IronWall AI: A Chunk-Level, Explainable and Batched Architecture for Real-Time Audio Deepfake Detection

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Generative speech synthesis has advanced to a point where a short sample of a person’s voice is sufficient to produce convincing arbitrary utterances. The resulting threat surface is no longer hypothetical: voice-cloned social engineering, fraudulent executive instructions and fabricated political audio are documented attack patterns. Most publicly available countermeasures answer only a narrow question — given a complete audio file, is it real or fake? — and answer it with a single opaque number. Three practical gaps follow from that framing. First, a clip-level verdict is blind to partially synthetic audio, in which authentic speech is spliced with a short generated segment. Second, an unexplained probability is difficult to act on and impossible to audit. Third, a batch-oriented offline classifier cannot be placed inside a live call, which is precisely where the attack occurs. This paper presents IronWall AI, a full-stack detection system that addresses all three gaps in a single deployable architecture. The system wraps a TensorFlow SavedModel built on the multilingual wav2vec2-large-XLSR-53 self-supervised speech backbone and serves it through an asynchronous FastAPI inference server paired with a React single-page dashboard. Four design contributions are described and analysed. (i) A sliding-window analysis front end decomposes an utterance into overlapping five-second windows at a 2.5-second stride, so that localised synthetic content cannot be diluted by surrounding authentic speech. (ii) A multi-criteria aggregation rule replaces the conventional maximum-score or mean-score decision with a joint test over prevalence, strength and temporal persistence of synthetic evidence, coupled with a twenty-percent trimmed mean for the reported score; the explicit aim is suppression of false positives arising from coughs, clipping and codec artifacts. (iii) A template-driven natural-language explanation is generated from the same three statistics that produce the verdict, so the prose and the numbers can never disagree. (iv) A crossrequest dynamic batching scheduler packs windows drawn from different concurrent HTTP requests into a single forward pass, decoupling accelerator utilisation from request arrival patterns. A browserside streaming subsystem extends the same model to live microphone monitoring using client-side resampling, WAV encoding and in-flight backpressure. The system additionally ships a client-side evaluation module that computes confusion matrices, precision, recall, specificity, F1, ROC, AUC, DET curves and equal error rate — but computes them strictly from user-supplied ground-truth labels and returns null rather than substituting model predictions when labels are absent. We report a full architectural and source-level analysis of the implementation, present the aggregation mathematics in closed form, and document a set of IronWall AI | 1 divergences between the repository documentation and the executing code that materially affect reproducibility. We conclude by specifying the benchmark protocol required to convert the present engineering contribution into a validated empirical claim.