tropa-voice: AI voice detection, hosted and open weights

What tropa-voice and tropa-voice-mini are

tropa-voice is the voice model behind the deepfake detector and the voice detection API (it appears as tropa-voice-1 in API responses). It scores the speech in a recording in short segments over the first 5 minutes and returns a calibrated AI probability from 0 to 100, a verdict (ai, uncertain or human) and a timeline of segment scores, so you can see where in a clip a synthetic voice occurs. A check of an MP3, WAV, OGG, AAC or FLAC file also asks OpenAI's content provenance service whether the clip carries OpenAI's audio watermark; OpenAI checks clips of up to 60 seconds.

tropa-voice-mini is a compact open-weights model (1.26 GB) on Hugging Face. It returns the same kind of output, a calibrated score, a verdict and per-window scores, and runs on a normal CPU, with 0.0% EER on ElevenLabs v3 and v4, MiniMax, OpenAI, Gemini and xAI voices. It is the strongest open-weights voice detector you can use commercially.

Results

Equal error rate (lower is better) on five public benchmarks, every model scored by us with the same code on the same files. Method and all five detectors are in the voice detector benchmark.

tropa-voice has a lower error rate than every open-source detector on all five benchmarks. On real 2024 social-media deepfakes it makes about 30% fewer errors than the best open detector, and on In-the-Wild it catches 99.4% of deepfakes at a 1% false-positive rate. On commercial voices from ElevenLabs, MiniMax, OpenAI, Gemini, xAI, HeyGen and others, both models make about ten times fewer errors than the next open detector. The evaluation code for every number is on GitHub.

Hosted API or open model?

The API gives you our most accurate system, continuously updated for new voice generators, with nothing to host. tropa-voice-mini is the right choice when audio must stay on your own servers or you want to start for free.

How to use tropa-voice (API)

Create an account, take the API key from the dashboard, and send a file or a URL. The response contains verdict, score (0–100), verified (OpenAI watermark confirmed) and detailed.segments with a score per segment. Each check costs 1,000 credits and is charged only when the analysis succeeds.

curl -X POST https://www.wasitaigenerated.com/api/v1/detect/audio \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -F "file=@suspicious_clip.mp3"

# Files over 4 MB (up to 100 MB): pass a URL instead
curl -X POST https://www.wasitaigenerated.com/api/v1/detect/audio \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url": "https://example.com/voice.mp3"}'

Field reference and code in Python, Node.js and PHP: API docs and the voice detection API page.

How to use tropa-voice-mini (Hugging Face)

# pip install torch transformers safetensors soundfile huggingface_hub  (+ ffmpeg for mp3/m4a/mp4)
from huggingface_hub import hf_hub_download
import importlib.util

repo = "wasitaigeneratedcom/ai-voice-detector-small"
spec = importlib.util.spec_from_file_location("detector", hf_hub_download(repo, "detector.py"))
detector = importlib.util.module_from_spec(spec); spec.loader.exec_module(detector)

det = detector.VoiceDetector.from_pretrained(repo)   # GPU if available, else CPU
result = det("suspicious_clip.mp3")
print(result["ai_score"], result["verdict"])         # e.g. 0.9986 'ai'
for seg in result["segments"]:                       # 6 s windows, 3 s hop
    print(seg["start"], seg["end"], seg["ai_score"])

ai_score is a calibrated probability from 0 to 1. The default ai threshold sits at a 0.5% false-positive rate on human speech; serving_head.json also lists thresholds for 1% and 0.1%. A 30-second clip takes about 0.06 s on a GPU and about 10 s on 4 CPU cores. Model card and files: wasitaigeneratedcom/ai-voice-detector-small.

FAQ