spkrec-ecapa-voxceleb, 192-dim) with cosine similarity.
Methodology
- Genuine pairs: Three recordings of the target speaker (enrolled + two verify samples)
- Impostor condition A — human: 12 recordings from different speakers from the same audio corpus
- Impostor condition B — pre-existing AI: 13 recordings from pre-existing synthetic/cloned voices (non-IVC)
- Metric: Equal Error Rate (EER) — the threshold where FAR (false accept rate) equals FRR (false reject rate)
Results
Genuine speaker scores and impostor scores show no overlap at the operating threshold, with a wide separation margin between the two distributions.
Configuration
- Similarity metric: Cosine similarity on speaker embeddings
- Threshold: Tuned per deployment environment
- Liveness: f0 variation analysis (not applied in this test run)
Upcoming conditions
The following conditions are planned for future test runs:- XTTS-v2 open-source clone (local, $0 cost) — IVC using the target speaker’s enrollment audio
- Hesitation signature analysis — comparing disfluency/pause patterns in real speech vs TTS output