AIRealNet-Audio
by Modotte
A 95M-parameter detector that scores speech as AI-generated or human, 1.21% EER on ASVspoof5
Modotte/AIRealNet-AudioDeploy AIRealNet-Audio
Single-tenantMixpeek has no managed extractor for this model. On a single-tenant deployment you upload the weights and a custom plugin serves them next to the rest of your pipeline.
Overview
AIRealNet-Audio is a binary audio classifier from Modotte that labels speech as AI-generated (text-to-speech, voice cloning, voice conversion) or real human speech. It was released in September 2026 under the MIT license. It reads 16 kHz audio in 12-second chunks with a wav2vec-style encoder and applies L1 normalization twice, before and after a projection layer, so the classifier cannot use embedding length as a shortcut.\n\nThe card reports equal error rates of 1.21% on ASVspoof5, 3.14% on ASVspoof2019 DF and 3.64% on In-the-Wild, with clips under 3 seconds removed. Its default decision threshold is 0.5, and the card says to calibrate it on your own audio.
Architecture
Wav2vec-based audio encoder at 16 kHz producing 768-dimensional features pooled over time, a projection to 256 dimensions, and a classification head, with L1 unit-norm applied to the base features and again to the projected features. About 95M parameters. Trained on speech from more than 100 TTS and voice-cloning systems plus real speech, with codec, bitrate and noise augmentation.
Mixpeek SDK Integration
# AIRealNet-Audio runs on your side. Score each audio segment and store the
# score as metadata next to the segment, so a retriever can filter on it.
from transformers import pipeline
import requests
detect = pipeline("audio-classification", model="Modotte/AIRealNet-Audio", trust_remote_code=True)
for seg in segments: # e.g. {"key": "calls/2026-10-12/0412.wav", "start": 24.0, "path": "seg_0412_024.wav"}
scores = {r["label"]: r["score"] for r in detect(seg["path"])}
requests.post(
"https://api.mixpeek.com/v1/buckets/bkt_your_bucket/objects",
headers={"Authorization": "Bearer API_KEY", "X-Namespace": "ns_your_namespace"},
json={"blobs": [{"property": "audio", "type": "audio", "data": f"s3://your-bucket/{seg['key']}"}],
"metadata": {"start": seg["start"], "ai_voice_score": scores.get("AIVoice", 0.0)}},
)Capabilities
- Scores audio as AI-generated or real human speech
- Reads 12-second chunks at 16 kHz
- Adjustable decision threshold, 0.5 by default
- Runs through the Transformers audio-classification pipeline
- MIT license
Use Cases on Mixpeek
Benchmarks
| Dataset | Metric | Score | Source |
|---|---|---|---|
| ASVspoof5 | EER | 1.21% | Model card: Modotte/AIRealNet-Audio (self-reported; clips under 3 s removed) |
| Voxness | EER | 1.48% | Model card (self-reported) |
| Malaad | EER | 2.11% | Model card (self-reported) |
| ASVspoof2019 DF | EER | 3.14% | Model card (self-reported) |
| In-the-Wild | EER | 3.64% | Model card (self-reported) |
Performance
The card lists limits: very short clips, adversarial edits of real audio, unseen generators and heavy telephony distortion can lower accuracy, and it should not be the only evidence in forensic or legal settings. We have not measured it.
Common Pipeline Companions
Frequently Asked Questions
What is AIRealNet-Audio?
An open audio classifier that scores speech as AI-generated or real human speech. It reads 12-second chunks of 16 kHz audio and is MIT licensed.
How accurate is AIRealNet-Audio?
The card reports equal error rates from 1.21% on ASVspoof5 to 3.64% on In-the-Wild, with clips under 3 seconds excluded. Accuracy on your own audio depends on the voice generators and recording conditions you see, so test it there.
What threshold should I use to flag AI voices?
The default is 0.5, and the card says to tune it per use case. Score a labeled sample of your own real and synthetic audio and pick the cutoff where false flags are rare enough for your review team.
Can it detect a cloned voice in a phone call?
It can score call audio, but the card notes that heavy telephony distortion it did not see in training can reduce accuracy, and that it should not be the only evidence in high-stakes decisions.
Specification
Research Paper
AIRealNet-Audio model card
arxiv.orgBuild a pipeline with AIRealNet-Audio
Add this model to a processing pipeline alongside other extractors. Combine with retrieval stages for end-to-end search.
Run it on your own data, free