Voice Biometric Authentication using Machine Learning

Passwords are the weakest link in everyday security: people reuse them, write them down and share them, while OTPs add friction and SIM-swap fraud keeps defeating SMS codes. Fingerprint and face biometrics work well, but they need dedicated sensors or a camera with good lighting, and they cannot authenticate someone over a plain voice call. The human voice carries a distinctive acoustic signature from the shape of each speaker's vocal tract That leaves a gap between the classroom and real biometric pipelines: enrollment, feature extraction, model scoring, threshold setting and the false-accept versus false-reject trade-off. This project closes that gap with a working voiceprint system, so those security concepts are demonstrated on real audio instead of defined on slides.

1. During enrollment, the user speaks for ~10 seconds into a microphone; the audio is captured at 16 kHz.
2. Pre-processing applies a noise gate, voice-activity detection (VAD) to drop silence, and amplitude normalization.
3. MFCC features (40 coefficients plus delta and delta-delta) are extracted in 25 ms frames with a 10 ms hop using librosa.
4. A GMM (16–64 mixtures) is trained per speaker on the enrollment features; alternatively, the CNN maps each utterance to a fixed-length 128-dim voiceprint embedding.
5. At login, the spoken sample is converted to the same features and scored against the claimed user's model — log-likelihood ratio for GMM, cosine similarity for the CNN embedding.
6. The score is compared with the threshold (set at the equal-error-rate point): above it the login is accepted, below it rejected. All attempts are logged with scores in the VoiceKey app.

• Voiceprint enrollment from ~10 seconds of spoken audio per user
• MFCC feature extraction with librosa (40 coefficients + delta/delta-delta, 25 ms frames)
• GMM-based speaker matching — one Gaussian Mixture Model trained per enrolled user
• CNN voiceprint classifier option (TensorFlow/Keras embedding + cosine similarity scoring)
• Text-independent verification — no fixed passphrase required
• Similarity score with configurable accept/reject threshold set at the EER point
• Anti-spoofing checks: replay-attack heuristics from spectral and energy cues
• VoiceKey Streamlit app: enroll, verify, and view the score gauge live
• Multi-user enrollment database with per-user models and attempt history
• Full metrics suite: FAR/FRR tables, DET curve and score histograms for the report

• Complete, commented Python source code (feature extraction, GMM + CNN training, Streamlit app)
• Trained speaker models and voiceprint database schema
• VoiceKey Streamlit web application with live demo mode
• Sample voice dataset for training and testing
• Project report PDF (literature survey, methodology, FAR/FRR results, conclusion)
• PPT presentation for final review
• Viva Q&A preparation document (MFCC theory, GMM scoring, biometric metrics)
• Installation and run guide

FAQs

1. How is this different from speech recognition? Speech recognition asks what was said; speaker verification asks who said it. This project models the speaker's voice characteristics — vocal-tract resonances captured in MFCCs — so the same phrase spoken by two people produces different scores.
2. Which model is better — GMM or CNN? GMMs are the classic, explainable baseline and train in seconds on CPU; the CNN embedding approach generalizes better across sessions and channels but needs more enrollment data. The report includes the procedure for comparing both on the same trial set.
3. What are FAR, FRR and EER? False Accept Rate (impostor accepted), False Reject Rate (genuine user rejected), and the Equal Error Rate point where the two balance — the standard biometric metrics. The app plots the DET curve and sets the default threshold at EER.
4. Can someone fool it by playing a recording? Casual replay is partially caught by the anti-spoofing heuristics (spectral flatness and energy-variation anomalies typical of loudspeaker playback). Full protection against voice cloning is listed as future work in the report.
5. Can I demo it live in my review? Yes — the VoiceKey app enrolls a voice in ~10 seconds and verifies attempts instantly with a visible score gauge. Enroll yourself and a friend, then show an impostor attempt being rejected.
6. How can I extend this project? Common extensions include: text-dependent passphrase mode, channel-robustness with data augmentation, on-device inference with TensorFlow Lite, multi-factor login (voice + OTP), and liveness detection.

• Background noise degrades accuracy — quiet-room enrollment and verification is recommended for the demo.
• Very short enrollment samples (<5 seconds) weaken the speaker model; ~10 seconds is the recommended minimum.
• Microphone quality matters — the same headset for enrollment and login gives the most consistent scores.
• Anti-spoofing uses statistical replay heuristics, not full liveness detection against advanced voice-cloning attacks.
• Voice characteristics shift with illness, fatigue or age — the model reflects the voice as enrolled.