Answers are grounded in this paper and cite their sources.
Fraud, impersonation and misinformation are being increasingly leveraged using synthetic speech generated by stateof-the-art speech-to-text and voice conversion systems, but the majority of the detectors are designed for English corpora. This paper introduces an end-to-end VoiceShield system which classifies speech as real or fake for five Indian languages (Hindi, Marathi, Urdu, Kannada and Bengali) using a single multilingual classifier. Audio is standardized to 16 kHz mono, segmented in 4-s chunks with 2-s hop, and transformed to 128-band spectrograms on a log Mel scale, and classified using a compact, 0.39 million parameter convolutional neural network (CNN). By recording-level stratified splitting, leakage of segments does not occur, and by using the Gradient-weighted Class Activation Mapping (Grad-CAM), the time–frequency regions responsible for each decision can be highlighted. The trained network is exported to TorchScript and served using a FastAPI back end which has 2 modes: file upload with per-segment analysis, and WebSocket based real-time monitoring, and a React front end. The curated catalogue covers 10,448 recordings (6,751 real, 3,697 fake). The average time spent on model inference on a CPU is 4.03 ms (99th percentile 6.17 ms) with an interactive budget of 200 ms per 4-s segment. We also describe the evaluation protocol to train the whole corpus and talk about threats to validity, especially the source imbalance between real and fake classes.
No takes yet. Share an insight, caveat, or question.
Mungelwar et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: