Hybrid Spoofed Audio Dataset (HSAD)

• HSAD is a benchmark paradigm that combines human, cloned, AI-generated, and hybrid (mixed) audio samples to evaluate spoofing beyond binary labels.
• It features detailed annotations—including speaker ID, spoofing method, and segment-level metadata—to support precise anti-spoofing evaluations.
• HSAD reveals calibration challenges in pretrained models, highlighting the need for dataset-specific adaptations and fine-tuning to improve detection performance.

Searching arXiv for the relevant papers and closely related dataset work. arXiv search query: "Hybrid Spoofed Audio Dataset HSAD audio spoof detection hybrid" Hybrid Spoofed Audio Dataset (HSAD) denotes a line of spoofed-audio benchmarks organized around the observation that realistic attacks are often not fully synthetic utterances. In the literature, the term is used explicitly for a benchmark containing human, cloned, zero-shot AI-generated, and mixed human–synthetic speech, and more broadly for hybrid corpora that combine multiple spoof types, sources, or annotation layers such as replay conditions or linguistically informed labels (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025, Khanjani et al., 2024). The unifying idea is that spoof detection should be evaluated beyond the conventional binary setting in which an entire file is either bona fide or wholly fake.

1. Terminology and conceptual scope

The acronym HSAD is not uniformly used across the literature. One 2024 study consistently refers only to “a hybrid dataset” or “our hybrid dataset”, describing it as “a hybrid dataset comprised of multiple types of spoofed audio augmented with sociolinguistic annotations,” but it does not introduce the name Hybrid Spoofed Audio Dataset (HSAD) or any HSAD acronym (Khanjani et al., 2024). By contrast, later 2025 work explicitly introduces and uses Hybrid Spoofed Audio Dataset (HSAD) as the formal name of a benchmark for mixed AI–human speech (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025).

The term “hybrid” itself is used in several technically distinct senses. In the named HSAD work, it primarily denotes utterances that mix genuine human segments with cloned or AI-generated segments inside one file, together with parallel clean and degraded conditions (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025). In the 2024 linguistically annotated hybrid dataset, “hybrid” refers to a corpus that combines replay attacks, mimicry, Text-to-Speech (TTS) deepfakes, Voice Conversion (VC) deepfakes, and genuine speech, while also adding Expert Defined Linguistic Features (EDLFs) (Khanjani et al., 2024). Related research extends the same idea to partially spoofed utterances that mix bona fide and spoofed segments (Zhang et al., 2021), replay-aware corpora containing bona fide, replayed bona fide, fake, and replayed fake speech (Zhang et al., 22 Oct 2025), and component-level mixtures where speech and environmental sound can be spoofed independently (Zhang et al., 19 Sep 2025).

This suggests that HSAD is best understood not as a single immutable dataset specification, but as a benchmark paradigm for mixed-source, multi-condition, or partially manipulated audio in anti-spoofing research.

2. Canonical HSAD formulation in 2025 work

The most explicit HSAD formulation is given in 2025 work that constructs a dataset around 12 participants, ages 19–38, with gender balanced recording, in an acoustically controlled indoor environment using studio-grade USB microphones, recorded at 44.1 kHz, then downsampled to 16 kHz (Huang et al., 21 May 2025). Each speaker records 104 scripted sentences across 8 categories: Alphanumeric combinations (8), Pure alphabetic sentences (8), Numeric sequences (8), Natural English sentences (16), Semantically coherent sentence pairs (16), Semantically unrelated sentences (16), Sentences with grammatical errors (16), and Complex sentences with semantic + grammatical anomalies (16) (Huang et al., 21 May 2025).

The core HSAD class structure comprises four balanced classes: Human, Cloned, AI-generated, and Hybrid, with the Clean Hybrid Dataset containing 1,248 utterances total, namely 312 utterances per class (Huang et al., 21 May 2025). A parallel Noisy/Degraded Dataset contains another 1,248 utterances, obtained by creating one distorted version per clean utterance (Huang et al., 21 May 2025). A later paper on HSAD reports the same four classes and the same 1,248 clean utterances, but its abstract also states “1,248 clean and 41,044 degraded utterances,” while its detailed construction section still describes a 1,248-clean + 1,248-degraded setup and does not reconcile the larger degraded count (Hu et al., 9 Sep 2025). This discrepancy is one of the clearest textual ambiguities surrounding the named HSAD.

The synthetic portion of HSAD is divided into two principal categories. Cloned speech is generated with a Tacotron 2-based TTS system plus speaker embeddings, using four cloning fidelity levels: $C_1$ for a single sentence reference, $C_2$ for a 16-sentence subset, $C_3$ for the full 104-sentence speaker corpus, and $C_4$ for the target sentence embedding (Huang et al., 21 May 2025). Pure AI-generated speech is produced with zero-shot TTS systems without explicit speaker identity preservation (Huang et al., 21 May 2025).

HSAD’s defining novelty is the hybrid utterance, constructed by segment-level concatenation under three composition patterns: H→S, S→H, and H↔S↔H, where transitions are smoothed with 10 ms cross-fading (Huang et al., 21 May 2025). Later analytical work on HSAD retains the same core idea and further refines evaluation with a six-group taxonomy: G1 – Genuine Human, G2 – Pure AI Clone, G3 – Pure AI Generated, G4 – Mixed: AI Generated + Human, G5 – Mixed: AI Clone + AI Generated, and G6 – Human Recombined (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025).

3. Annotation, metadata, and protocol design

HSAD is structured as more than a four-way label set. Each utterance is annotated with speaker ID, gender, spoofing method, and, for mixed examples, segment arrangement such as H→S, S→H, or H↔S↔H (Huang et al., 21 May 2025). In the 2025 evaluation-oriented HSAD description, hybrid examples are also annotated with segment start and end boundaries, source type per segment, and degradation metadata (Hu et al., 9 Sep 2025). At the same time, both papers emphasize that the benchmark is still used primarily for utterance-level classification, and that segment-level timestamps are a planned extension rather than a fully realized localization protocol (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025).

The degraded partition is produced through three explicit perturbation families. First, additive noise is injected using noise types such as street, café ambiance, and white noise, at 10, 15, 20, 30 dB SNR (Huang et al., 21 May 2025). Second, channel simulation includes low-pass filtering at 4 kHz (Huang et al., 21 May 2025). Third, compression artifacts are introduced via Opus at 16 kbps and 24 kbps (Huang et al., 21 May 2025). Each distorted utterance inherits the original class label and additionally carries SNR level, noise type, and degradation parameters such as codec and filter settings (Huang et al., 21 May 2025).

The main experimental protocol is speaker-disjoint. The papers describe an 80/20 train–test split, with 10% of training held out for validation, so that speakers in test are different from speakers in train (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025). This design is intended to suppress trivial voice memorization and to emphasize cross-speaker generalization under hybrid tampering. A later HSAD paper also states that the test set is subdivided into seen and unseen spoof types for cross-attack generalization, though it does not provide exact per-split counts beyond the 80/20 ratio and speaker disjointness (Hu et al., 9 Sep 2025).

A broader annotation tradition around hybrid spoof data appears in linguistically augmented corpora. The 2024 unnamed hybrid dataset and the later Linguistically Augmented Audio Speech Data (LinguAS) add five Expert Defined Linguistic Features (EDLFs)—pitch, pause, word-initial and word-final release bursts of consonant stops, audible intake or outtake of breath, and overall audio quality—encoded as binary indicators per clip (Khanjani et al., 2024, Keaton et al., 8 Jun 2026). In that line of work, hybrid spoof datasets are not only mixed by source type but also enriched with human-discernible linguistic metadata.

4. Benchmarking results and analytical use

HSAD is explicitly positioned as a stress test for pretrained audio models that appear strong on standard spoof benchmarks but fail under mixed-source conditions. In the HSAD analysis centered on transformer models, pretrained models overgeneralize and collapse under hybrid conditions, while spoof-specific fine-tuning improves separability but struggles with unseen compositions, and dataset-specific adaptation on HSAD yields large performance gains (Hu et al., 9 Sep 2025).

The strongest quantitative results reported for HSAD come from fine-tuned Audio Spectrogram Transformer (AST) models. In one study, Model A—a fine-tuned MIT-AST—reaches 97.88% validation accuracy, while Model B—a fine-tuned MattyB95 AST—reaches 98.08% validation accuracy; both achieve 97% test accuracy and F1-score ≈ 99% on the HSAD test split (Huang et al., 21 May 2025). The same study reports that both models correctly identified hybrid audio samples with 100% precision, that recall exceeds 88% for AI-generated and AI-cloned classes, and that human speech accuracy climbs to 84% after HSAD fine-tuning (Huang et al., 21 May 2025). It further reports a 71% reduction in false positives for human speech for Model A, from 1,207 to 197, and a 95% reduction in false negatives for AI-generated speech for Model B, from 14,117 to 684 (Huang et al., 21 May 2025).

The later HSAD analysis broadens the model set to six transformer-based architectures, including MIT-AST, MattyB95-AST, Wav2Vec2, and HuBERT (Hu et al., 9 Sep 2025). It reports a striking failure mode for pretrained MIT-AST: despite 93.67% aggregate accuracy, the model fails to correctly identify any genuine human utterances, effectively treating everything as spoofed (Hu et al., 9 Sep 2025). The same paper reports that pretrained Wav2Vec2-base-960h, Wav2Vec2-large-960h, and HuBERT-base-ls960 achieve only 20.06%, 3.52%, and 1.30% overall accuracy, respectively, before HSAD fine-tuning, and then rise to 97.94%, 97.26%, and 96.76% after HSAD-specific adaptation (Hu et al., 9 Sep 2025). The intended analytical conclusion is not merely that fine-tuning helps, but that HSAD reveals calibration failures, class collapse, and bias toward global spoof cues that standard binary datasets do not expose (Hu et al., 9 Sep 2025).

This empirical pattern closely aligns with the more general claim that models trained only on homogeneous real-versus-fake corpora tend to overestimate their robustness and fail in hybrid scenarios, especially when only part of the signal is spoofed, when spoofed segments are short, or when the utterance is carefully blended (Huang et al., 21 May 2025).

5. Relation to adjacent hybrid and partially spoofed corpora

HSAD belongs to a broader family of datasets that redefine spoofing as a mixed, structured, or partially localized phenomenon rather than a whole-utterance binary label.

Resource	Hybrid notion	Key structure
PartialSpoof	bona fide and spoofed segments in one utterance	segment- and utterance-level labels
EchoFake	synthetic plus physical replay	B, RB, F, RF
CompSpoof	spoofed speech and/or spoofed environment	5 component-level classes
LinguAS / 2024 hybrid dataset	multi-attack audio plus EDLFs	linguistic anomaly labels
Named HSAD	human, cloned, AI-generated, mixed audio	4 classes with degraded variants

PartialSpoof introduces a database in which utterances are only partially spoofed, with bona fide and spoofed segments mixed inside one file, and shows that countermeasures trained on fully spoofed utterances degrade substantially when tested on partially spoofed data, whereas training on partially spoofed data is reliable for both fully and partially spoofed utterances (Zhang et al., 2021). This is conceptually close to HSAD’s mixed human–synthetic utterances, although PartialSpoof is built from ASVspoof 2019 LA and is explicitly segment-labeled (Zhang et al., 2021).

EchoFake extends the hybrid idea from within-utterance mixing to attack-mode mixing, combining bona fide (B), replayed bona fide (RB), fake (F), and replayed fake (RF) speech under real acoustic playback and re-recording conditions (Zhang et al., 22 Oct 2025). Its role in the HSAD landscape is to show that hybridization can arise not only from segment concatenation but also from the interaction between synthesis and physical replay.

CompSpoof pushes the notion further toward component-level audio spoofing, where only the speech component, only the environmental sound component, both, or neither are spoofed (Zhang et al., 19 Sep 2025). Its five classes—original, bonafide_bonafide, spoof_bonafide, bonafide_spoof, and spoof_spoof—generalize the HSAD premise from segment mixtures to overlapping source components (Zhang et al., 19 Sep 2025).

LinguAS and the earlier unnamed hybrid dataset demonstrate that hybrid spoof corpora can also be organized around annotation richness rather than only mixed-source synthesis. LinguAS contains over 800 audio samples, each annotated with five EDLFs and accompanied by gender and generator metadata, while the 2024 unnamed hybrid dataset contains 344 audio samples balanced in terms of spoofed vs genuine, with 30% VC, 30% TTS, 25% replay attacks, and 15% mimicry (Keaton et al., 8 Jun 2026, Khanjani et al., 2024). In both cases, the hybrid property lies in combining multiple spoof families and adding a human-interpretable annotation layer.

At the design level, the major antecedent is ASVspoof 2019, which unifies Logical Access (LA) and Physical Access (PA) scenarios and treats synthetic, converted, and replayed speech within a common evaluation framework (Todisco et al., 2019, Wang et al., 2019). This multi-attack, known-versus-unknown philosophy has become a template for later hybrid spoof datasets, including HSAD-style resources.

6. Ambiguities, limitations, and future directions

The named HSAD literature is explicit about several limitations. The benchmark is built from only 12 speakers, is English only, and derives its clean recordings from a controlled indoor studio-like setup rather than in-the-wild capture (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025). Its spoof coverage is centered on TTS-based synthesis and AI voice cloning, with limited or no explicit inclusion of voice conversion, GAN-based attacks, prosody-only manipulations, or adversarial examples (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025). The current evaluation remains largely utterance-level, even though hybrid constructions motivate finer-grained temporal localization (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025).

A second limitation is terminological and documentary. One line of work describes a hybrid spoof corpus with rich sociolinguistic annotations but does not use the HSAD acronym at all (Khanjani et al., 2024). Conversely, the 2025 HSAD literature uses the acronym explicitly but contains a visible discrepancy between an abstract claiming 41,044 degraded utterances and a detailed construction section that consistently describes 1,248 degraded utterances (Hu et al., 9 Sep 2025). The existence of both statements in the same paper means that any encyclopedic treatment of HSAD must distinguish the dataset’s intended design from its unresolved reporting inconsistency.

The research directions proposed around HSAD are comparatively consistent. They include expanding spoof modalities to voice conversion, more TTS engines, multilingual speech, GAN-based synthesis, and adversarially crafted attacks; adding real-world recordings from smartphones, smart speakers, telephony systems, and varied acoustic environments; releasing segment-level annotations for localization; and conducting cross-corpus benchmarking against resources such as HAD and ADD2023-PF (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025). A related line of work on codec-resynthesized speech argues that future hybrid spoof datasets should also separate clean human speech, benign codec/vocoder processing, and fully synthetic codec-based deepfakes, because different labeling choices materially change detector behavior (Xiao et al., 18 Feb 2026). This suggests that the long-term evolution of HSAD is likely to involve not only larger scale and more attack families, but also more granular label taxonomies for ambiguous or borderline cases.

Within anti-spoofing research, HSAD is therefore best viewed as both a named benchmark and a broader design philosophy: a move away from homogeneous real-versus-fake corpora toward datasets that encode mixed provenance, partial manipulation, cross-condition degradation, and, increasingly, interpretable metadata that expose what current detectors actually learn (Huang et al., 21 May 2025, Hu et al., 9 Sep 2025, Khanjani et al., 2024).