Leveraging NLP and LLMs for Music Industry Royalty Recovery

love_without_sound · Media & Entertainment · 2025 · source Agentic: no · promotional bias: low · source-backed · extracted with prompt 0.5.6 (schema v2)

Generated from extraction.json — do not edit by hand. Every claim cites verbatim source evidence; stated means the quote directly asserts the claim, implied means it supports it indirectly.

Problem

Domain: music royalty recovery and legal NLP

Love Without Sound needed to recover lost music royalties and streamline legal negotiations despite messy metadata, fragmented usage channels, high-volume correspondence, and strict privacy requirements.

Constraints:

• Music metadata is inconsistent across featured artists, live versions, remixes, and auto-formatted titles.
• The system must handle a two-billion-row metadata database, millions of tracks, legal emails, attachments, agreements, and case citations.
• Legal documents and financial information are confidential, so models and applications need to run locally in a data-private environment.
• The current production system is mostly custom NLP with spaCy and Prodigy; LLM/RAG work is emerging and should not be overstated.

Architecture

Love Without Sound built a modular NLP platform using spaCy pipelines, Prodigy annotation loops, transformer and CNN components, legal-document classifiers, case-citation extraction, request tracking, music-reference linking, and local private deployment. The production system standardizes music metadata, links song references to unique IDs, classifies legal correspondence and attachments, detects settlement details, predicts negotiation direction, and is expanding toward a private RAG/SQL query layer and audio embeddings.

Diagram confidence: explicit (topology described by the source).

flowchart LR
  Metadata_standardization_pipeline["Metadata standardization pipeline"] --> Music_metadata_database["Music metadata database"]
  Prodigy_annotation_loop["Prodigy annotation loop"] --> Metadata_standardization_pipeline
  Prodigy_annotation_loop --> Model_training_runtime["Model training runtime"]
  Email_body_extractor["Email body extractor"] --> Legal_correspondence_classifier["Legal correspondence classifier"]
  Legal_correspondence_classifier --> Request_tracking_dashboard["Request tracking dashboard"]
  Case_citation_detector["Case citation detector"] --> Negotiation_intelligence["Negotiation intelligence"]
  Music_reference_linker["Music reference linker"] --> Music_metadata_database
  Music_reference_linker --> Negotiation_intelligence
  Attachment_processor["Attachment processor"] --> Negotiation_intelligence
  RAG_and_SQL_query_layer["RAG and SQL query layer"] --> Music_metadata_database

Components

Data flow

Raw music metadata is normalized by spaCy NER and text classification into song, artist, and modifier structures with hierarchical IDs. Legal emails and attachments pass through body extraction, correspondence classification, citation extraction, agreement section classification, and request tracking. Extracted song references are linked to unique music identifiers so licensing claims can be verified. Prodigy keeps humans in the annotation loop for new examples and edge cases, while Modal supports training and batch processing at scale.

Why this architecture

The architecture decomposes a complex royalty-recovery business problem into modular, separately evaluated NLP components. Local deployment preserves privacy, Prodigy lets the domain expert iteratively correct schemas and edge cases, and the future RAG/SQL layer is deliberately scoped to translate questions into queries rather than use a broad generative system.

Models used: spaCy · Prodigy · transformer-based NER · CNN-based text classification · span categorization · Modal · RAG pipeline · small on-premise LLM · audio embedding model

Scale

• Love Without Sound standardizes metadata across a two-billion-row database.
  “normalize and standardize song and artist information across a 2 billion-row database”article, stated
• Spotify receives about 40,000 new tracks per day and 15% contain incorrect metadata.
  “Spotify is receiving approximately 40,000 new tracks per day and 15% of them contain incorrect metadata.”article, stated
• An estimated $2.5 billion in U.S. royalties remained unallocated between 2016 and 2018 due to metadata issues.
  “Estimates suggest $2.5 billion in royalties remained unallocated in the U.S. alone between 2016 and 2018 due to metadata issues.”article, stated
• Legal workflows can involve thousands of emails per day.
  “This process can also involve extensive legal correspondence and negotiations, with thousands of emails being sent per day.”article, stated
• The metadata modifier classifier reports 0.99 AUC at 447,493 words per second.

  "Metadata Extraction (Modifiers)3

textcat

5

0.99

447,493"

— article, stated

Alternatives considered (roads not taken)

Treat metadata errors as manual cleanup

Rejected because: The problem spans billions of rows, thousands of daily emails, and millions of lost royalties, making manual monitoring impractical.

Build one monolithic model

Rejected because: The case chose separate modular components that can be combined, improved, and evaluated independently.

Use a broad LLM for private legal and catalog data

Rejected because: The source emphasizes local data-private deployment and scopes future LLM use to SQL translation.

Patterns

human-in-the-loop-gate (data development)

Prodigy keeps the model in the loop while Jordan reviews predictions, corrects edge cases, and retrains as new data arrives.

data-quality-loop (data)

Annotation, correction, schema iteration, and retraining are used to keep extraction definitions and model outputs current as catalogs and client requests change.

batch-offline-processing (operations)

Modal provides serverless CPU/GPU resources for model training and large batch jobs without local infrastructure management.

privacy-preserving-processing (governance)

Models and applications run locally in a data-private environment because legal and financial information is confidential.

rag-retrieval (retrieval)

A work-in-progress private RAG pipeline will answer natural-language questions about case history and artist information by generating SQL queries.

Tradeoffs

Model architecture

Chose Separate modular NLP models over One combined system for all legal and music tasks.

Rationale: Separate models allow each feature to be combined freely and evaluated independently while preserving speed and privacy.

LLM role in future RAG

Chose Natural-language-to-SQL translation with a smaller private model over Broad generative reasoning over confidential data.

Rationale: The LLM job is tightly scoped so it can run privately and support controlled querying of case history and artist information.

Annotation workflow

Chose Model-in-the-loop correction over Static one-time annotation.

Rationale: New client requests and catalog additions require a repeatable path for adding edge cases and refreshing models.

Gotchas (what bit them)

Music metadata errors directly misroute royalties.

Root cause: There is no standard format for featured artists, live versions, remixes, or even title formatting across tools.

Resolution: Normalize song, artist, and modifier fields with a dedicated NLP metadata pipeline and hierarchical IDs.

Legal emails contain important content buried in boilerplate.

Root cause: Negotiation emails are sandwiched between multiple disclaimer blocks and mixed with newsletters or meeting invitations.

Resolution: First detect message body boundaries, then classify substantive correspondence and extract structured fields.

Initial label schemes are often under-specified.

Root cause: The domain expert discovers ambiguous definitions only after interacting with real examples.

Resolution: Use Prodigy to iterate on schemas, definitions, and corrected examples before retraining.

Evaluation

Each NLP component is evaluated separately with published accuracy and speed metrics; Prodigy supports iterative correction, and Modal supports training and batch processing.

Metrics: NER F-score · text classification AUC · words per second · component-level accuracy · legal research time reduction

ner

1

0.98"

Outcomes

• The platform helps publishers recover royalties and streamline legal negotiations.
• The legal citation system reduced research time by nearly 50%.
• The published component table reports strong accuracy and speed across metadata, correspondence, attachment, agreement, and citation tasks.

The system operates across a two-billion-row metadata database, millions of tracks, thousands of daily emails, and highly confidential legal/financial data.