AI Audio Detector
Checks whether speech in an audio clip is AI-generated or voice-cloned.
| Slug | ID | Status | Tagged | Model version |
|---|---|---|---|---|
ai-audio-detection | 2 | Active | Yes | v2-MILE-Ensemble |
What it detects
| Detects | Does not detect |
|---|---|
| AI-generated or voice-cloned speech | Background-noise editing, simple splicing of real recordings, or non-speech AI-generated sounds |
The clip is analyzed in 4-second windows, and the overall score combines them.
Request
| Slot | Accepted contentType |
|---|---|
original | audio/mpeg, audio/wav, audio/m4a |
{
"taskTypeId": "2",
"inputs": [
{ "slotName": "original", "fileName": "call.mp3", "contentType": "audio/mpeg", "sizeBytes": 1048576 }
]
}See Running a job for the full create, upload and finalize flow.
Limits
- File size: up to the account upload limit, 10 MB by default.
Result
Uses the standard result fields, plus:
| Field | Type | Description |
|---|---|---|
windows | object | One entry per 4-second window, keyed by time range (e.g. "0-4s"). Each has RealConfidencePercent, the real confidence (0–100) for that window |
Each window entry, and the result itself, also has a spectrogram path. These paths are internal and can't be downloaded; use the image artifacts instead.
{
"isFake": false,
"confidencePercent": 54.18,
"fakeConfidencePercent": 45.82,
"realConfidencePercent": 54.18,
"tag": "likely_real",
"tagBoundaries": { "real": 30, "likelyReal": 50, "likelyFake": 75 },
"modelVersion": "v2-MILE-Ensemble",
"windows": {
"0-4s": { "RealConfidencePercent": 15.62, "spectrogram": "/app/jobs/560/output/spectrogram_window_0.png" },
"4-8s": { "RealConfidencePercent": 98.02, "spectrogram": "/app/jobs/560/output/spectrogram_window_1.png" },
"8-12s": { "RealConfidencePercent": 15.29, "spectrogram": "/app/jobs/560/output/spectrogram_window_2.png" }
},
"spectrogram": "/app/jobs/560/output/spectrogram.png"
}Low window scores show where in the clip the model found synthetic speech.
Output artifacts
kind | File | Description |
|---|---|---|
result | result.json | Full result with the raw probabilities |
image | spectrogram.png | Spectrogram of the whole clip with per-window confidence overlays |
image | spectrogram_window_N.png | One spectrogram per 4-second window (metadata.window gives the time range) |
Download them from artifacts on GET /jobs/{id}.
Cost and availability
- Fixed price per job, set by your plan. The exact amount is returned as
costwhen you create the job. See Quotas & Credits. - Available on the cloud API; your plan decides which task types you can use.
