Skip to content

Try Resonance 2 early.

Tell us a little about yourself and what you’re building.

A brief description of your use case.

We’ll use these details to review your application and contact you about access. Privacy policy

Explore oruk

OrukZero · Preview

Synthetic speech detection API

OrukZero analyzes English recordings for patterns associated with human and AI-generated speech. Send an audio file and receive a prediction, a clip score and timed segment scores. Use it as one signal in an audio review workflow. API access is free with an active Oruk account; scores do not prove authenticity.

Contract checked . The public upload demo works without an account.

Inputs, outputs and price

OrukZero HTTP file API
EndpointPOST /v1/audio/synthetic-speech · model=orukzero
AudioEnglish speech; WAV, FLAC, OGG or MP3; mono or stereo at 8–48 kHz
Limits1–60 seconds; at most 12 MiB; one multipart file per request
Outputhuman/synthetic prediction, uncalibrated synthetic_score, threshold, duration and timed segments
AccessActive account, approved account API access and valid Bearer key. No paid subscription or separate OrukZero model approval required.
Price$0 extra, including free and trial accounts. No credits or plan minutes consumed.
Rate limit60 requests per minute per API key; capacity limits also apply
Unsupported inputM4A/AAC, remote audio URLs and silent audio

Check access first with GET /v1/models/orukzero/access and the same Bearer key. This reads eligibility without uploading audio. Other speech models have their own subscription allowances.

Run a Python request

Use the supplied public recording to check your integration, then pass your own file. The program sends one request and prints the JSON returned by the service. Keep the key on your server; the request ID helps investigation but does not guarantee idempotency.

Terminal · setup and public sample
python3 -m venv .venv
source .venv/bin/activate
python -m pip install requests
curl --fail --location --max-time 30 \
  https://oruk.ai/samples/oruk-quickstart.wav -o speech.wav
# Set ORUK_API_KEY in your server environment.
# Save the Python below as detect.py, then:
python detect.py speech.wav
Python · detect.py
import json
import os
from pathlib import Path
import sys
import uuid
import requests

key = os.environ.get("ORUK_API_KEY")
if not key:
    raise SystemExit("Set ORUK_API_KEY in the environment.")
if len(sys.argv) != 2:
    raise SystemExit("Usage: python detect.py speech.wav")
audio_path = Path(sys.argv[1])
if not audio_path.is_file() or not 0 < audio_path.stat().st_size <= 12 * 1024 * 1024:
    raise SystemExit("Provide a nonempty audio file of at most 12 MiB.")
request_id = str(uuid.uuid4())
print("Request ID:", request_id, file=sys.stderr)
try:
    with audio_path.open("rb") as audio:
        response = requests.post(
            "https://speech-api.oruk.ai/v1/audio/synthetic-speech",
            headers={"Authorization": "Bearer " + key, "X-Request-ID": request_id},
            data={"model": "orukzero"},
            files={"file": (audio_path.name, audio)},
            timeout=(10, 120),
        )
except (requests.RequestException, OSError):
    raise SystemExit("Request failed; no automatic retry. Keep the request ID for investigation.")
if not response.ok:
    print("HTTP", response.status_code, file=sys.stderr)
    if response.status_code in (429, 503):
        print("Retry-After:", response.headers.get("Retry-After", "not supplied"), file=sys.stderr)
    raise SystemExit("Check account access, input limits or service availability before retrying.")
try:
    print(json.dumps(response.json(), indent=2))
except ValueError:
    raise SystemExit("The response was not valid JSON. Keep the request ID.")

Read the returned fields

This abridged shape comes from the dedicated API reference. Its example score is illustrative, not an expected answer for the supplied recording. Your complete response also contains segment results and revision information.

JSON · documented example, abridged
{
  "model": "orukzero",
  "prediction": "human",
  "synthetic_score": 0.00156,
  "threshold": 0.5,
  "duration": 14.45,
  "usage": {
    "billed": false,
    "billing_mode": "included_with_account"
  }
}

The fixed classification threshold is 0.5. segments covers windows of up to six seconds; longer clips use experimental aggregation. Inspect disagreements across segments instead of relying only on the aggregate score. Preserve the returned model revision with any evaluation.

Handle failure without losing the context

  • 401: check the key, its expiry or revocation, and account API access. Purchasing a plan is not an OrukZero requirement.
  • 400, 413, 415 or 422: check model ID, multipart encoding, file size, duration, sample rate and supported format.
  • 429 or 503: retain the request ID, honor Retry-After where supplied and use bounded backoff. This example deliberately makes no automatic retry.
  • Timeout: completion is uncertain. A request ID is a trace identifier, not permission to replay indefinitely.

Evidence and limitations

A human prediction cannot certify that a recording is genuine, and a synthetic prediction does not establish fraud or identify a generator. Short, processed, noisy, mixed-origin and non-English recordings need separate evaluation. This detector does not provide transcription, speaker identity, music detection or a guarantee that it will recognize an unseen voice generator.

Evaluate false positives and false negatives on consented recordings from your intended workflow. Keep a held-out evaluation set, include recording and codec changes, and inspect segment-level failures. We do not advertise a universal accuracy figure for this endpoint. Uploaded audio is processed transiently and is not retained by the endpoint; see the privacy policy for operational data handling.

Sources: current input and access contract, machine-readable response schema and service health.