Proposal: Comprehensive Music Originality Verification & Legal Risk Mitigation#
The Problem#
Oshun generates music at scale — through Calliope (AI artists), Euterpe (music production), Isis (generative factory), and Nous (audio AI). Every generated track is a potential copyright liability. A single infringement lawsuit can cost $150,000+ per work in statutory damages (US Copyright Act §504), destroy brand trust, and trigger platform-wide takedowns.
Current AI music generators (Suno, Udio) are already facing RIAA lawsuits alleging systematic reproduction of copyrighted material. The legal landscape is hostile and precedent is being set right now. We cannot afford to be reactive — we must build the most comprehensive originality verification system in the industry, integrated into every music generation pipeline, operating as an immovable gate between generation and publication.
Why "As Good as Humanly Possible" Matters#
- Melodic plagiarism is the #1 litigation vector. The "Blurred Lines" verdict ($5.3M) established that even feel can be actionable. The "Dark Horse" case ($2.8M, later reversed) showed that short melodic fragments trigger lawsuits even when the legal outcome is uncertain. The litigation cost alone is destructive.
- AI memorization is a proven phenomenon. Large language models and music models demonstrably memorize and reproduce training data verbatim. We must assume our generation models can and will reproduce copyrighted material unless actively prevented.
- Scale amplifies risk. If we generate 10,000 tracks/day and 0.1% contain infringing material, that's 10 potential lawsuits per day. At scale, even tiny false-negative rates are catastrophic.
System Architecture Overview#
┌─────────────────────────────────────────────────────────────────────┐
│ MUSIC GENERATION PIPELINE │
│ (Calliope / Euterpe / Isis / Nous / any generation source) │
└──────────────────────────┬──────────────────────────────────────────┘
│
▼
┌─────────────────────────────────────────────────────────────────────┐
│ THEMIS MUSIC SHIELD GATE │
│ │
│ ┌─────────────┐ ┌─────────────┐ ┌─────────────┐ ┌──────────────┐ │
│ │ Melodic │ │ Harmonic │ │ Rhythmic │ │ Spectral │ │
│ │ Similarity │ │ Similarity │ │ Similarity │ │ Similarity │ │
│ │ Engine │ │ Engine │ │ Engine │ │ Engine │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬───────┘ │
│ │ │ │ │ │
│ ┌──────┴──────┐ ┌──────┴──────┐ ┌──────┴──────┐ ┌─────┴────────┐ │
│ │ Lyric │ │ Vocal │ │ Structural │ │ Audio │ │
│ │ Similarity │ │ Similarity │ │ Similarity │ │ Fingerprint │ │
│ │ Engine │ │ Engine │ │ Engine │ │ Engine │ │
│ └──────┬──────┘ └──────┬──────┘ └──────┬──────┘ └──────┬───────┘ │
│ │ │ │ │ │
│ └───────────┬───┴───────┬───────┴───────────────┘ │
│ │ │ │
│ ┌──────▼──────┐ ┌──▼──────────────┐ │
│ │ Risk │ │ Provenance │ │
│ │ Scoring │ │ Audit Trail │ │
│ │ Engine │ │ │ │
│ └──────┬──────┘ └──────┬──────────┘ │
│ │ │ │
│ ┌──────▼───────────────▼──────┐ │
│ │ Decision Engine │ │
│ │ PASS │ FLAG │ REJECT │ HOLD │ │
│ └──────┬──────────────────────┘ │
│ │ │
└─────────────────────┼───────────────────────────────────────────────┘
│
┌───────────┼───────────┐
▼ ▼ ▼
┌─────────┐ ┌─────────┐ ┌──────────┐
│ PASS │ │ FLAG │ │ REJECT │
│ Publish │ │ Human │ │ Auto- │
│ │ │ Review │ │ Regen │
└─────────┘ └─────────┘ └──────────┘
Core Technology Pillars#
1. Melodic Similarity Engine (@themis/melodic-analysis)#
Melody is the most legally protected element of music. This is the single most critical subsystem.
1.1 Pitch Sequence Extraction#
- Monophonic melody extraction — Isolate the dominant melodic line from polyphonic audio using pYIN (probabilistic YIN), CREPE, or SPICE neural pitch trackers. Extract fundamental frequency (f0) contours at frame-level resolution (10ms frames).
- Polyphonic melody extraction — For tracks with multiple melodic lines (harmonies, countermelodies), extract all salient pitch streams using multi-pitch estimation (DeepSalience, BasicPitch). Each stream is analyzed independently.
- MIDI-level quantization — Convert continuous pitch contours to discrete note events with onset, offset, pitch (MIDI number), and velocity. Apply adaptive quantization that respects the track's tempo grid while preserving expressive timing.
- Vocal melody isolation — Use source separation (Demucs, HTDemucs) to isolate vocal tracks before melody extraction, since vocal melodies are the primary litigation target.
- Instrumental melody extraction — Separately extract lead instrument melodies (guitar solos, synth leads, piano riffs) which are independently protectable.
1.2 Melodic Contour Analysis#
- Parsons Code representation — Encode melodies as Up/Down/Repeat sequences
(e.g.,
*UUDDUDDR). This captures the gross contour shape and enables rapid pre-filtering before expensive comparison. Parsons codes are transposition-invariant by definition. - Interval sequence representation — Encode melodies as sequences of
intervals in semitones (e.g.,
+2, +2, +1, -2, -2, -1, -2). This is transposition-invariant and more precise than Parsons coding. - Relative pitch encoding — Normalize all melodies to start on pitch 0 and express subsequent notes as cumulative intervals. Eliminates key dependency entirely.
- Contour complexity scoring — Measure the entropy/complexity of a melodic contour. Simple contours (stepwise, narrow range) are more likely to match existing works by chance; complex contours are stronger evidence of copying when matched.
- Segment-level contour comparison — Compare contours at multiple granularities: phrase-level (8-16 bars), motif-level (1-4 bars), and fragment-level (3-8 notes). Different match lengths have different legal significance.
1.3 N-gram Melodic Matching#
- Pitch n-gram extraction — Extract all contiguous pitch subsequences of length n (typically n = 5 through 12) from both the query melody and the reference database. Store as hash tables for O(1) lookup.
- Interval n-gram matching — Same as above but using interval sequences rather than absolute pitches. This is transposition-invariant — the same melody in C major and E♭ major produces identical interval n-grams.
- Weighted n-gram scoring — Not all n-gram matches are equal:
- Matches in the chorus/hook carry 3x weight (hooks are the most protectable and recognizable element)
- Matches at phrase boundaries carry 2x weight (beginnings and endings of melodic phrases are more memorable)
- Matches in verses carry 1x weight
- Matches in transitions/fills carry 0.5x weight
- Common melody filtering — Maintain a database of melodic fragments that are so common they are considered public domain or unprotectable (e.g., do-re-mi, pentatonic scale runs, standard blues licks, Alberti bass patterns). Filter these from match results to reduce false positives.
- Rhythmically-aware n-grams — Encode both pitch AND duration in n-grams
(e.g.,
C4:q, D4:e, E4:e= C quarter note, D eighth note, E eighth note). A melody with the same pitches but entirely different rhythm is a much weaker match.
1.4 Similarity Metrics#
- Edit distance (Levenshtein) — Minimum number of insertions, deletions, and substitutions to transform one melody into another. Normalized by length for cross-length comparison.
- Dynamic Time Warping (DTW) — Elastic matching that handles tempo variations, rubato, and slight timing differences. Essential for comparing melodies performed at different tempos or with different rhythmic feels.
- Longest Common Subsequence (LCS) — Finds the longest shared melodic fragment allowing gaps. A long LCS between two melodies is strong evidence of similarity even if surrounding material differs.
- Smith-Waterman local alignment — Adapted from bioinformatics. Finds the best local alignment between two melodies, identifying the region of highest similarity. Especially useful when one melody is embedded within a longer work.
- Cosine similarity on pitch histograms — Compare the statistical distribution of pitches (pitch class profiles). Two melodies using the same notes in the same proportions score high even if the ordering differs. Useful as a fast pre-filter.
- Earth Mover's Distance on contour features — Measure the "work" required to transform one melodic contour distribution into another. Captures global shape similarity.
1.5 Transposition & Tempo Invariance#
- All-key comparison — For every query melody, generate all 12 transpositions and compare against the reference database. Since interval-based representations are inherently transposition-invariant, this is handled implicitly for interval n-grams but must be explicit for absolute pitch comparisons.
- Tempo normalization — Normalize all melodies to a canonical tempo (120 BPM) before comparison. This ensures that the same melody played at 80 BPM and 140 BPM produces the same comparison result.
- Mode-invariant comparison — The same melody in major and natural minor (e.g., "Happy Birthday" in C major vs. A minor) should be detected. Use interval sequences modulo mode to capture this.
- Rhythmic augmentation/diminution detection — Detect when a melody has been rhythmically doubled (augmented) or halved (diminished) in duration while preserving pitch content.
1.6 Reference Database Integration#
- Proprietary melody database — Index millions of melodies extracted from commercially released music. Sources: MusicBrainz metadata, AcoustID fingerprints, licensed melody databases (e.g., MIDI databases, sheet music repositories).
- Public domain melody corpus — Separately index all known public domain melodies (folk songs, hymns, classical themes, traditional music from all cultures) so that matches against public domain works can be automatically cleared.
- Real-time ingestion — As new commercially released music appears (via chart tracking, release monitoring, Content ID partner feeds), automatically extract and index its melodies within 24 hours of release.
- Cultural melody databases — Index melodies from non-Western traditions (ragas, maqamat, gamelan scales, West African melodic patterns, Chinese pentatonic traditions) to avoid inadvertent appropriation claims.
2. Harmonic Similarity Engine (@themis/harmonic-analysis)#
Chord progressions have weaker copyright protection than melodies (the I-V-vi-IV progression is used in thousands of songs and is not protectable), but unusual harmonic sequences combined with other similar elements can strengthen an infringement claim.
2.1 Chord Progression Extraction#
- Audio-to-chord transcription — Extract chord labels from audio using neural chord recognition models (e.g., BTC, HarmonyTransformer). Output chord symbols at beat or half-beat resolution.
- Roman numeral analysis — Convert chord progressions to Roman numeral notation relative to the detected key (e.g., C-Am-F-G → I-vi-IV-V). This enables key-invariant comparison.
- Extended chord vocabulary — Recognize not just triads (major/minor) but 7ths, 9ths, 11ths, 13ths, sus chords, augmented, diminished, altered dominants, polychords, and slash chords. The specificity of unusual chord voicings increases protectability.
- Harmonic rhythm extraction — Capture not just which chords appear but when they change. The rhythmic placement of chord changes is part of the musical identity.
2.2 Harmonic Similarity Comparison#
- N-gram chord progression matching — Extract chord progression n-grams (n = 4 through 8) and compare against the reference database.
- Common progression filtering — Maintain an exhaustive database of
non-protectable progressions:
- I-V-vi-IV (most common pop progression)
- I-IV-V-I (basic tonal cadence)
- i-VII-VI-VII (Andalusian cadence)
- I-vi-IV-V (50s progression / "Heart and Soul")
- vi-IV-I-V (modern pop variant)
- 12-bar blues (I-I-I-I-IV-IV-I-I-V-IV-I-V)
- ii-V-I (jazz standard cadence)
- Hundreds more cataloged common progressions
- Any progression appearing in 50+ commercially released songs is automatically classified as non-protectable
- Harmonic function comparison — Compare progressions at the level of harmonic function (tonic/subdominant/dominant) rather than specific chords. This captures the deeper harmonic logic.
- Modulation pattern comparison — Detect and compare key changes (modulations). The specific pattern of modulations through a song is a more distinctive fingerprint than any single progression within a key.
- Voice leading comparison — Compare the specific voice leading (how individual notes move between chords) rather than just chord labels. Two songs using I-V-vi-IV with identical voice leading are more similar than two songs using the same progression with different voicings.
2.3 Harmonic Protectability Scoring#
- Uniqueness scoring — Score each chord progression against the reference database. Progressions found in fewer existing songs receive higher protectability scores.
- Complexity weighting — More complex harmonic language (jazz harmony, chromaticism, non-functional harmony) is more likely to be protectable because it's less likely to arise independently.
- Context-dependent scoring — A I-V-vi-IV progression alone is not protectable, but I-V-vi-IV with a specific bass line, specific voicing, and specific rhythm might be.
3. Rhythmic Similarity Engine (@themis/rhythmic-analysis)#
Rhythmic patterns have the weakest copyright protection of all musical elements, but distinctive rhythmic signatures (drum breaks, beat patterns) can be protectable, especially when sampled.
3.1 Beat & Rhythm Extraction#
- Beat grid detection — Extract the metrical grid (downbeats, beats, subdivisions) using joint beat/downbeat tracking models (madmom, BeatNet).
- Drum pattern transcription — Extract drum patterns from audio using automatic drum transcription. Decompose into kick, snare, hi-hat, and other percussion elements on a quantized grid.
- Rhythmic motif extraction — Identify repeated rhythmic patterns (1-4 bar loops) and extract them as binary onset patterns on a 16th-note or 32nd-note grid.
- Groove analysis — Extract micro-timing deviations from the grid (swing, shuffle, humanization patterns). These "groove templates" are distinct fingerprints.
- Polyrhythmic detection — Identify layered rhythmic patterns (e.g., 3-against-4) that may constitute protectable arrangements.
3.2 Rhythmic Comparison#
- Onset pattern matching — Compare binary onset patterns using Hamming distance (number of differing positions).
- Groove template similarity — Compare micro-timing profiles using correlation coefficients. Two tracks using the same drum machine pattern with the same swing setting will have highly correlated groove templates.
- Drum break fingerprinting — Maintain a database of iconic drum breaks (Amen break, Funky Drummer, Impeach the President, Think break, Apache, etc.) and detect their use or close derivatives.
- Tempo-invariant rhythm comparison — Normalize all rhythmic patterns to a canonical tempo before comparison.
- Cross-rhythm detection — Detect when a rhythmic pattern from one genre tradition appears in another (e.g., a dembow rhythm in a pop song).
4. Spectral & Timbral Similarity Engine (@themis/spectral-analysis)#
Timbre and spectral characteristics are generally not copyrightable, but they are crucial for detecting direct sampling, close reproduction of production styles, and "sound-alike" issues that may not constitute legal infringement but create brand/reputation risk.
4.1 Spectral Feature Extraction#
- Mel-Frequency Cepstral Coefficients (MFCCs) — Extract the standard 13-40 MFCCs per frame, capturing the spectral envelope shape. MFCCs are the most widely used feature for audio similarity.
- Chroma features (HPCP) — Extract Harmonic Pitch Class Profiles. These capture the harmonic content while being timbre-invariant.
- Spectral centroid, bandwidth, rolloff — Extract low-level spectral descriptors that characterize the "brightness" and frequency distribution.
- Spectral flux — Measure spectral change over time. Useful for detecting similar production dynamics (e.g., filter sweeps, build-ups).
- Tonnetz representation — Project harmonic content onto the Tonnetz (pitch-class network) for topology-aware harmonic comparison.
- Mel spectrogram embeddings — Use pre-trained neural networks (VGGish, PANNs, CLAP, MERT, MusicGen encoder) to extract high-level spectral embeddings that capture perceptual similarity.
4.2 Timbral Fingerprinting#
- Production style fingerprinting — Extract features that characterize the mix: EQ curves, dynamic range, stereo width, reverb characteristics, bass weight. Two tracks with the same production style will have similar fingerprints.
- Instrument timbre identification — Identify specific instrument sounds and their processing chain. Useful for detecting when the same synthesizer patches, sample libraries, or processing chains produce overly similar results.
- Vocal timbre analysis — See Vocal Similarity Engine (Section 6).
- Mastering characteristics — Compare loudness (LUFS), dynamic range, peak characteristics, and frequency balance.
4.3 Sample Detection#
- Direct sample detection — Detect when an audio segment has been directly sampled (copied) from another recording. This is distinct from musical similarity — it's identical audio content, potentially time-stretched or pitch-shifted.
- Sample library identification — Maintain fingerprints of popular sample libraries (Splice, Loopmasters, Native Instruments, etc.) and detect when generated audio reproduces exact samples. This prevents inadvertent claims against sample library content.
- Micro-sample detection — Detect very short samples (< 1 second) such as drum hits, vocal chops, and sound effects that may have been memorized from training data.
5. Lyric Similarity Engine (@themis/lyric-analysis)#
Lyrics have strong copyright protection as literary works. Even partial lyric reproduction can constitute infringement.
5.1 Text-Level Analysis#
- N-gram text matching — Extract word-level n-grams (n = 3 through 8) from generated lyrics and compare against a database of copyrighted lyrics. Direct text overlap beyond common phrases is flagged.
- Semantic similarity — Use sentence embeddings (all-MiniLM, BGE, GTE) to compare the meaning of lyric lines. Two lines expressing the same specific idea in different words may still infringe if the idea is sufficiently specific and original.
- Phonetic similarity (sounds-alike detection) — Convert lyrics to phonetic representations (IPA, Soundex, Metaphone) and compare. Lyrics that sound the same when sung but use different spellings are still potentially infringing. Example: "all the single ladies" vs. "all the simple ladies" sound nearly identical when sung.
- Rhyme scheme comparison — Extract rhyme schemes (ABAB, AABB, etc.) and compare. Identical rhyme schemes with similar content are a stronger match.
- Meter and scansion matching — Analyze the rhythmic meter of lyrics (stressed/unstressed syllable patterns) and compare. Lyrics with identical meter, rhyme scheme, and similar content are a strong match.
5.2 Filtering & Thresholds#
- Common phrase database — Maintain an exhaustive database of lyric phrases
that are too common to be protectable:
- Generic love expressions ("I love you", "baby come back", "hold me tight")
- Common idioms and proverbs used in songs
- Genre-specific clichés ("ride or die", "on the floor", "hands in the air")
- Phrases appearing in 100+ commercially released songs
- Title protection — Song titles are generally not copyrightable, but using a distinctive title phrase in lyrics (e.g., "Bohemian Rhapsody", "Smells Like Teen Spirit") triggers elevated scrutiny.
- Hook/chorus weighting — Lyric matches in the chorus/hook carry higher weight than matches in verses, as choruses are the most memorable and recognizable portions.
- Multi-language support — Analyze lyrics in all target languages. Detect when lyrics translate or transliterate copyrighted lyrics from another language.
5.3 Lyric-Melody Joint Analysis#
- Combined lyric-melody matching — The strongest infringement evidence is when both the lyrics AND the melody match simultaneously. A two-dimensional match (same words sung to the same tune) is far more significant than either alone.
- Syllable-pitch alignment — Analyze how specific syllables align with specific pitches. Even with different lyrics, if the syllable stress patterns align with the same melodic contour, this is a compound similarity signal.
6. Vocal Similarity Engine (@themis/vocal-analysis)#
Voice similarity is increasingly litigious. The NO FAKES Act (US), and emerging international regulations specifically address unauthorized voice replication.
6.1 Speaker Embedding Comparison#
- Speaker verification embeddings — Extract d-vectors or x-vectors using pre-trained speaker verification models (ECAPA-TDNN, TitaNet, WavLM). These embeddings capture the unique characteristics of a voice independent of what is being said.
- Cosine similarity threshold — Compare speaker embeddings against a database of protected voices (real artists who have not consented to voice replication). Flag any generated voice with cosine similarity > 0.75 to any protected voice.
- Voice print database — Maintain a database of voice prints for:
- All major-label recording artists (extracted from publicly available music)
- Artists who have explicitly opted out of AI voice replication
- Artists covered by the NO FAKES Act or equivalent legislation
- Public figures with voice right protections
6.2 Vocal Characteristic Analysis#
- Formant analysis — Compare vocal formant frequencies (F1-F5) which characterize the unique resonance of each voice.
- Vibrato analysis — Compare vibrato rate, depth, and onset patterns. Distinctive vibrato is a key identifying characteristic.
- Breathiness/rasp quantification — Measure and compare voice quality parameters (breathiness, roughness, strain).
- Singing technique fingerprinting — Identify signature techniques: melismatic runs (Mariah Carey, Whitney Houston), vocal fry (Britney Spears), growling (metal vocalists), yodeling (various traditions), falsetto transitions, etc.
- Phrasing and timing analysis — Compare how a voice phrases lyrics (anticipated/delayed beats, rubato, breath placement).
6.3 Anti-Impersonation System#
- Real-time voice comparison — During voice synthesis, continuously compare the generated voice against the protected voice database.
- Distinctiveness scoring — Score each generated voice on how distinct it is from all known voices, not just protected ones. Higher distinctiveness = lower risk.
- Voice modification suggestions — When a generated voice is too similar to a protected voice, suggest specific modifications (shift formants, alter vibrato, change breathiness) to increase distinctiveness while preserving quality.
- Consent verification — Integrate with a consent database: if an artist has explicitly consented to voice replication (e.g., for a licensed collaboration), bypass the similarity flag for that specific artist.
7. Structural Similarity Engine (@themis/structural-analysis)#
Song structure alone is rarely protectable (verse-chorus-verse is universal), but structural similarity combined with other similar elements strengthens infringement claims.
7.1 Song Form Analysis#
- Section segmentation — Automatically segment songs into structural sections (intro, verse, pre-chorus, chorus, bridge, outro, instrumental break) using MSAF, SALAMI, or neural segmentation models.
- Section label sequence — Extract the ordered sequence of section types as a structural fingerprint (e.g., I-V-V-PC-C-V-PC-C-B-C-C-O).
- Section duration ratios — Compare the relative durations of each section. Two songs with identical structure AND identical section proportions are more similar than two songs with just the same section sequence.
- Energy contour — Extract the loudness/energy curve over the entire song. Compare the overall dynamic shape (build-ups, drops, quiet sections).
7.2 Arrangement Analysis#
- Instrumentation timeline — Track which instruments enter and exit throughout the song. The arrangement fingerprint captures the orchestration choices.
- Textural density curve — Measure the number of simultaneous sound sources over time. Compare the textural "shape" of the arrangement.
- Production event detection — Detect specific production events (filter sweeps, risers, drops, breakdowns, drum fills) and compare their placement and sequence.
8. Audio Fingerprint Engine (@themis/audio-fingerprint)#
Extends the existing @nous/audio fingerprinting system with copyright-specific
capabilities. This is the fastest pre-filter — catching direct copies and very
close reproductions before the more expensive analysis engines run.
8.1 Multi-Algorithm Fingerprinting#
- Landmark-based fingerprinting — Shazam-style spectrogram peak
constellation matching. Highly robust to noise, encoding artifacts, and
time-shifting. Catches direct copies, covers, and close reproductions.
(Extends existing
@nous/audiolandmark fingerprinting.) - Chroma-based fingerprinting — Captures harmonic content independent of
timbre. Catches transposed covers, re-orchestrations, and instrumental
versions. (Extends existing
@nous/audiochroma fingerprinting.) - Neural audio fingerprinting — Use learned fingerprints from pre-trained models (Neural Audio Fingerprint, COLA) that capture perceptual similarity at a deeper level than hand-crafted features.
- Robust hashing — Generate perceptual hashes that are invariant to common transformations: MP3 encoding, volume changes, EQ adjustments, slight pitch shifts. Two versions of the "same" recording should produce the same hash.
8.2 Content ID Integration#
- YouTube Content ID proxy — Before publication, submit audio to YouTube as private/unlisted and monitor for Content ID claims. This leverages the industry's most comprehensive fingerprint database (100M+ reference tracks) at zero cost.
- Audible Magic / DistroKid integration — Pre-screen against additional fingerprint databases used by major distribution platforms.
- ISRC/ISWC cross-reference — When a fingerprint match is found, resolve to the specific copyrighted work via International Standard Recording Code (ISRC) or International Standard Musical Work Code (ISWC).
8.3 Near-Duplicate Detection#
- Segment-level matching — Don't just match whole tracks; match segments (e.g., 10-second windows). A generated track that inadvertently reproduces 15 seconds of a copyrighted chorus should be caught even if the rest is original.
- Speed/pitch-shifted matching — Detect copies that have been time-stretched or pitch-shifted. Common AI model failure mode: reproducing training data at a different speed/pitch.
- Mix element matching — After source separation (stems), fingerprint individual stems (vocals, drums, bass, instruments) and match each independently. A track might have an original melody but inadvertently reproduce a copyrighted drum pattern.
9. Reference Database System (@themis/reference-db)#
The entire system is only as good as its reference database. A comprehensive, current, and well-organized reference database is the foundation.
9.1 Database Architecture#
- Melody database — Indexed melodic contours, pitch n-grams, and interval sequences for millions of songs. Optimized for fast n-gram lookup (hash tables) and approximate nearest neighbor search (HNSW, IVF).
- Audio fingerprint database — Landmark and chroma fingerprints for all indexed audio. Optimized for constellation matching.
- Lyric database — Full lyrics with word-level n-gram indices and sentence embeddings. Sources: licensed lyric APIs (Musixmatch, Genius), public domain collections.
- Voice print database — Speaker embeddings for protected voices.
- Harmonic database — Chord progressions in Roman numeral form, indexed by n-gram.
- Rhythm database — Quantized drum patterns and rhythmic motifs.
- Neural embedding database — High-dimensional audio embeddings in a vector database (pgvector, Qdrant) for fast approximate similarity search.
9.2 Data Sources & Ingestion#
- Commercial music catalogs — License or extract features from:
- MusicBrainz / AcoustID (open metadata + fingerprints for 20M+ recordings)
- Spotify audio features API (tempo, key, loudness, etc. for 100M+ tracks)
- Billboard / chart tracking services (prioritize popular/high-risk works)
- MIDI databases (Lakh MIDI Dataset: 176,000+ MIDI files with melody data)
- Public domain collections — Index all known public domain music to enable
automatic clearance of public domain matches:
- IMSLP (600,000+ classical scores)
- Musopen (public domain recordings)
- Traditional folk music archives
- Pre-1928 published music (US public domain)
- Real-time ingestion pipeline — Monitor new releases via:
- Spotify new release radar
- Apple Music new releases
- Billboard chart entries
- DDEX feed from major labels (if partnership established)
- Social media viral music detection (TikTok trending sounds)
- Update cadence — New commercially released music must be indexed within 48 hours of release. Chart-topping music within 24 hours. Viral content within 12 hours.
9.3 Database Maintenance#
- Deduplication — Prevent the same work from appearing multiple times under different metadata (remasters, re-releases, compilation appearances).
- Metadata enrichment — Link fingerprint entries to comprehensive metadata: songwriter credits, publisher, label, ISRC, ISWC, copyright holder, known licensing terms.
- Protectability annotation — For each indexed work, score which elements are likely protectable based on originality analysis. Common progressions, generic lyrics, and standard drum patterns are marked as low-protectability.
- Jurisdiction mapping — Track copyright status per jurisdiction. A work in the public domain in one country may still be protected in another.
10. Legal Risk Scoring Engine (@themis/risk-scoring)#
The decision engine that synthesizes all similarity signals into a single actionable risk assessment.
10.1 Multi-Dimensional Similarity Score#
-
Element-wise scoring — Generate independent similarity scores for each musical element: | Element | Weight | Rationale | |---|---|---| | Melody | 0.35 | Strongest copyright protection | | Lyrics | 0.25 | Strong literary copyright | | Vocal identity | 0.15 | NO FAKES Act, voice rights | | Harmony | 0.08 | Weak protection unless unusual | | Production/timbre | 0.07 | Generally not protectable | | Structure | 0.05 | Almost never protectable alone | | Rhythm | 0.05 | Weakest protection |
-
Compound similarity detection — When multiple elements simultaneously match the same reference work, the risk escalates non-linearly. Two independent low-similarity matches (melody 30% + harmony 40%) are far less concerning than a compound match (melody 60% to Song X AND lyrics 50% to Song X). Compound matches against the same work receive a multiplier.
-
Confidence intervals — All scores include confidence intervals. A 70% similarity score with ±20% confidence is treated differently than 70% with ±3% confidence.
10.2 Legal Precedent Database#
- Case law database — Maintain a structured database of music copyright infringement cases with their outcomes, the specific elements found similar, and the threshold the court applied: | Case | Year | Outcome | Key Element | Threshold Applied | |---|---|---|---|---| | Bright Tunes v. Harrisongs ("My Sweet Lord") | 1976 | Infringement | Melody | Subconscious copying | | Three Boys Music v. Bolton ("Love Is a Wonderful Thing") | 2000 | Infringement ($5.4M) | Melody + lyrics | General audience test | | Williams v. Gaye ("Blurred Lines") | 2015 | Infringement ($5.3M) | Overall feel/groove | Extrinsic + intrinsic test | | Gray v. Perry ("Dark Horse") | 2019 | Reversed on appeal | Ostinato | Too commonplace | | Griffin v. Sheeran ("Thinking Out Loud") | 2023 | No infringement | Harmony + rhythm | Common elements | | Structured Asset Sales v. Sheeran ("Let's Get It On") | 2023 | No infringement | Harmony + rhythm | Chord progressions not protectable |
- Threshold calibration — Use case law outcomes to calibrate risk thresholds. The system should flag at levels below where courts have found infringement, providing a safety margin.
10.3 Risk Classification#
- PASS (Green) — Similarity scores below the minimum actionable threshold across all elements. No matches to any reference work exceed background noise levels. Auto-approved for publication.
- ADVISORY (Yellow-Green) — Low similarity detected to one or more reference works, but well below actionable thresholds. Logged for audit trail but auto-approved. Notification sent to content team for awareness.
- FLAG (Yellow) — Moderate similarity detected. Requires human review before publication. Content is held in review queue.
- HIGH RISK (Orange) — Significant similarity to a specific copyrighted work across one or more elements. Requires musicologist review AND legal review before publication. Remediation recommended.
- REJECT (Red) — Similarity exceeds safe thresholds with high confidence. Automatic rejection. Content is blocked from publication and routed to auto-regeneration pipeline.
10.4 Jurisdiction-Aware Assessment#
- US Copyright Act analysis — Apply the Arnstein test (probative similarity → access → substantial similarity) and the Ninth Circuit's extrinsic/intrinsic test.
- EU Copyright Directive — Apply the originality threshold standard (author's own intellectual creation).
- UK CDPA analysis — Apply the "substantial part" test.
- Multi-jurisdiction reporting — When a work will be distributed globally, assess risk under all major jurisdictions and report the highest risk.
- Moral rights considerations — In jurisdictions with moral rights (most of the world outside the US), assess whether generated content could constitute a "distortion" of an existing work.
11. Provenance & Audit Trail System (@themis/provenance)#
Comprehensive documentation of every generated work's origins. Essential for legal defense ("we took reasonable steps") and regulatory compliance.
11.1 Generation Provenance#
- Prompt capture — Record the exact prompt/parameters used to generate every piece of music. This includes text prompts, reference audio, style parameters, and any other generation inputs.
- Model versioning — Record the exact model version, checkpoint, and configuration used for generation. If a model is later found to have memorization issues, this enables retroactive auditing.
- Generation parameters — Record temperature, top-k/top-p, guidance scale, seed, and any other generation parameters.
- Timestamp and operator — Record when each piece was generated and by whom (automated pipeline vs. human operator vs. specific AI artist agent).
- Multi-stage provenance — For content that goes through multiple generation stages (e.g., melody → arrangement → production → mastering), record provenance at each stage.
11.2 Training Data Provenance#
- Model training data manifests — For each model used in generation, maintain a manifest of its training data: what datasets were used, what licenses apply, what opt-outs were honored.
- Dataset audit — Regular audits of training datasets to verify license compliance, opt-out honoring, and data quality.
- Opt-out registry — Maintain a registry of artists/rights holders who have opted out of AI training. Verify that all generation models respect these opt-outs.
- Data lineage graph — Track the full lineage from training data → model → generated output. If a specific training data source is later found to be unlicensed, identify all generated outputs that may be affected.
11.3 Decision Audit Trail#
- Similarity scan results — Store the complete output of every similarity scan, including all element-wise scores, matched reference works, and the final risk classification.
- Human review decisions — Record all human review decisions with reviewer identity, timestamp, rationale, and any conditions attached.
- Override logging — If a human reviewer overrides an automated rejection, record the full justification. All overrides are auditable.
- Publication decision chain — For every published work, maintain the complete decision chain: generation → scan → classification → review (if applicable) → approval → publication.
12. Human Review Pipeline (@themis/review-pipeline)#
Automated systems catch the obvious cases; human experts catch the subtle ones. A robust human review pipeline is essential for the "flag" zone between clear pass and clear reject.
12.1 Review Queue System#
- Priority queue — Flagged content enters a review queue prioritized by: risk score (higher risk → faster review), publication deadline, artist tier (higher-profile artists → faster review), and similarity confidence.
- Reviewer assignment — Route flagged content to qualified reviewers based
on expertise:
- Musicologist reviewers — For melodic, harmonic, and structural similarity disputes
- Lyricist reviewers — For lyric similarity disputes
- Legal reviewers — For high-risk cases requiring legal analysis
- Genre specialists — Reviewers with expertise in specific genres (hip-hop sampling culture, electronic music production, country songwriting conventions)
- SLA targets — 90% of flagged content reviewed within 4 hours. 99% within 24 hours. High-risk content reviewed within 1 hour.
12.2 Review Interface#
- A/B comparison tool — Side-by-side playback of the generated track and the matched reference track, with synchronized waveform display, aligned to the matching segments.
- Element isolation — Ability to solo/mute individual elements (melody, harmony, rhythm, vocals) in both the generated and reference tracks for focused comparison.
- Similarity heatmap — Visual timeline showing where similarity is concentrated (e.g., the chorus of the generated track matches bars 33-48 of the reference track).
- Legal context panel — For each flagged match, display relevant legal context: the reference work's copyright holder, known litigiousness of the rights holder, relevant case law, and jurisdiction-specific risk.
- One-click decisions — Approve, reject, or request modification with a single click. Approved and rejected decisions include mandatory rationale text.
12.3 Musicologist Network#
- Expert panel — Maintain relationships with 10-20 professional musicologists who can provide expert opinions on difficult cases.
- Escalation path — Content that internal reviewers cannot confidently assess is escalated to the expert panel.
- Expert opinion documentation — Expert opinions are documented and stored as part of the audit trail, providing evidence of due diligence.
13. Remediation Pipeline (@themis/remediation)#
When content is flagged or rejected, automatic remediation generates alternative content that preserves the creative intent while eliminating the similarity.
13.1 Targeted Element Modification#
- Melody variation — When the melody is the flagged element, automatically
generate variations that preserve the overall contour and emotional character
while altering specific note sequences to break the similarity match:
- Pitch substitution (replace matching notes with neighbor tones)
- Rhythmic variation (alter the rhythm while keeping the pitch contour)
- Intervallic variation (expand or compress intervals)
- Ornamental variation (add grace notes, passing tones, turns)
- Harmonic reharmonization — When chord progressions are flagged, suggest alternative harmonizations that preserve the harmonic function but use different specific chords (e.g., replace IV with ii, replace V with viio).
- Lyric reformulation — When lyrics are flagged, regenerate the matching lines while preserving the semantic meaning, meter, and rhyme scheme.
- Vocal modification — When vocal similarity is flagged, apply targeted voice modifications (formant shifting, vibrato alteration, breathiness adjustment) to increase distinctiveness.
- Rhythmic variation — When rhythmic patterns are flagged, introduce micro-timing variations, swap drum sounds, or alter the rhythmic density.
13.2 Regeneration Pipeline#
- Constrained regeneration — When targeted modification is insufficient, regenerate the entire section with negative constraints: "generate a chorus that is NOT similar to [reference work]." The similarity engine runs as a real-time constraint during generation.
- Iterative refinement — Regeneration runs in a loop: generate → scan → if flagged, regenerate with tighter constraints → scan again. Maximum 5 iterations before escalating to human review.
- Creative intent preservation — Track the original creative intent (mood, energy, genre, tempo, key, lyrical theme) and ensure that remediation preserves these while eliminating the similarity.
13.3 Pre-Generation Screening#
- Prompt risk assessment — Before generation even begins, assess the prompt
for copyright risk:
- References to specific artists ("make it sound like Drake") → high risk
- References to specific songs ("similar to Bohemian Rhapsody") → reject
- Genre/mood descriptions ("upbeat pop in C major") → normal risk
- Specific musical descriptions ("use a I-V-vi-IV progression") → low risk (common elements)
- Negative prompting — Automatically inject negative constraints into generation prompts: "do not reproduce any existing copyrighted melodies."
- Training data exclusion — If the prompt references a specific genre or era, cross-reference against known memorization hotspots in the generation model and inject additional guardrails.
14. Model Safety & Training Data Governance (@themis/model-governance)#
Prevention is better than detection. Ensuring that generation models are safe before deployment reduces the detection burden.
14.1 Training Data Audit#
- License verification — For every dataset used to train a music generation model, verify and document the license terms. Maintain a ledger: | Dataset | License | Covers Generative Use? | Opt-Out Mechanism | |---|---|---|---| | FMA (Free Music Archive) | CC-BY/CC0 | Yes | N/A | | MUSDB18 | Research only | No — cannot use for commercial generation | N/A | | Licensed catalog | Custom | Verify clause-by-clause | Artist request |
- Opt-out compliance — Honor all artist/rights-holder opt-out requests. Maintain an opt-out registry. When a model is trained, verify that all opt-out requests are honored in the training data.
- Deduplication audit — Verify that training data has been properly deduplicated. Duplicate training examples increase memorization risk.
14.2 Memorization Detection#
- Membership inference testing — After model training, run membership inference attacks against the model to detect which training examples it has memorized. Flag models with high memorization rates for additional deduplication or differential privacy.
- Verbatim reproduction testing — Feed the model prompts designed to elicit verbatim reproduction of training data (e.g., first 5 seconds of known training songs as continuation prompts). Measure reproduction rates.
- Canary testing — Insert known "canary" sequences into training data and test whether the model reproduces them. High canary reproduction = high memorization risk.
- Differential privacy verification — If differential privacy was used during training, verify the privacy budget (ε, δ) and ensure it provides meaningful protection against memorization.
14.3 Model Red-Teaming#
- Adversarial prompt testing — Before deploying a model, red-team it with
prompts specifically designed to elicit copyrighted content:
- "Continue this melody: [first 4 bars of a famous song]"
- "Generate a song in the style of [famous artist] about [specific topic of their famous song]"
- "Generate a beat that sounds like [famous producer]'s signature style"
- Systematic coverage — Red-team against a representative sample of the Billboard Hot 100, Grammy winners, and most-streamed tracks from each decade.
- Ongoing monitoring — Continuously red-team deployed models as new music is released. A model that didn't reproduce "Song X" last month might reproduce "Song Y" this month due to different prompt patterns.
15. Platform Integration & API (@themis/music-shield-api)#
The shield must be seamlessly integrated into every music generation pipeline.
15.1 Gating API#
interface MusicShieldScanRequest {
audio: Float32Array | Uint8Array | string; // audio data or URL
lyrics?: string; // lyrics text if available
metadata?: {
genre?: string;
tempo?: number;
key?: string;
generationPrompt?: string;
modelId?: string;
modelVersion?: string;
artistId?: string;
};
scanConfig?: {
engines?: ScanEngine[]; // which engines to run (default: all)
jurisdiction?: string[]; // jurisdictions to assess
riskThreshold?: RiskLevel;
maxLatencyMs?: number; // time budget for scan
};
}
interface MusicShieldScanResult {
overallRisk: RiskLevel; // PASS | ADVISORY | FLAG | HIGH_RISK | REJECT
overallScore: number; // 0.0 - 1.0 composite score
confidence: number; // confidence in the assessment
elementScores: {
melody: ElementScore;
lyrics: ElementScore;
vocal: ElementScore;
harmony: ElementScore;
spectral: ElementScore;
structure: ElementScore;
rhythm: ElementScore;
};
matches: SimilarityMatch[]; // matched reference works
provenance: ProvenanceRecord; // audit trail
recommendations: Recommendation[]; // suggested actions
scanDurationMs: number;
}
type RiskLevel = 'pass' | 'advisory' | 'flag' | 'high_risk' | 'reject';
interface SimilarityMatch {
referenceWorkId: string;
referenceTitle: string;
referenceArtist: string;
copyrightHolder: string;
matchedElements: MatchedElement[];
compoundScore: number; // compound similarity score
legalRiskAssessment: string; // natural language risk summary
timeSegments: TimeSegment[]; // where in the tracks the match occurs
}
15.2 Integration Points#
- Calliope integration — Every AI artist's music passes through the shield before publication. The shield is configured per-artist based on their genre, influences, and vocal identity.
- Euterpe integration — The DAW/studio engine calls the shield API before exporting any generated or AI-assisted content.
- Isis integration — The generative factory's music production pipeline includes the shield as a mandatory post-generation step.
- Nous integration — Leverages existing
@nous/audiofingerprinting and@nous/coresimilarity search infrastructure. - Real-time mode — For interactive generation (live jamming, real-time composition), a lightweight version of the shield runs in real-time, providing instant feedback with lower fidelity. Full scan runs asynchronously.
15.3 Performance Requirements#
- Full scan latency — Complete multi-engine scan in < 30 seconds for a standard 3-minute track. This is acceptable for batch/publication workflows.
- Quick scan latency — Melody + fingerprint scan in < 5 seconds. Used for interactive feedback during composition.
- Real-time mode — Continuous melody comparison at < 100ms latency. Used during live generation.
- Throughput — Support scanning 10,000+ tracks per day at full scan depth.
- Scalability — Horizontally scalable across scan engines. Each engine runs independently and can be scaled based on load.
16. Monitoring, Analytics & Continuous Improvement (@themis/shield-analytics)#
The system must learn and improve continuously.
16.1 False Positive/Negative Tracking#
- False positive tracking — When human reviewers override a FLAG or REJECT decision, record this as a false positive. Track false positive rates per engine, per genre, per reference work.
- False negative tracking — When a published work receives a third-party copyright claim that the shield did not catch, record this as a false negative. This is the most critical metric.
- Threshold tuning — Use accumulated false positive/negative data to continuously calibrate risk thresholds. The goal: zero false negatives with minimal false positives.
16.2 Analytics Dashboard#
- Scan volume metrics — Tracks scanned/day, scans by risk level, average scan latency.
- Match analytics — Most frequently matched reference works, most common match elements, match trends over time.
- Genre risk heatmap — Risk levels by genre. Some genres (pop, hip-hop) have higher match rates due to more formulaic structure; others (experimental electronic) have lower match rates.
- Model risk tracking — Track which generation models produce the most flagged content. Models with high flag rates may need additional training data deduplication or fine-tuning.
- Legal landscape tracker — Monitor new copyright infringement cases, legislative changes, and regulatory guidance that may require threshold adjustments.
16.3 Feedback Loop#
- Reviewer feedback → model improvement — Patterns identified by human reviewers (e.g., "this model consistently generates melodies similar to 2010s pop hooks") feed back into model training and prompt engineering.
- Post-publication monitoring — Monitor published tracks for third-party claims (Content ID, DMCA) and feed results back into the reference database and threshold calibration.
- Industry benchmarking — Compare our flag/reject rates against industry norms. If we're flagging 50% of content while competitors publish freely, our thresholds may be too aggressive.
Library Architecture#
libs/themis/
├── music-shield/ # Core shield orchestration
│ ├── src/
│ │ ├── shield.ts # Main MusicShield class
│ │ ├── config.ts # Shield configuration
│ │ ├── pipeline.ts # Scan pipeline orchestration
│ │ ├── decision-engine.ts # Risk classification logic
│ │ └── index.ts
│ ├── project.json
│ ├── package.json
│ └── tsconfig.json
│
├── melodic-analysis/ # Melodic similarity detection
│ ├── src/
│ │ ├── pitch-extraction.ts
│ │ ├── contour-analysis.ts
│ │ ├── ngram-matching.ts
│ │ ├── similarity-metrics.ts
│ │ ├── transposition.ts
│ │ ├── common-melody-filter.ts
│ │ └── index.ts
│ └── ...
│
├── harmonic-analysis/ # Harmonic similarity detection
│ ├── src/
│ │ ├── chord-extraction.ts
│ │ ├── progression-matching.ts
│ │ ├── common-progression-filter.ts
│ │ ├── protectability-scoring.ts
│ │ └── index.ts
│ └── ...
│
├── rhythmic-analysis/ # Rhythmic similarity detection
│ ├── src/
│ │ ├── beat-extraction.ts
│ │ ├── pattern-matching.ts
│ │ ├── groove-analysis.ts
│ │ ├── drum-break-db.ts
│ │ └── index.ts
│ └── ...
│
├── spectral-analysis/ # Spectral & timbral similarity
│ ├── src/
│ │ ├── feature-extraction.ts
│ │ ├── timbral-fingerprint.ts
│ │ ├── sample-detection.ts
│ │ ├── embedding-comparison.ts
│ │ └── index.ts
│ └── ...
│
├── lyric-analysis/ # Lyric similarity detection
│ ├── src/
│ │ ├── text-matching.ts
│ │ ├── semantic-similarity.ts
│ │ ├── phonetic-matching.ts
│ │ ├── common-phrase-filter.ts
│ │ ├── lyric-melody-joint.ts
│ │ └── index.ts
│ └── ...
│
├── vocal-analysis/ # Voice similarity & anti-impersonation
│ ├── src/
│ │ ├── speaker-embedding.ts
│ │ ├── vocal-characteristic.ts
│ │ ├── anti-impersonation.ts
│ │ ├── voice-modification.ts
│ │ ├── consent-registry.ts
│ │ └── index.ts
│ └── ...
│
├── structural-analysis/ # Song structure similarity
│ ├── src/
│ │ ├── section-segmentation.ts
│ │ ├── arrangement-analysis.ts
│ │ ├── energy-contour.ts
│ │ └── index.ts
│ └── ...
│
├── audio-fingerprint/ # Extended fingerprinting
│ ├── src/
│ │ ├── neural-fingerprint.ts
│ │ ├── content-id-proxy.ts
│ │ ├── near-duplicate.ts
│ │ ├── segment-matching.ts
│ │ └── index.ts
│ └── ...
│
├── reference-db/ # Reference database management
│ ├── src/
│ │ ├── melody-db.ts
│ │ ├── fingerprint-db.ts
│ │ ├── lyric-db.ts
│ │ ├── voice-db.ts
│ │ ├── ingestion-pipeline.ts
│ │ ├── deduplication.ts
│ │ ├── metadata-enrichment.ts
│ │ └── index.ts
│ └── ...
│
├── risk-scoring/ # Legal risk assessment
│ ├── src/
│ │ ├── composite-scorer.ts
│ │ ├── legal-precedent-db.ts
│ │ ├── jurisdiction-rules.ts
│ │ ├── compound-detection.ts
│ │ ├── protectability-analysis.ts
│ │ └── index.ts
│ └── ...
│
├── provenance/ # Audit trail & provenance
│ ├── src/
│ │ ├── generation-provenance.ts
│ │ ├── training-data-audit.ts
│ │ ├── decision-trail.ts
│ │ ├── opt-out-registry.ts
│ │ └── index.ts
│ └── ...
│
├── review-pipeline/ # Human review workflow
│ ├── src/
│ │ ├── review-queue.ts
│ │ ├── reviewer-assignment.ts
│ │ ├── comparison-tools.ts
│ │ ├── escalation.ts
│ │ └── index.ts
│ └── ...
│
├── remediation/ # Auto-fix & regeneration
│ ├── src/
│ │ ├── melody-variation.ts
│ │ ├── harmonic-reharmonization.ts
│ │ ├── lyric-reformulation.ts
│ │ ├── vocal-modification.ts
│ │ ├── constrained-regeneration.ts
│ │ ├── prompt-risk-assessment.ts
│ │ └── index.ts
│ └── ...
│
├── model-governance/ # Training data & model safety
│ ├── src/
│ │ ├── license-verification.ts
│ │ ├── memorization-detection.ts
│ │ ├── model-red-team.ts
│ │ ├── canary-testing.ts
│ │ └── index.ts
│ └── ...
│
├── shield-api/ # Platform integration API
│ ├── src/
│ │ ├── scan-api.ts
│ │ ├── real-time-mode.ts
│ │ ├── batch-mode.ts
│ │ ├── webhook-notifications.ts
│ │ └── index.ts
│ └── ...
│
└── shield-analytics/ # Monitoring & continuous improvement
├── src/
│ ├── false-positive-tracker.ts
│ ├── false-negative-tracker.ts
│ ├── analytics-dashboard.ts
│ ├── threshold-tuning.ts
│ ├── legal-landscape-monitor.ts
│ └── index.ts
└── ...
Technology Stack#
| Component | Technology | Rationale |
|---|---|---|
| Core similarity engines | TypeScript + ONNX Runtime | ML inference with type safety |
| Audio DSP (melody/chroma extraction) | Rust via WASM / napi-rs | Real-time performance for audio processing |
| Vector similarity search | pgvector + Qdrant | Approximate nearest neighbor for embeddings |
| Audio fingerprint matching | Rust (constellation matching) | High-throughput hash matching |
| Lyric NLP | TypeScript + sentence-transformers (ONNX) | Semantic embedding comparison |
| Reference database | PostgreSQL + Redis | Structured metadata + caching |
| Message queue (async scans) | BullMQ | Job queue for scan pipeline |
| API layer | TypeScript (Hono/Fastify) | Low-latency HTTP API |
| Review UI | React (within Yemaya admin) | Internal review interface |
| Monitoring | Prometheus + Grafana | Metrics and alerting |
Integration with Existing Infrastructure#
| Existing System | Integration |
|---|---|
@nous/audio (fingerprinting) |
Extend landmark/chroma/hybrid fingerprinting with copyright-specific matching |
@nous/core (similarity search) |
Use similarity search engine for embedding-based audio comparison |
@nous/audio (audio-search) |
Extend audio search with copyright reference database |
@calliope/voice (anti-similarity) |
Plug shield's vocal analysis into Calliope's voice creation pipeline |
@euterpe/genesis (generation) |
Shield gate on all generated music output |
@euterpe/theory (music theory) |
Use theory primitives for melodic/harmonic analysis |
@isis/audio-generation |
Shield gate on factory music generation |
@yemaya/agents (audio-style-enforcer) |
Feed shield signals into style enforcement |
Development Phases#
Phase 1: Foundation (Weeks 1-4)#
@themis/melodic-analysis— Pitch extraction, contour analysis, n-gram matching@themis/audio-fingerprint— Extend@nous/audiofingerprinting for copyright matching@themis/reference-db— Core database schema, MusicBrainz/AcoustID integration@themis/risk-scoring— Basic composite scoring with melody + fingerprint signals@themis/music-shield— Core pipeline orchestration, PASS/FLAG/REJECT decisions- Deliverable: Working end-to-end pipeline for melody + fingerprint scanning
Phase 2: Depth (Weeks 5-8)#
@themis/harmonic-analysis— Chord extraction, progression matching, common filtering@themis/lyric-analysis— Text matching, semantic similarity, phonetic matching@themis/spectral-analysis— MFCC/chroma features, timbral fingerprinting, sample detection@themis/provenance— Generation provenance, audit trail, decision logging- Deliverable: Multi-element scanning with lyric + harmonic + spectral analysis
Phase 3: Voice & Legal (Weeks 9-12)#
@themis/vocal-analysis— Speaker embeddings, anti-impersonation, consent registry@themis/structural-analysis— Section segmentation, arrangement analysis@themis/risk-scoring— Legal precedent database, jurisdiction-aware assessment, compound detection@themis/shield-api— Full REST API, webhook notifications, batch mode- Deliverable: Complete multi-engine scanning with legal risk assessment
Phase 4: Human Loop & Remediation (Weeks 13-16)#
@themis/review-pipeline— Review queue, reviewer assignment, comparison tools@themis/remediation— Melody variation, reharmonization, lyric reformulation, constrained regeneration@themis/model-governance— License verification, memorization detection, red-teaming- Deliverable: Full human-in-the-loop review and automatic remediation
Phase 5: Scale & Intelligence (Weeks 17-20)#
@themis/shield-analytics— False positive/negative tracking, threshold tuning, dashboards@themis/reference-db— Real-time ingestion pipeline, chart monitoring, viral content tracking- Performance optimization — Rust WASM for hot paths, caching, parallel scan engines
- Calliope/Euterpe/Isis integration — Deploy shield into all production music pipelines
- Deliverable: Production-grade, continuously improving shield at scale
Success Metrics#
| Metric | Target | Measurement |
|---|---|---|
| False negative rate | < 0.01% | Third-party claims on published works that shield missed |
| False positive rate | < 5% | Human review overrides of FLAG/REJECT decisions |
| Full scan latency (3-min track) | < 30s | P95 scan duration |
| Quick scan latency | < 5s | P95 melody + fingerprint scan |
| Reference database coverage | > 10M works | Indexed commercial recordings |
| Review SLA (flagged content) | < 4 hours (90%) | Time from flag to human decision |
| Remediation success rate | > 90% | % of flagged content successfully remediated without human intervention |
| Uptime | 99.9% | Shield API availability |
Legal Disclaimer#
This system is a risk mitigation tool, not a legal opinion generator. All HIGH_RISK and FLAG decisions should be reviewed by qualified music copyright attorneys before final publication decisions. The system's risk assessments are based on analysis of publicly available case law and musicological principles but do not constitute legal advice.
References#
Case Law#
- Arnstein v. Porter, 154 F.2d 464 (2d Cir. 1946) — foundational access + similarity test
- Bright Tunes Music Corp. v. Harrisongs Music, Ltd., 420 F. Supp. 177 (S.D.N.Y. 1976) — subconscious copying
- Three Boys Music Corp. v. Bolton, 212 F.3d 477 (9th Cir. 2000) — audience test for infringement
- Williams v. Gaye, 895 F.3d 1106 (9th Cir. 2018) — "Blurred Lines" groove/feel
- Gray v. Hudson, No. 2:15-cv-05642 (C.D. Cal. 2020) — "Dark Horse" reversal on commonplace elements
- Griffin v. Sheeran, No. 1:17-cv-05221 (S.D.N.Y. 2023) — "Thinking Out Loud" no infringement
- Structured Asset Sales v. Sheeran, No. 1:18-cv-05839 (S.D.N.Y. 2023) — chord progressions not protectable
Legislation#
- US Copyright Act, 17 U.S.C. §§ 101-1332
- NO FAKES Act (proposed) — digital replica rights
- EU AI Act (2024) — AI system transparency and risk classification
- EU Copyright Directive, Art. 3-4 — text and data mining exceptions
Technical References#
- Wang, A. (2003). "An Industrial-Strength Audio Search Algorithm" (Shazam fingerprinting)
- Müller, M. (2015). "Fundamentals of Music Processing" (DTW, chroma features)
- Serra, J. et al. (2009). "Cross recurrence quantification for cover song identification"
- Doras, G. & Peeters, G. (2020). "Audio-based detection of explicit content in music"
This proposal establishes Oshun's Themis Music Shield as the most comprehensive AI music copyright protection system in the industry — not as a competitive advantage, but as a moral and legal necessity. Every generated note must be defensible.