campplus-gender-classifier

Estimates whether a voice sounds female or male from a few turns of 16 kHz speech, with a calibrated confidence meant for gating decisions. It predicts perceived voice characteristics, not a person's gender identity.

• Base: WeSpeaker CAM++ (VoxCeleb, Apache-2.0); fine-tuned: xvector.block3, xvector.transit3, xvector.out_nonlinear, xvector.dense, head
• Output: p = sigmoid(mean turn logit / T), T = 0.820; confidence = |2p - 1|
• Training audio augmented with noise, babble, an opposite-sex interferer, gain/clipping and an 8 kHz mu-law phone path
• Sources: edacc, kathbath_hi, kathbath_ml, libritts, vctk (speaker-disjoint fit / calibration / test)

Gate on simulated calls (held-out speakers)

confidence >=	answers on	wrong
0.8	90.8%	2.40%
0.9	86.1%	1.97%
0.95	81.0%	1.40%
0.98	71.7%	1.17%

The test labels include known LibriTTS speaker-metadata errors (speakers whose recorded gender is flipped, confirmed by listening), so these error rates are upper bounds. Per-source and per-condition results: metrics.json.

Use

from inference import GenderClassifierModel
model = GenderClassifierModel(".")
model.predict([turn1_int16, turn2_int16])   # -> (p_female, confidence) or None

PyTorch weights for further fine-tuning: model.safetensors, loaded with model.GenderClassifier.from_folder(".").

Training data

LibriTTS-R (gender from parler-tts speaker descriptions), VCTK, EdAcc and Kathbath (Hindi, Malayalam). Each dataset is under its own licence; see its page.