not state whether the final metrics in Table X are computed on the remaining 90% or on the entire test set. If the entire test set is used, the threshold selection leaks test information and the open-set attribution numbers are optimistically biased. The authors should define a dedicated validation split for threshold selection and evaluate only on a fully disjoint test split, or clearly report both the threshold-tuning portion and the final evaluation portion.","section":"§VI, Experiment 2 (Open-Set Attribution)"}],"minor_comments":[{"comment":"Typo: \"contains both synthetic and and real speech\" should read \"contains both synthetic and real speech.\"","section":"§III-C"},{"comment":"Typo: \"V oice Cloning\" should be \"Voice Cloning.\"","section":"§I"},{"comment":"Reference [47] is a duplicate of reference [19] (same title, same authors, same venue). Please merge or differentiate.","section":"References"},{"comment":"Inconsistent naming of FishSpeech vs Fish-Speech across the text and tables; please unify.","section":"§III-C, Table II"},{"comment":"The relation between the detection subset (which uses exact transcripts of real signals) and the attribution subset (which uses ChatGPT-generated text) should be clarified, since content mismatch can affect both detection and attribution generalization.","section":"§III-C"},{"comment":"The confusion matrix figure is referenced but the axes and normalization are not fully described in the text; please ensure the figure has clear labels and captions.","section":"Fig. 1"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The short version: this dataset is a real contribution, but the detection numbers are less solid than they look. HISPASpoof is the first large-scale Spanish corpus for both synthetic speech detection and attribution, with 535k signals, six accents, and six TTS systems. That genuinely fills a gap. The paper is honest about the limited ODSS Spanish material and designs proper unseen-speaker/unseen-generator splits, which is good practice.\n\nWhat it does well: content-matched synthetic speech (same transcript), gender-balanced speakers, open-set attribution setup with two held-out generators, and evaluation across five different detector families. The finding that English-trained detectors do badly on Spanish is plausible and consistent with prior cross-lingual results.\n\nThe soft spot: real speech comes from VoxPopuli, CIEMPIESS, and crowdsourced Latin American corpora, with sampling rates of 16–48 kHz, while synthetic speech is clean TTS output at 16–44.1 kHz. The paper does not describe resampling, channel normalization, or noise matching. So the low EERs on unseen generators — 0.72% for Spec-ResNet, 1.57% for LFCC-GMM — could reflect bandwidth/noise/recording-condition cues rather than synthesis artifacts. That doesn't sink the dataset, but it means the headline detection improvements are not yet established as deepfake detection ability. The open-set attribution also fits the threshold δ on a 10% held-out slice of the test set, and the unseen test speakers are just six people, so the generalization claims rest on a narrow base.\n\nThe paper also reports single-run EERs with no error bars. That is common in this corner of the literature, but it makes the fine-grained comparisons (e.g., 1.57% vs 0.72%) less meaningful.\n\nBottom line: if you work on speech forensics, this is worth having as a benchmark. I would send it to review, but I'd ask the authors to address the channel mismatch directly or soften the detection claims.","headline":"Useful new Spanish deepfake dataset, but the detection gains may partly be a channel artifact; worth engaging as a benchmark, not as proof of detector skill.","tokens_in":12988,"tokens_out":2847,"would_cite":true,"duration_ms":34975,"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":"HISPASpoof introduces the first large-scale Spanish dataset for synthetic speech detection and attribution, and training on it substantially improves Spanish deepfake detection.","keywords":["synthetic speech detection","speech forensics","Spanish deepfake detection","text-to-speech","voice cloning","attribution","zero-shot TTS","benchmark dataset"],"falsifier":"Train a detector on HISPASpoof and evaluate it on Spanish real speech that has been passed through the same codec, resampling, and channel simulation as the synthetic samples, while keeping speakers and generators unseen. If equal error rate jumps toward chance levels, the original gains came from channel artifacts. Alternatively, train on HISPASpoof and test on a matched recording of the same speakers under identical microphone conditions.","tokens_in":12122,"feed_emoji":"🎙️","tokens_out":4576,"duration_ms":45474,"temperature":0.7,"pith_summary":"The paper claims that Spanish synthetic speech is poorly handled by detectors trained on English or on small multilingual corpora, and that the gap is a data problem rather than an unsolvable technical one. To close it, the authors build HISPASpoof, a benchmark of real and synthetic Spanish speech spanning six accents and six modern zero-shot voice-cloning systems, with both detection and attribution subsets. They show that every tested detector improves sharply when trained on HISPASpoof, with error rates dropping from over 40% to under 5% on the Spanish test set. The dataset also enables attribution, identifying which synthesizer generated a clip, near-perfectly when the generator is known and with usable accuracy for unseen generators. If the benchmark holds up, it gives Spanish-speaking communities a foundation for detecting voice deepfakes and for standardizing future forensics research.","feed_headline":"First Spanish deepfake speech dataset beats English-trained models","feed_subtitle":"Six accents and six synthesizers show English-trained detectors miss Spanish deepfakes.","key_machinery":"The load-bearing object is the dataset itself, engineered for controlled generalization tests. HISPASpoof pairs real utterances from public corpora with synthetic versions of the exact same transcripts, removing content bias, and separates speakers and generators into seen/unseen groups so that test performance reflects genuine generalization to new voices and new synthesizers. The evaluation machinery is the battery of five detector architectures spanning feature-based (LFCC-GMM), image-based (MFCC-ResNet, Spec-ResNet, PaSST), and waveform-based (Wav2Vec2-AASIST) approaches, trained under four language conditions and scored by equal error rate for detection and accuracy/F1 for attribution.","core_discovery":"On the paper's own terms, the central discovery is that synthetic speech detection is language-sensitive, and a large-scale language-specific dataset can overcome that sensitivity. The authors introduce HISPASpoof with 6,241 real Spanish speech signals and 37,446 synthetic signals for detection (43,687 total), plus an attribution subset with 492,000 generated samples. Real speech is drawn from six Spanish accents (Peninsular, Argentinian, Colombian, Mexican, Chilean, Peruvian); synthetic speech is produced by six zero-shot TTS systems (ElevenLabs, F5-Spanish, FishSpeech, XTTS-v1.1, XTTS-v2, YourTTS). Using disjoint unseen speakers and two unseen generators, they evaluate five representative","pith_inferences":["Editorial inference: because real speech comes from different public corpora with different recording conditions and sampling rates, part of the reported gains may come from detectors learning channel or noise cues rather than synthesis artifacts; a channel-matched test would quantify this.","Editorial inference: the confusion between XTTS-v1 and XTTS-v2 in open-set attribution suggests architecture-level fingerprints; probing a continuum of XTTS checkpoints could reveal how attribution confidence degrades with model similarity.","Editorial inference: the accent-balanced design invites accent-sensitivity studies, e.g., training on Peninsular and testing on Mexican Spanish, to identify which synthesis artifacts are accent-invariant."],"forward_implications":["Detectors trained only on English should not be trusted for Spanish speech; HISPASpoof-trained models cut EER from over 40% to under 5%.","Language-specific data is a practical lever: even classical GMM and small ResNet detectors become strong Spanish detectors with enough matched training data.","Attribution is feasible: known generators are identified near-perfectly, and unseen generators can be flagged as unknown with about 78% accuracy by the best model.","The benchmark's seen/unseen split makes it a reusable testbed for measuring generalization to new voices and new synthesis methods.","Cross-lingual generalization is asymmetric: training on Spanish and testing on English hurts less than the reverse, suggesting shared acoustic knowledge partially transfers."],"fun_headline_variants":["Spanish deepfake audio: English detectors fail, new dataset fixes it","New Spanish speech dataset exposes English-trained AI blind spots","HISPASpoof: 43k Spanish speech samples outsmart English detectors","First Spanish deepfake detection dataset: 6 accents, 6 synthesizers","English-trained deepfake detectors stumble on Spanish—new dataset helps"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The benchmark's validity assumes that real speech from public corpora and the TTS-generated speech differ primarily in synthesis artifacts rather than in recording conditions, channels, or speaker-matching quality; if detectors exploit those incidental differences, the reported improvements would not measure deepfake-detection ability.","fun_headline_variants_meta":{"raw":{"variants":["Spanish deepfake audio: English detectors fail, new dataset fixes it","New Spanish speech dataset exposes English-trained AI blind spots","HISPASpoof: 43k Spanish speech samples outsmart English detectors","First Spanish deepfake detection dataset: 6 accents, 6 synthesizers","English-trained deepfake detectors stumble on Spanish—new dataset helps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000172,"raw_usage":{"total_tokens":1091,"prompt_tokens":700,"completion_tokens":391,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":444,"completion_tokens_details":{"reasoning_tokens":313}},"tokens_in":444,"tokens_out":391,"duration_ms":4627,"temperature":1.0,"reasoning_tokens":313,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T19:34:26.399312+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Train a detector on HISPASpoof and evaluate it on Spanish real speech that has been passed through the same codec, resampling, and channel simulation as the synthetic samples, while keeping speakers and generators unseen. If equal error rate jumps toward chance levels, the original gains came from channel artifacts. Alternatively, train on HISPASpoof and test on a matched recording of the same speakers under identical microphone conditions.","supporting_citations":[],"review_version":1}