Detect API: find out if a voice recording is AI‑generated.Protect API: stop AI from transcribing a recording, while people can still understand it.
1,000 free credits to start. No card needed.
Send a file, get an answer. Audio is processed in memory and never stored.
POST/v1/detect
The model behind Voice Check, as an API. Send a call, a voice note or a clip and get a verdict with a confidence, plus a second-by-second timeline.
POST/v1/protect
The engine behind Off the Record, for files. You get back a recording people can still understand, while AI transcription gets little or nothing usable from it.
balanced, strong, maxFlag cloned voices in uploads, voice notes and livestream clips before they spread.
Check recorded calls for synthetic callers posing as customers or executives.
Check a viral clip before you run it, and see which seconds sound generated.
Share recordings people can listen to that AI notetakers can’t turn into text.
Credits work for both endpoints and never expire. 1 credit is 1 second of audio checked; protecting costs 3 credits a second.
To try it out
$0
1,000 credits, once
For new accounts
For side projects
$10once
50,000 credits
$0.012 a minute checked
For apps in production
$50once
300,000 credits
$0.010 a minute checked
For high volume
$200once
1,500,000 credits
$0.008 a minute checked
Volume pricing, higher rate limits, streaming and private deployments, and support from the team that builds it.
Detect: 1 credit / second · Protect: 3 credits / second · Billed per started second · Prices in USD; tax may apply at checkout
/v1/detect or /v1/protect.export WM_API_KEY="wm_live_…"
curl https://api.weirdmachines.ai/v1/detect \
-H "Authorization: Bearer $WM_API_KEY" \
-F file=@voice-note.m4aThe base URL is https://api.weirdmachines.ai/v1. Every response is JSON, except a successful /v1/protect, which is the audio itself.
Pass your key as a bearer token on every request. Keys start with wm_live_. Create, name and revoke them in the console; a key is shown once, when you create it.
Authorization: Bearer wm_live_…Keep keys on your server. Anyone with a key can spend its account’s credits.
/v1/detectTells whether the voice in a recording is a person or AI. The recording is scored in overlapping 3-second windows; the verdict weighs every window that has a voice in it.
Send the file as multipart/form-data in a field named file, or as the raw request body with any Content-Type.
curl https://api.weirdmachines.ai/v1/detect \
-H "Authorization: Bearer $WM_API_KEY" \
--data-binary @call.wav| Field | Description |
|---|---|
verdict | "ai", "human", or "no_voice" when no speech was found (music, noise, silence). |
confidence | 0–1: how strongly the recording leans toward the verdict. null for no_voice. |
scores.ai, scores.human | The AI and human shares of the voiced windows, adding up to 1. |
audio | duration and voice_seconds in seconds, plus the input’s sample_rate and channels. |
timeline[] | One entry per window, a second apart: start, end, voice (0–1, how likely there’s a voice) and ai (0–1, null when the window is silent). |
credits | Credits used by this request and remaining on the account. |
Detection can make mistakes. Use it as one signal among others, and don’t base decisions that matter on it alone.
/v1/protectReturns the recording protected from AI transcription. Sound quality drops somewhat, but people can still understand what’s said; transcription services get little or nothing usable from it.
| Parameter | Description |
|---|---|
file | The recording, as with /v1/detect. |
strength | Optional, as a query or form field: balanced (default), strong or max. Stronger levels cost more sound quality. |
curl "https://api.weirdmachines.ai/v1/protect?strength=strong" \
-H "Authorization: Bearer $WM_API_KEY" \
-F file=@interview.mp3 \
-o interview-protected.wavA 16-bit mono WAV (audio/wav) at the input’s sample rate (up to 48 kHz), the same length as the input. Details come in headers:
| Header | Description |
|---|---|
X-Request-Id | The request’s id, for support. |
X-Audio-Duration | Length in seconds. |
X-Audio-Sample-Rate | The WAV’s sample rate. |
X-Credits-Used, X-Credits-Remaining | Credits for this request, and what’s left. |
Protection makes recordings much harder to transcribe; it isn’t a guarantee against every system. Heavy re-encoding, like low-bitrate phone codecs, weakens it.
/v1/accountYour balance and rates.
{
"object": "account",
"credits": { "total": 50820, "free": 820, "paid": 50000, "paid_purchased": 50000, "free_granted": 1000 },
"rates": { "detect_per_second": 1, "protect_per_second": 3 }
}New accounts start with 1,000 free credits, once. Credits never expire, and free ones are used before bought ones.
Errors are JSON with a stable type and a message you can show to people.
{ "error": { "type": "insufficient_credits", "message": "This needs 94 credits and you have 12. Buy more in the console", "credits_needed": 94, "credits_remaining": 12 } }| Status | Type | Meaning |
|---|---|---|
| 400 | invalid_request | Missing file or a bad parameter. |
| 401 | unauthorized | Missing, wrong or revoked key. |
| 402 | insufficient_credits | Not enough credits; nothing was charged. |
| 413 | file_too_large, audio_too_long | Over 25 MB, or longer than the endpoint takes. |
| 415 | unsupported_audio | The file couldn’t be read as audio. |
| 422 | no_audio | The file has no audio in it. |
| 429 | rate_limited | Over 60 requests a minute for this key. |
| 5xx | internal_error | Our side. Retry with backoff; you weren’t charged. |
/v1/detect, 5 minutes for /v1/protect.Audio is decoded and processed in memory, then discarded. We never store it, listen to it, or train on it. For each request we keep only its kind, length, cost and, for detection, the verdict, so you can see your usage; that’s deleted after 90 days.
Sign in with your email, create a key, and make your first call in a minute.