AI Music Detection

Detect AI-generated music — vocals, instrumentals, and everything in between.

Most detectors return a single pass/fail for an entire track. Modulate AI Music Detect scores every 4-second segment, separately evaluating vocal and instrumental content — so you know where in a track AI was used, not just whether it was.

Get free API access

TALK TO SALES

How It Works

Clip-level verdicts built from per-window evidence.

Send audio, get back a primary verdict, per-window scores, and confidence values. Batch or real-time streaming — same structured output either way.

Batch API

Send a complete audio file, receive a clip-level primary_verdict plus a per-window breakdown. Supported formats: .aac, .flac, .m4a, .mp3, .mp4, .ogg, .opus, .wav — up to 100 MB.

Streaming API (WebSocket)

Connect over WebSocket and receive per-window vocal AI verdicts as audio arrives. Instrumental AI detection and the final clip-level verdict are delivered in the done message at end of stream.

Structured output

Every response includes primary_verdict, vocal_ai_percentage, vocal_ai_confidence, instrumental_ai_percentage, instrumental_ai_confidence, and a full per-window breakdown.

Tunable thresholds

Adjust the precision/recall tradeoff without retraining. Set vocal and instrumental thresholds independently to match your use case.

AI Detection Capabilities

Two detection paths. One API.

Most AI music detectors treat a track as a monolith — score it once, return a number, move on. That approach breaks down on hybrid tracks, multi-part productions, and anything where AI was used for only part of the composition.

Vocal detection

Identifies AI-generated singing, rap, or any vocal performance within a 4-second window — including AI vocals over organic human music, a common pattern in fraudulent or policy-violating uploads. Each window is scored independently, returning vocal_ai_percentage and vocal_ai_confidence.

Instrumental detection

Identifies AI-generated instrumental content when no sufficient vocal content is present — including fully AI-generated tracks with no vocals at all. Results are delivered per-window and aggregated into the clip-level verdict.

Modulate vs. industry standard

A side-by-side comparison for teams evaluating AI music detection.

Modulate

Industry Standard

Detection paths

Vocal + instrumental, independent

Single combined score per track

Output granularity

Per-4-second window + clip-level verdict

Clip-level only

Speech signals

Per-4s window for vocal AI; instrumental AI at end-of-stream

Single score at end of file

Hybrid content handling

Window-level visibility into mixed tracks

Not supported

Threshold tuning

98.9% accuracyYes — no retraining needed

Fixed or not offered

Streaming support

Real-time WebSocket streaming

Batch only

Pricing

$0.07/hr

$0.15–0.45/hr (estimated)

Additional Velma models

Transcription, Deepfake Detection, PII Redaction, Emotion, Accent

Music detection only

Self-serve API

Drop-inYes — sign up and start in minutes

Varies; some require enterprise partnership negotiation

Cost

$0.07/hr

Not publicly disclosed or per-track pricing

How It Works

Clip-level verdicts built from per-window evidence.

Send audio

Point Velma AI Music Detect at a file or a live stream. One endpoint — no pipeline to assemble. Supported formats: .aac, .flac, .m4a, .mp3, .mp4, .ogg, .opus, .wav.

Get per-window verdicts

Every 4-second window is scored independently for vocal AI and instrumental AI content. Streaming returns vocal results in real time; instrumental results and the clip-level verdict arrive at end of stream.

Receive structured output

Get back a primary_verdict of ai-vocal-music, ai-instrumental, or not-ai-music — plus confidence scores and a full per-window breakdown ready to plug into your pipeline.

Streaming, start to finish:

# 1 · open a connection 2 · stream audio 3 · read results

ws = connect("wss://platform.modulate.ai/api/velma-2-ai-music-detection-streaming?api_key=…")

ws.send(audio_chunk) # stream your audio

ws.send("") # signal end of stream

for event in ws: # window verdicts + final primary_verdict

print(event["primary_verdict"])