Send an audio or video file and know within seconds whether the voice is AI-generated or cloned. tropa-voice beats every open-source detector on all five public benchmarks, shows where in the clip the synthetic voice occurs, and confirms OpenAI-generated audio by its watermark.
1,000 free credits · no credit card · MP3, WAV, M4A, OGG, FLAC, WebM, MP4, MOV · up to 100 MB
/api/v1/detect/audio# File up to 4MB, uploaded directly
curl -X POST https://www.wasitaigenerated.com/api/v1/detect/audio \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "file=@/path/to/voice-message.mp3"
# Larger file (up to 100MB), passed by URL
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/interview.mp4"}'file / url — file (multipart, up to 4MB) or url (JSON, up to 100MB)
Analyzed — the first 5 minutes of speech
Cost — 1,000 credits per file, only on success
{
"isAI": true,
"confidence": 0.981,
"score": 98.1,
"verdict": "ai",
"verified": false,
"openaiProvenance": null,
"patterns": ["synthetic_voice"],
"warnings": [],
"analysis": {
"likelihood": "high",
"reasoning": "AI-generated voice detected with
98.1% confidence. 2 of 4 segments flagged
as AI-generated."
},
"detailed": {
"scores": { "ai": 98.1, "human": 1.9 },
"model": "tropa-voice-1",
"durationSec": 24.0,
"segments": [
{ "startSec": 0, "endSec": 6, "aiScore": 0.02 },
{ "startSec": 6, "endSec": 12, "aiScore": 0.04 },
{ "startSec": 12, "endSec": 18, "aiScore": 0.97 },
{ "startSec": 18, "endSec": 24, "aiScore": 0.99 }
],
"segmentsFlaggedAi": 2,
"processingMs": 6120
},
"creditsCharged": 1000
}Built for the clips you actually receive: voice messages, call recordings, viral videos. Measured against the strongest open-source detectors with the same code on the same files.
Lower error rate than every open-source voice detector on all five public benchmarks we ran. On real 2024 social-media deepfakes it makes about 30% fewer errors than the best open alternative, and it catches 99.4% of In-the-Wild deepfakes at a 1% false-positive rate.
0.23% equal error rate across 14 commercial TTS systems, including ElevenLabs, MiniMax, OpenAI, Gemini and xAI, against 2.0–7.1% for open alternatives.
Clips are also checked against OpenAI’s audio provenance watermark. When OpenAI confirms it, the response says verified: true and the verdict is ai: a fact, not a probability.
Every segment gets its own score, so a cloned sentence spliced into a real recording shows up exactly where it happens.
tropa-voice is updated continuously as new TTS and voice-cloning systems appear. Your integration stays the same.
Audio or video up to 100 MB, the first 5 minutes analyzed, a calibrated score and a clear verdict. No GPU on your side, and zero retention for enterprise agreements.
| Benchmark | tropa-voice (this API) | tropa-voice-mini (open) | Best other open model |
|---|---|---|---|
| In-the-Wild | 0.66% | 0.98% | 0.96% |
| Deepfake-Eval-2024 | 11.9% | 13.5% | 17.2% |
| Commercial-Voices-2026-mini | 0.23% | 0.19% | 2.01% |
| SONAR | 0.05% | 0.18% | 0.43% |
| MLAAD-tiny | 0.34% | 2.33% | 0.71% |
Five detectors, five public benchmarks, same evaluation code and files. Full benchmark and method · Full results and reproducible benchmark code on GitHub
Want to run detection on your own servers? tropa-voice-mini is free under Apache-2.0. Hosted vs open compared
Multipart upload or a URL, Bearer auth, JSON back
# File up to 4MB, uploaded directly
curl -X POST https://www.wasitaigenerated.com/api/v1/detect/audio \
-H "Authorization: Bearer YOUR_API_KEY" \
-F "file=@/path/to/voice-message.mp3"
# Larger file (up to 100MB), passed by URL
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/interview.mp4"}'verdict — ai, uncertain or human. ai starts at a score of about 89, set for a 1% false-positive rate on real speech; human is 20 or below.score — calibrated AI probability for the whole clip, 0–100. confidence is the confidence in isAI (0–1).verified — true when OpenAI confirmed its own audio watermark in the clip.detailed.segments[] — startSec, endSec and aiScore (0–1) per segment: the timeline of where the synthetic voice occurs.warnings[] — reliability notes, e.g. for clips shorter than 2 seconds.Check voice messages and call recordings that ask for money or credentials, before anyone acts on a cloned voice of a relative, a CEO or a bank.
Verify leaked audio and viral clips of public figures before publication, with a segment timeline you can cite in the correction or the story. See the use case
Score uploaded audio and video as it arrives and route synthetic voices into the moderation queue you already run. See the use case
Screen recorded voice verifications and callback recordings for synthetic speech, alongside your existing identity checks. See the use case
Check podcasts, interviews and video soundtracks for cloned voices before licensing, airing or using them as evidence.
Every failure is a JSON body with an error field
1,000 credits per file, charged only on success. No subscription needed to start.
1,000
credits on email verification
1 voice file to try the endpoint. No card required.
from $10
one-time, never expires
50,000 credits ≈ 50 voice files. 1,500,000 credits ≈ 1,500 voice files.
$299/mo
3M API & MCP credits every month
Everything in Pro plus the API and MCP server, white-label PDF reports and priority support. Cancel anytime.
A JSON object with verdict (ai, uncertain or human), score (the calibrated AI probability for the whole clip, 0–100), isAI and confidence, verified (true when OpenAI confirmed its watermark), warnings, a short analysis, and detailed.segments with a start time, end time and AI score for each segment. creditsCharged tells you what the call cost.
tropa-voice beats every open-source detector we tested on all five public benchmarks, with equal error rates of 0.66% on In-the-Wild, 11.9% on Deepfake-Eval-2024, 0.23% on 14 commercial TTS systems, 0.05% on SONAR and 0.34% on MLAAD-tiny. On In-the-Wild it catches 99.4% of deepfakes at a 1% false-positive rate. On real 2024 social-media deepfakes it makes about 30% fewer errors than the best open alternative (17.2%).
For MP3, WAV, OGG, AAC and FLAC files, each check also asks OpenAI’s content provenance service whether the clip carries OpenAI’s audio watermark (OpenAI checks clips of up to 60 seconds). If it is confirmed, verified is true and the verdict is ai: the clip was made with an OpenAI voice. A missing watermark proves nothing, because other generators do not use it and re-encoding can remove it.
That is what the segment timeline is for. Each segment of the clip gets its own AI score, so a synthetic sentence spliced into an otherwise real recording shows up at the time it occurs. detailed.segmentsFlaggedAi counts the segments scored 0.5 or above.
MP3, WAV, M4A/AAC, OGG/Opus, FLAC and WebM audio, and MP4, MOV or WebM video, whose audio track is analyzed. Upload up to 4 MB as multipart/form-data, or pass a public URL for files up to 100 MB. The first 5 minutes are scored. Clips need at least a few seconds of speech; under 2 seconds the result comes with a warning.
1,000 credits per file, charged only when the analysis succeeds. A new verified account starts with 1,000 free credits, enough for one test file. The $10 pack of 50,000 credits covers 50 files, the $49 pack of 300,000 covers 300, and the $199 pack of 1,500,000 covers 1,500. API Business is $299/month with 3,000,000 credits every month (about 3,000 files).
tropa-voice-mini is free under Apache-2.0 and runs on your own hardware, which is ideal when audio must stay on your servers. The API runs tropa-voice, our most accurate system: it is continuously updated for new voice generators, checks OpenAI’s watermark, returns the full segment timeline and needs no infrastructure. The comparison is on the tropa-voice model page.
Audio sent through the API is processed to produce a result and is not used to train our models. For institutional and enterprise agreements we offer zero retention, which switches off detection history entirely. See the privacy policy for the full detail.
Create an account, verify your email, and your API key is waiting in the dashboard.
Volume pricing, SLAs, custom rate limits, zero retention and private deployments.