Technical Overview
From Generated Audio to Attributable Recordings
Musical AI runs a four-stage pipeline that separates, fingerprints, and matches generated audio against your catalog to return ranked attribution evidence with rights-ready metadata.
Explore the APIThe Four Stages
Each Stage Adds a Layer of Attribution Precision
Source Separation
Isolate vocal, instrumental, rhythmic, and harmonic components using deep learning stem separation.
Spectral Analysis
Build per-stem frequency fingerprints that persist across tempo changes, pitch shifts, and EQ processing.
Catalog Matching
Run approximate-nearest-neighbor search across your indexed ISRC catalog at scale, with configurable confidence thresholds.
Attribution Response
Return ranked matches with confidence scores, ISRC identifiers, rights-holder metadata, and an audit trail.
Deep Dive
Stage 1: Source Separation
Modern AI music generation produces outputs that are blends of learned features from many source recordings. Separating those layers is the first requirement for accurate attribution.
Musical AI uses a proprietary source-separation model trained on annotated recorded-music datasets, producing up to eight discrete stems per submitted track. Each stem is independently fingerprinted, which means a strong rhythmic match can still surface even when the harmonic content was heavily transformed during generation.
The separation stage processes WAV and MP3 inputs up to 10 minutes at 44.1 kHz stereo.
Supported stem types
- Vocals (lead and backing)
- Lead melodic instrument
- Harmonic / chordal layers
- Bass frequency band
- Percussion and rhythm
- Ambient and texture stems
Catalog Integration
Index Your Catalog Once. Query It Continuously.
Musical AI ingests catalog data via a secure bulk-upload endpoint. ISRC-keyed metadata, rights-holder assignments, and audio references are indexed into a high-throughput vector store that enables millisecond-latency matching across catalogs of any size.
Bulk Catalog Ingest
Upload ISRC-keyed catalog segments via a single REST endpoint. Audio URLs, metadata JSON, and ownership trees are all accepted.
Incremental Updates
Add new releases, update rights-holder assignments, or remove catalog entries without re-indexing the full catalog.
Isolated Catalog Namespaces
Each organization's catalog is stored in a logically isolated namespace. Cross-client data access is architecturally prevented, not policy-controlled.
Confidence Scoring
Configurable Thresholds for Your Workflow
Every match returns a score from 0.0 to 1.0. You set the minimum threshold. High-confidence mode returns only matches above 0.75. Broad-sweep mode returns all matches above 0.20 for exploratory analysis. The audit trail logs every query and result for evidentiary documentation.
Ready to Build Your Attribution Infrastructure?
Request access to the Musical AI API and start indexing your catalog today.