Scam AI raised $2.6M to combat AI-powered scams→
scam.ai

~/docs cat detect-audio.md

Detect audio

Checks a recording for synthetic or cloned voice, and answers with one verdict for the recording plus a timeline with a value for every window of speech.

Send the file

curl -X POST \
  "https://api.scam.ai/v1/detections" \
  -H "x-api-key: <YOUR_API_KEY>" \
  -F "file=@/path/to/voicemail.mp3"

Recordings are sent as a file. A link to a recording is fetched too; see Detect from a link.

Parameters

Request parameters
FieldTypeMeaning
filebinaryThe media file. MP3, WAV, FLAC, M4A, AAC or OGG, at most 50 MB.

The response

idstring
The detection id. Pass it to GET /v1/detections/{id} to read the run back.
statusstring
Always "completed". The answer arrives in the same request.
created_atstring
When the run finished. RFC 3339, millisecond precision, always Z.
mediaobject
What was checked: type (image, video or audio), filename, mime_type and bytes. The last three can be null when they were not known. mime_type is what the upload declared, so a client that sends no content type sees application/octet-stream here; type comes from the bytes themselves.
modelstring
The Eva version that answered this run, "Eva V1.5" or "Eva V1.6". Which version answers is set per account.
verdictstring
LIKELY_REAL, ALERT or LIKELY_AI. Branch on this.
scorenumber | null
0 to 1, the probability the media is AI-made or manipulated. null when nothing could be scored.
summarystring
One sentence a person can read, matching the verdict.
credits_usednumber
What this run cost, an integer. 0 on a sandbox key, or when the same file was sent byte for byte from this account within 30 days.
zero_charge_reasonstring
Only when credits_used is 0: "sandbox" or "duplicate".
duration_msnumber
The end of the last window in milliseconds, scored or not. Audio is billed by length.
segmentsarray
One entry per window: start_ms, end_ms and value (0 to 1, the probability that window is synthetic, or null when the window was not scored).
segments[].statusstring
scored when the window has a value, abstained when it does not.
segments[].reason_codesstring[]
Only on an abstained window: why it was not scored, for example no_voice_content.
response · 200
{  "id": "c3a9e1d0-7f24-4b6a-8e15-9d0c2b7a4f61",  "status": "completed",  "created_at": "2026-09-15T10:21:40.088Z",  "media": {    "type": "audio",    "filename": "voicemail.mp3",    "mime_type": "audio/mpeg",    "bytes": 131072  },  "model": "Eva V1.6",  "verdict": "LIKELY_AI",  "score": 0.96,  "summary": "This recording shows strong signs of synthetic or cloned voice.",  "credits_used": 1,  "duration_ms": 8000,  "segments": [    { "start_ms": 0, "end_ms": 4000, "value": 0.9612, "status": "scored" },    {      "start_ms": 4000,      "end_ms": 8000,      "value": null,      "status": "abstained",      "reason_codes": ["no_voice_content"]    }  ]}

Questions

What is a window?
The recording is scored in fixed-length windows. Each entry in segments[] is one window, with its start and end in milliseconds and its own value. The score at the top of the response is the verdict for the whole recording.
Why is a window's value null?
Nothing was scored there: status is abstained and reason_codes says why, for example no_voice_content when the window holds no speech. A null is never a 0.
Why does audio never answer ALERT?
A recording is judged on one line, so it answers LIKELY_REAL or LIKELY_AI. Your code should still handle all three words, because the same integration will see ALERT on images and videos.
← Detect a videoDetect from a link →

Start building

Create an API key in the dashboard and send your first file. To go through an integration or an on-prem deployment with our team, request a demo.

Get an API keyRequest a demo