VivaGuard · isAI API

Is this voice AI?

isAI detects whether a voice recording was AI-generated by ChatGPT, Claude, or Gemini. Send audio, get a verdict — AI detected, human review, or no AI detected — with a confidence level. On-device, zero egress, audio never stored.

Base URL

https://isai-501449824361.asia-south1.run.app

Content-Type: multipart/form-data

Authentication

Every /api/* call (except /api/health) requires your API key, sent as either header:

x-api-key: <your-key>
Authorization: Bearer <your-key>

Keys are per-customer and can be rotated or revoked. Keep them server-side — never ship a key in client code.

POST /api/isai

Classify one audio recording.

Request — multipart/form-data

field	required	notes
`file`	yes	the audio clip

Formats: mp3 · wav · m4a · mp4 · ogg · flac · aac · webm

Input limits

limit	value	behaviour
Min length	≥ 0.5 s	shorter → `400`. For a reliable verdict send ≥ 3–4 s of speech.
Max size	50 MB	larger → `413`. ≈ 20–50 min depending on bitrate.
Best range	under ~2 min	the whole clip is analysed, but the pass that detects AI spliced into part of a longer answer scans the first 120 s.

Audio is resampled to 16 kHz mono internally. A candidate's spoken answer (≈30 s–2 min) sits comfortably in the ideal range.

Response — 200 application/json

{
  "is_ai": false,
  "confidence": "high",
  "ai_score": 35,
  "advisory": "No AI-generated voice detected.",
  "explanation": "No AI-generated voice was detected in this recording.",
  "mixed": false,
  "filename": "answer.m4a"
}

field	type	meaning
`is_ai`	bool	`true` = an AI-generated voice (ChatGPT / Claude / Gemini) was detected
`confidence`	string	`high`/`medium`/`low` when `is_ai`; `review` for borderline; `high`/`low` on a clear result (confidence that no AI-generated voice was detected)
`ai_score`	int	0–100 AI-likelihood score — `≥70` AI detected · `50–69` human review · `<50` clear. Use it to rank results or set your own cutoff (see note below).
`advisory`	string	one-line customer-facing summary
`explanation`	string	short human-readable detail
`mixed`	bool	AI speech spliced into an otherwise human answer
`filename`	string	echo of the uploaded name

Verdict guide

what you see	meaning	suggested action
AI DETECTED `is_ai:true`	AI-generated voice (ChatGPT / Claude / Gemini)	treat the answer as AI-assisted
HUMAN REVIEW `confidence:"review"`	borderline, possible AI voice	route to a human reviewer
NO AI DETECTED clear	no AI-generated voice detected	none

A "no AI detected" result means no ChatGPT / Claude / Gemini-generated speech was detected — it is not positive proof the voice is human. AI from other tools, or a heavily disguised voice, may not be flagged. Use isAI as one corroborating signal, not the sole decision.

About ai_score

ai_score is a 0–100 indicator that always agrees with the verdict:

ai_score	verdict
70–100	AI DETECTED
50–69	HUMAN REVIEW
0–49	NO AI DETECTED

Use it to rank recordings or apply your own cutoff — e.g. escalate anything ≥ 60.

It is an ordinal indicator, not a calibrated probability. A score of 72 does not mean "72% likely AI", and small gaps (68 vs 72) are not meaningful — real-world recording conditions (microphone, room, phone codec) move the score more than that. Treat the bands as the signal and the exact number only for ranking.

Errors

status	when
`400`	no `file`, unsupported audio, or too short (< 0.5 s)
`401`	missing or invalid API key
`403`	license invalid/expired
`413`	file larger than 50 MB
`429`	rate limit exceeded — slow down and retry
`500`	internal error

Rate limit: requests per key per minute (default 120). Retry a 429 after a short backoff.

GET /api/health

No auth. Returns service status, version, and license validity.

{ "status":"healthy", "service":"isai", "version":"1.4.2",
  "auth_required":true, "egress":"none", "license":{ "valid":true } }

Example

curl -X POST \
  -H "x-api-key: $ISAI_KEY" \
  -F "file=@candidate-answer.m4a" \
  https://isai-501449824361.asia-south1.run.app/api/isai

{ "is_ai": true, "confidence": "high",
  "advisory": "Confirmed AI voice.",
  "explanation": "Voice characteristics match AI-generated speech.",
  "mixed": false, "filename": "candidate-answer.m4a" }

Privacy: audio is processed in memory and never written to disk. No egress — the recording does not leave the service.