tropa-voice-mini: Speech Anti-Spoofing Arena wrapper
Packaged for the Speech Anti-Spoofing Arena from the authors' code.
No paper: the model sits in the Arena's unpublished tier (shown, not ranked).
tropa-voice-mini (wasitaigenerated.com, 2026): the open-weights AI-voice detector behind the wasitaigenerated API's small tier. A wav2vec 2.0 XLS-R 300M encoder fine-tuned end-to-end, with a multi-layer attention head (SLS head): softmax gate over all 25 hidden states, LayerNorm + Linear(1024,256) + GELU, attentive statistics pooling, 2-layer MLP to one AI logit. Per-window zero-mean / unit-std waveform normalisation inside the model. Authors' released checkpoint; training data undisclosed. Input: 16 kHz, first 64600 samples (Arena 4 s cap), shorter clips repeat-padded (the authors score 6 s windows with a 3 s hop over the whole clip and calibrate the top-30 % mean logit). score = -(pre-calibration AI logit), higher = human. FP32 (TF32 off).
- Weights: re-hosted here under the upstream license (model.safetensors, commit-pinned); upstream: https://huggingface.co/wasitaigeneratedcom/ai-voice-detector-small/blob/4858b4064b1f8530def904e6340d2780ad847160/detector.py
- Params: 315.87 M. Trained on undisclosed data, so no Arena cell is marked n/a*, but leakage cannot be ruled out.
- Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; most other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; FP32 on the RTX 5060 Ti), which matched PyTorch on the full J-SPAW_LA set (|dEER| <= 0.25, Spearman >= 0.999); the RTX 4060 Ti cards ran PyTorch.
Arena results
EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's
DeepVoice EER threshold. Per-dataset scores are under .eval_results/.
- Downloads last month
- 212