SPARC: Anchored Residual Calibration for Speaker Verification
· Paper · Speaker Verification / Score Calibration / PyTorch / VAD / Language ID
SPARC is a portable calibration plug-in for frozen speaker-verification systems. It combines encoder scores with automatically derived duration, voice-activity, and language context, then learns a small residual around a frozen context-affine anchor.
The anchor-first design keeps the reference solution intact at initialization and adds capacity only when validation supports it. The resulting deployment bundle is small, explicit about encoder and context order, and does not require changing the upstream speaker encoders.