spectra-aasist3-int8-audio-deepfake

Spectra-AASIST3 INT8 — Academic Deployment Derivative

Base model: lab260/Spectra-AASIST3 Original revision: bc0ded888080ddad493177bb53aa6f5b95219d7c Original license: Apache-2.0

This repository hosts a dynamic INT8 ONNX derivative of lab260/Spectra-AASIST3, produced for a university Final Year Project ("Detecting AI-Generated and Cloned Voices: A Deep Learning System for Robust Audio Deepfake Detection", GitHub: munz-aftab/audio-deepfake-detection-fyp), to make CPU-only browser-based deployment (Streamlit Community Cloud) practical without a GPU.

Derivation

• No retraining. No fine-tuning. This is a numerical transform of the original FP32 checkpoint's weights only.
• Produced with ONNX Runtime dynamic INT8 quantization (onnxruntime.quantization.quantize_dynamic, op_types_to_quantize=['MatMul'], weight_type=QInt8, per_channel=False) — primarily quantizing the wav2vec2 XLS-R-300m transformer's MatMul/linear-projection weights, which carry 95.4% of the original model's weight bytes. Convolutional layers (the RawNet2-style encoder, ~4% of weight bytes) are left in FP32.
• Generated from the original FP32 ONNX artifact (spectra-aasist3.onnx, SHA256 5f05c29a01ad80c702b32654db87c2aa6e467c11c67b6d47f2fac873f846cae9), itself an export of lab260/Spectra-AASIST3's model.safetensors.
• CPU deployment target. onnxruntime CPU execution provider only; never tested with or intended for CUDA/GPU.

Architecture

wav2vec 2.0 XLS-R-300m self-supervised front-end → single-layer MLP bridge → KAN-enhanced AASIST (KAN-AASIST) back-end. See the base model card for full architectural detail; this repository changes only the numerical precision of the exported ONNX graph's weights, not the architecture.

Input

• Mono audio, 16 kHz
• Pre-emphasis, coefficient 0.97, applied to the full waveform (y[0] = x[0], y[n] = x[n] - 0.97*x[n-1] for n ≥ 1)
• Exactly 64,600 samples (~4.04s) required by the graph
• Long clips: deterministic first-64,600-sample window (author- compatible scoring convention)
• Short clips: tile-repeat (not zero-pad) until ≥64,600 samples, then truncated to exactly 64,600

Output

• 2 logits, wav → logits
• Class/logit index 1 = bona fide. Higher bona-fide score/logit means more likely genuine speech. (Verified directly from the base model's own source code — see the base model's model.py, SpectraAASIST3.classify().)

Evaluation — FYP project-measured results on a fixed balanced In-the-Wild evaluation subset

These are NOT the original model authors' benchmark results, and are not claimed to represent universal real-world performance. They were measured by this FYP on a small (200-clip), balanced, deterministically sampled subset of the SpeechAntiSpoofingBenchmarks/InTheWild dataset, disjoint from a separate calibration subset used to select the operating threshold. Full methodology: docs/spectra_aasist3_evaluation.md and docs/spectra_production_optimization.md in the project's GitHub repository.

Do not treat 3.0% EER as a universal, dataset-independent performance guarantee — it is a single-subset, project-measured number, reported for transparency and reproducibility, not as a general benchmark claim.

Limitations

• The base model, lab260/Spectra-AASIST3, is a pre-release/unpublished detector — there is no peer-reviewed paper for this exact architecture as of this writing.
• Domain shift (languages, recording conditions, unseen synthesis methods) remains possible; the evaluation above covers one real-world corpus at one point in time.
• Not forensic evidence. Not suitable as the sole basis for legal, security, or identity-verification decisions.
• Output scores are raw logits/softmax values, not inherently calibrated probabilities — a "0.9" score is not literally "90% likely genuine."
• Long-audio (multi-window) aggregation is an application-level extension built by this FYP, not part of the original model or its authors' evaluation methodology.

Files

• spectra-aasist3-int8-dynamic.onnx — the quantized model
• quantization_metadata.json — full reproducibility metadata (base revision/hash, ONNX Runtime version, quantized operator counts, resulting hash/size, evaluation summary, date)

Attribution

Base model and architecture: Kirill Borodin / lab260 (lab260/Spectra-AASIST3), Apache-2.0. Upstream SSL backbone: facebook/wav2vec2-xls-r-300m (Meta AI / Facebook, Apache-2.0). This derivative repository: produced as part of a university Final Year Project; not affiliated with the original authors.