Check a voice clip

How a voice check runsCopy Anchor Link

A voice clip is a job, as a video is. POST /audit answers 202 at once with its id and media: "audio", and you poll GET /audit/{id} until its status is succeeded or failed. The statuses and the polling rhythm are a video's, as Check and optimize a video describes.

Send exactly one of audio_url or audio_upload_id, as an MP3, WAV or M4A. A clip on disk goes through an upload with its audio content_type. A voice clip is only checked. POST /optimize refuses one with unsupported_media.

Send the lengthCopy Anchor Link

Pass seconds, the clip's length, when you know it. Credits are held for that length, and only that much of a longer clip is read. Without it, credits for the longest clip are held, and what was read is charged once the job finishes. A voice clip takes the video limits and costs what a video check costs.

Check a voice clipCopy Anchor Link

curl -X POST "https://realpng.com/api/v1/audit" \
  -H "X-API-Key: $REALPNG_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "audio_url": "https://audio.example.com/voice-note.mp3",
    "seconds": 18
  }'

{
  "id": "9a4c2e7f-1b6d-4e8a-b3f5-0d7c9e2a6b18",
  "status": "queued",
  "media": "audio"
}

Then poll the check until it finishes.

curl "https://realpng.com/api/v1/audit/9a4c2e7f-1b6d-4e8a-b3f5-0d7c9e2a6b18" \
  -H "X-API-Key: $REALPNG_API_KEY"

{
  "id": "9a4c2e7f-1b6d-4e8a-b3f5-0d7c9e2a6b18",
  "status": "succeeded",
  "verdict": "ai",
  "score": 91,
  "decided_by": "voice",
  "lanes": {
    "voice": {
      "verdict": "ai",
      "score": 91
    },
    "provenance": {
      "verdict": "nothing",
      "issuer": null
    }
  },
  "media": "audio",
  "audio": {
    "seconds": 18,
    "scored_seconds": 18,
    "readings": [
      {
        "t": 0,
        "score": 89
      },
      {
        "t": 4.038,
        "score": 92
      },
      {
        "t": 8.075,
        "score": 91
      },
      {
        "t": 12.113,
        "score": 90
      }
    ]
  }
}

One voice detector reads the clip in 4 second windows, and the clip's reading is the median of the windows'. A score under 30 reads real, from 50 it reads ai, and between them uncertain. A valid C2PA claim decides first: a signed recording reads real and a signed AI claim reads ai, whatever the detector reads.

decided_by is voice when the detector decides and provenance when a signature does. lanes has voice, the detector's reading, and provenance, with no watermark lane. audio adds the clip's length and each window's reading.

What a voice check cannot tellCopy Anchor Link

• A clip played through a speaker and recorded again can read as a real voice.
• Edited real audio, such as a pitch shift or heavy noise, can read as AI.
• Voices from Google and Microsoft are not in our test sets yet.
• The detector was built for speech. Music and silence get a reading it was never tested on.