Deepfake Audio Detection using Deep Learning (ASVspoof)

Voice cloning tools can now synthesize anyone's speech from seconds of audio, and replayed recordings can fool voice authentication — the "is this voice real?" question has become a genuine security problem. Detecting spoofed audio is a well-defined machine-learning task with a respected benchmark: ASVspoof 2019, which packages text-to-speech, voice-conversion and replay attacks into logical-access and physical-access partitions. This project trains a light CNN (LCNN) on LFCC acoustic features to classify utterances as bonafide or spoof, evaluates with the challenge's equal error rate (EER) and tandem detection cost metrics, and delivers an audio-forensics web demo: select a clip, inspect its waveform and mel spectrogram, and get a verdict with the contributing artifact signals explained. The full methodology is documented for the viva.

1. ASVspoof 2019 utterances (16 kHz) are loaded from the logical-access and physical-access partitions with train/dev/eval splits as defined by the challenge.
2. 60-dimensional LFCC features (with delta and double-delta) are extracted per utterance and truncated/padded to a fixed 400-frame window.
3. An LCNN — five conv blocks with Max-Feature-Map activation, max-pooling, dropout, a dense layer and a 2-way softmax — is trained with cross-entropy.
4. Development-set scores pick the EER operating threshold; evaluation-set scores are computed once and scored with the official EER and t-DCF routines.
5. Per-attack EER is broken down by attack ID (A01–A19, replay) to show which spoof families are hardest.
6. In the web demo, a clip's waveform and mel spectrogram are rendered, the feature pipeline runs, and the model returns P(spoof) with the top contributing artifact signals.

Tech stack:

• Python 3, PyTorch
• librosa (LFCC / mel features)
• ASVspoof 2019 dataset
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training & evaluation)
• HTML5 canvas + JavaScript (forensics demo)

Dataset & model details

• Dataset: ASVspoof 2019 — logical access: 121,461 utterances (TTS/VC attacks A01–A19, plus bonafide); physical access: replay attacks. 16 kHz, defined train/dev/eval splits. From the ASVspoof consortium challenge.
• Task: Binary utterance classification; input = 60×400 LFCC feature matrix, output = P(spoof) vs P(bonafide).
• Model: Light CNN (LCNN): 5 conv blocks with Max-Feature-Map activation → max-pool ×2 → dropout → Dense(128, ReLU) → Dense(2, softmax). Approximately 0.5M parameters.
• Metrics: Equal error rate ~4.5% on eval (design target for the built-to-order training run), t-DCF, per-attack EER breakdown. No EER is claimed as measured until the training run is executed for the order.

• [LCNN spoof classifier] Light CNN with Max-Feature-Map activations on 60-dim LFCC features — the architecture family that performed strongly in the ASVspoof 2019 challenge.
• [Audio forensics demo] Waveform and mel-spectrogram visualization per clip, with a bonafide/spoof verdict and an artifact breakdown (vocoder shimmer, F0 anomalies, replay echo).
• [19 attack types covered] Training spans neural TTS (A01–A06), voice conversion (A07–A13, A16–A19) and replay attacks from both challenge partitions.
• [Challenge-standard evaluation] Equal error rate (EER) and t-DCF computed per the ASVspoof 2019 protocol, not ad-hoc accuracy.
• [Full training notebook] Feature extraction, model definition, training, and EER/t-DCF scoring in one reproducible notebook.
• [Test history log] The demo keeps a scored history of analyzed clips with verdicts and top artifacts.
• [Threshold analysis] Score distributions for bonafide vs spoof with the EER operating point marked.

• Complete training & evaluation Jupyter notebook
• Trained LCNN model file with LFCC extraction code
• Audio forensics web demo (verdict, artifacts, history)
• EER/t-DCF scoring scripts and per-attack breakdown plots
• Project report PDF (background, spoofing taxonomy, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (LFCC, MFM, EER, t-DCF, attack families)